| method | what it scored |
|---|---|
| velocity 2026-08-29 · does how FAST it is moving predict? | yes, and BACKWARDS: the faster it moves when we buy, the worse we do ⚠ 5 factors now agree every 'has already moved' reading is inverted |
| rotation_btc_volume 2026-08-29 · does BTC volume predict money spreading to the board? | monotone; crossover ~5x normal BTC volume ⚠ stage 2->3 (rotation -> prices) does NOT pay in a flat regime and has never been tested in a BTC run |
| money_flow_ledger 2026-08-30 · net dollars per coin, (buy-sell)xprice | moves with price in 153 of 158; the 5 failures are stablecoins and FX |
| room_gauge 2026-09-01 · how big is the tradeable room right now? | BTC far off its day high => 1.7x the money, 1.34x buyable coins, 1.24x bigger moves, and NO directional difference. LARGE room carried 42 run events/7 big vs SMALL 28/3. ⚠ LIVE ON THE BOARD. A measurement, not a forecast. |
| run_anatomy 2026-09-01 · what does the tape show before and at the start of a run? | the 2h before is a FLUSH: turnover 8.15x its own normal, net selling -2.52x, price -1.05%. The first 30min IS the move - every decile negative after the leg, worst where the leg was biggest. ⚠ descriptive only; see flush_conditional for why it is not a filter |
| limit_at_the_pullback 2026-09-01 · is waiting for the pullback worth anything, or should we just buy when the setup fires? | ⭐ REPLICATED ALMOST EXACTLY ACROSS TWO INDEPENDENT WINDOWS AND TWO BAR SIZES: limit +2.726% vs market +1.021% = +1.704pp (30d hourly); limit +1.530% vs market -0.177% = +1.707pp (120d, 4h). Waiting for the pullback is worth about +1.7pp. ⚠ ⭐ this is the operator's 'we bought at the bottom of the pullback', isolated and measured. It is the most stable number the project has produced. ⚠️ It is the ENTRY TECHNIQUE, not a complete strategy — it says how to buy, not what to buy or when to sell. |
| missed_fills_are_the_runners 2026-09-01 · the 16% of pullback setups the limit never catches — are they the good ones? | YES. The 39 never filled went +11.149% over the same span; the 211 filled went -0.555%. The limit systematically declines the runners. ⭐ BUT scored PER SETUP with a no-fill counting as a real 0.0: LIMIT +0.716% vs MARKET (buy every signal) +0.590% — the limit still wins, by +0.126pp. ⚠ ⭐⭐ THIS RESIZES THE HEADLINE. The +1.704pp/+1.707pp 'waiting for the pullback is worth' number is about PRICE GIVEN A FILL. At the strategy level, counting the runners it declines, the advantage is +0.126pp. Both true; only the second is a strategy. |
| entry_exit_matrix 2026-09-01 · an array of entries x an array of exits, matched, with entries-not-made recorded | ⭐ RETRACE DEPTH IS MONOTONE AND THE 4-POINT GRID COULD NOT SEE IT: limit_75 +0.605% and positive against 22 OF 22 exits; limit_60 +0.210% (16/22); limit_50 -0.031% (10/22); limit_25 -0.562% (3/22); limit_15 -0.690% (2/22). The deeper the pullback you wait for, the better. ⭐ EXITS: hold2/3/4 all positive (+0.34..+0.41, 10-12 of 15 entries); trail +0.437% (11/15); hold24 -1.593% (1/15); every target negative; stop3 -0.294% and stop5 -0.647%. ⛔ CONFIRMATION IS EXPENSIVE: turn 2 of 22, turn2 5 of 22, reclaim_vwap 0 OF 22. Best cell limit_75+hold3 at +1.364%/setup is INSIDE a luckiest-of-330 null (+4.478%). 128 of 330 cells positive. ⚠ ⭐⭐ widening on operator instruction immediately paid: 0.25 (the first registered entry) sits in the WORST region and 0.75 is positive against every exit — the more literal reading of 'we bought at the bottom of the pullback'. ⛔ Still read as marginals, never as a best cell. |
| runup_window_sweep 2026-09-01 · the run-up window was fixed at 6 bars and never varied — is it fitted? | NOT FITTED — a BROAD PLATEAU. net by window: 3 bars -1.00%, 4 bars -0.22%, 6 bars +0.85% (registered), 8 bars +0.98%, 12 bars +1.07%, 18 bars +0.48%. Positive and stable from 6 to 18 bars; only the very short windows fail. ⚠ the registered 6 sits inside the plateau but not at its top (12 is). Deliberately NOT moved — changing it to the observed best is the exact fitting the pre-registration exists to prevent. |
| trade_journal 2026-09-02 · can a METHOD be distilled from outcomes rather than assembled from rules? | BUILT, not yet answered. First backfill (Aug 1 onward, 91 setups, 65 settled) immediately REPRODUCED THE CAPITAL FINDING ON A SECOND WINDOW: primary exit 30/65 up, mean +0.744% net unconstrained -- but only 28 were takeable one-at-a-time, mean -0.565% net. THE SIGN FLIPS AGAIN. Two independent windows now agree that the unconstrained number is positive and the realistic one is negative. ⚠ ⭐⭐⭐ THE PREREQUISITE LAW #3 NAMES. 'Distilling needs DECISION + OUTCOME' -- forward_suite records aggregates, so 'what happened during that trade that we could improve on' had no source. A mean and a win rate cannot be interrogated; a journal can. ⭐ The PATH column is the one that answers it: a loser that was +4% before it turned is a different problem from one that went straight down. ⛔ It records and decides nothing; the pre-registered candidates are untouched. |
| portfolio_sizing 2026-09-02 · what does the strategy do under a real capital model — % of equity per trade, capped by the coin, as many concurrent as qualify? | Monotone in risk up to ~50%, then it breaks (over-betting). Candidate A: +11.4% at 5% risk (maxDD -6.0%) · +22.9% at 10% (-11.8%) · +60.3% at 33% (-31.2%) · +89.4% at 50% (-43.8%) · -0.8% at 100%. C: +17.2% / +36.7% / +164.3% / +198.2% / +344.5%. ⭐ COIN CAPACITY NEVER BINDS at this fund size (median $3,206 vs a 10% ticket of $140); only CASH binds, and only above 33% risk. Max concurrent positions 9, but the MEDIAN while in market is 1. ⚠ ⭐⭐⭐ THIS CORRECTED A CONCLUSION I HAD REPORTED. Candidate A read -59% under one-at-a-time and +22.9% under the real model. With an outlier-driven return, HOW MANY POSITIONS YOU CAN HOLD IS THE STRATEGY -- holding six things means six chances to be in the one that runs. ⛔ The risk sweep is a SELECTION: read the shape (monotone then collapse), never the best cell. ⛔ IN-SAMPLE. |
| why_every_exit_rule_fails 2026-09-03 · WHY do targets and stops both lose, when losers are so often profitable first? | THE SKEW EXPLAINS IT. 77% of losing trades touched +1% at some point, 61% touched +2%, 45% touched +3%. Median MFE of a loser +2.51%; of a WINNER +11.03%. Median MAE of a loser -7.19%; of a winner -2.42%. So a target that rescues the 61% of losers which touched +2% must also cap the winners whose median run is +11% — and the winners are the entire return. Symmetrically a stop tight enough to cut the -7% losers also clips winners that routinely dip -2.4% first. ⚠ ⭐⭐⭐ THIS IS WORTH MORE THAN ANOTHER FAILED RULE: it says WHY the whole exit family fails, and predicts that any exit acting on the SMALL move destroys more winner-upside than it saves. The only exits that can help must discriminate on something other than the size of the move so far. That is the AI-decides-at-the-sticking-point frontier, and the journal now carries the paths to train it. |
| perfect_play_ceiling 2026-09-03 · working backwards — if the data we have had been traded PERFECTLY, what would the outcome have been? | MEAN PER TRADE: L0 our rule +1.00% · L1 perfect EXIT on our setups +8.66% · L2 perfect SELECTION (only the 45% that win), our exit +7.18% · L3 perfect both +16.39%. ⭐ THE GAP DECOMPOSES: perfect exit is worth +7.66pp per trade, perfect selection +6.18pp — THE EXIT IS THE BIGGER LEAK, and both are ~7x our entire current edge. The perfect ex-post hold has a MEDIAN of 5 bars (20h) against our registered 2 (8h). ⛔ BUT A FIXED LONGER HOLD DOES NOT CAPTURE IT: 1 bar +0.33% · 2 +1.00% · 3 +0.28% · 4 -0.04% · 5 -0.42% · 12 -0.70%. Hold 2 is already the best fixed choice. The perfect exit differs per trade and a fixed rule cannot chase it. L4 whole-universe oracle: 85 trades, 40 distinct coins, best picks ESPORTS +157%/+117%, AKE +85% — the compounded figure is meaningless fantasy. ⚠ ⭐⭐⭐ THIS IS THE ANSWER TO 'IS A DECISION LAYER WORTH BUILDING'. The prize is large — capturing even a SIXTH of the exit gap would double the current edge — and it is concentrated in the EXIT, which is also where a fixed rule provably cannot help (hold 2 is already optimal among fixed holds). That is the precise shape of a problem that needs a per-trade decision rather than a parameter. ⛔ Perfect foresight ceilings are always high; what this bounds is the PRIZE, not the achievable. |
| trade_taxonomy 2026-09-03 · before building an AI exit layer — are these trades one kind or several, and is the kind visible AT ENTRY? | THEY ARE EMPHATICALLY SEVERAL KINDS, and they want opposite handling: IMMEDIATE 8 (4%) best hold 1 · BUILDER 44 (21%) best hold 8 (+9.40%) · DIP-RUN 53 (25%) best hold 12 (+9.84%) · SPIKE-FADE 38 (18%) best hold 1 · DEAD 70 (33%) best hold 1 (-4.27%). The best fixed hold ranges from 1 bar to 12 bars ACROSS TYPES — the operator's 'subtle differences' are not subtle. ⚠ ⭐ this is why a single fixed hold cannot win: it is one answer to five questions. See the companion entry for whether the type is knowable at entry. |
| daily_winner_path 2026-09-03 · the day's biggest buyable winner — where did it start and how did it get to the high over the COURSE OF THE DAY? | ⭐ THE MOVE TAKES HALF A DAY, NOT AN HOUR: median 12h from start to peak, range 2-23h, and 96% take 4h or more. ⭐⭐ THE VOLUME IS U-SHAPED THROUGH THE CLIMB: first third 1.83x the coin's own normal, middle 1.07x, LAST THIRD 2.38x — heaviest at the TOP. ⭐⭐ THE CLIMB IS NOT A LINE: 42% of the hours during it close DOWN, the median deepest drawdown while climbing is -2.8% (worst -11.4%), and only 19% of the whole gain arrives in the first third. ⭐⭐⭐ SO THE ENTRY IS FORGIVING: entering 1h late still captures 84% of the run, 2h late 85%, 3h late 68%, 6h late 62%. ⚠ ⭐⭐⭐ THE OPERATOR'S CORRECTION WAS RIGHT AND IT CHANGES THE TARGET. 'Catch the bottom' was the wrong frame — only 19% of the gain is in the first third and you can be two hours late and keep 85%. The hard part is not the entry. ⭐ AND THE VOLUME SPIKE AT THE TOP (2.38x against 1.07x mid-climb) is an EXIT candidate, which is where the leak actually is. ⛔ DEFINITIONAL CAVEATS, stated: searching for the best buy/sell pair INSIDE a UTC day forces starts early and peaks late, so the clock clustering (starts 00-04h, peaks 18-23h) is an artefact, and the 12h median is bounded by the 24h window. The volume shape, the 42% down-hours, the -2.8% drawdown and the late-entry capture are NOT definitional. |
| buy_sell_pressure 2026-09-03 · are people buying or selling, and at what pace — the operator's lifelong constant, tested | ⭐ CONFIRMED AS A DESCRIPTION. Buy share through the trade — RAN vs FADED: the hour before 0.564/0.567 · at the trigger 0.660/0.633 · +15-60min 0.575/0.480 · +1-2h 0.530/0.467 · +2-4h 0.540/0.435. 78% of runs kept buy share above 0.5 the whole 4h. The fades FLIP to sellers; the runs do not. ⛔ BUT IT IS NOT PREDICTIVE AT THE TRIGGER: buy share prior-hour AUC 0.502, at-trigger 0.504, pace 0.530 — all INSIDE a day-shuffled null of 0.109. The divergence only opens 15-60 MINUTES AFTER entry. ⛔ AND AS AN EXIT CHECK IT DOES NOT PAY: hold-4h-no-check -1.06%; exit at +1h if buy share <0.50 gives -0.96%, INSIDE a null of -0.866%. ⚠ ⛔⭐ FIRST, THE MISTAKE HE CAUGHT: every study run on 2026-09-03 used Kraken OHLC candles, which have NO BUY/SELL SPLIT AT ALL — a coin being dumped and one being accumulated look identical there. The constant he describes was invisible in the data those studies used. ⭐⭐ THE REFRAME THAT SURVIVES: at the pop, two thirds of the money is buying in BOTH cases — everyone chases every pop. What separates a real move is whether buying CONTINUES after the excitement, and that is knowable within the hour. It is a MONITORING signal, not a selection one — which is exactly how he described using it ('something I always watched'). ⛔⛔ THE BINDING LIMIT: our buy/sell tape starts 2026-08-25 and the hot days were 08-19/20/21, so THERE IS NOT ONE HOT DAY IN THE ONLY DATA THAT CARRIES BUY/SELL. The base strategy loses in this ordinary-regime sample (-1.06%), so this tests a filter on something that does not work here anyway. The next hot day is the first chance to test the constant where it matters. |
| btc_up_amplifies_it_does_not_create 2026-09-03 · operator: 'this proves my thinking on the bitcoin being up part, doesn't it?' — does a rising board create the opportunity, or does it do something else to it? | ⛔ HALF YES, AND THE HALF THAT FAILS IS THE ONE THE GATE WAS BUILT ON. OPPORTUNITY DOES NOT NEED BTC UP: per-coin harvestable movement is +11.2% on the 5 up days and +11.7% on the 5 flat/down days — identical. corr(board move, per-coin movement) = +0.313 (spearman +0.261, n=10). The BEST day of the ten was 09-02, board -0.03%, at +22.1% per coin — MORE than today's hot +15.4%. Number of buyable coins is uncorrelated with the board (+0.025). ⭐⭐⭐ BUT THE PAYOFF TO PICKING RIGHT DOES SCALE WITH IT: corr(board move, top-half afternoon movement) = +0.720, corr(board move, good-minus-bad pick spread) = +0.532. 09-03 (+5.21%) had the WIDEST good-vs-bad spread of the ten days at +6.70pp. ⚠ ⭐⭐⭐ A RISING BOARD DOES NOT CREATE OPPORTUNITY — IT MAKES THE GOOD PICKS PAY MORE. That is precisely 'a following wind', and it is a POSITION-SIZE dial applied to coins already identified, not an entry gate deciding whether to trade at all. ⭐⭐ IT ALSO REVIVES THE LEVERAGE DIAL: leverage_dial_test S3 failed because nothing we own forecasts the board, so the dial had nothing to multiply. This needs no forecast — BTC-from-open is knowable LIVE, and the persistent coin ranking is the thing it multiplies. ⛔ n=10 DAYS. Correlations on ten points are weak evidence and both of these are best-of-several readings on the same small sample. ⛔ It also revises the standing doctrine that the strategy only pays on hot days: opportunity is FLAT across regimes, so waiting for 2.3 days a month costs the other 27 for no measured gain. |
| oscillation_persistence_survives_liquidity 2026-09-03 · is the morning->afternoon oscillation persistence just a thinness ranking (the trap that killed the 8/30 net-flow finding, which collapsed in the liquid half), and does PACE explain it? | ⭐⭐⭐ IT SURVIVES EVERYTHING AND GETS STRONGER WHERE IT IS EASIEST TO TRADE. Pooled spearman(morning sum-of-ups, afternoon sum-of-ups) = +0.728 over 325 coin-days. It is NOT a liquidity ranking: spearman(morning turnover, morning sum-of-ups) = -0.134 and turnover->afternoon = -0.255. By liquidity half: THIN +0.653, LIQUID +0.796 — the OPPOSITE of the net-flow finding, which died in the liquid half. Monotone across turnover quintiles: +0.616 / +0.681 / +0.757 / +0.768 / +0.772, STRONGEST in the most liquid quintile. ⛔ AND PACE DOES NOT EXPLAIN IT: spearman(morning pace, afternoon sum-of-ups) has median -0.117 and is positive on only 2 of 10 days. The thing that persists is the MOVEMENT ITSELF, not the activity underneath it. ⚠ ⭐⭐⭐ THE FIRST READING IN THIS PROJECT THAT IS STRONGER IN THE LIQUID HALF. Every previous candidate was a thinness tilt (8/30 net flow), a beta ride (8/31 realised vol) or a variance tell — all of which weaken or reverse exactly where a real order can be filled. ⭐ The connection the operator asked for is real but it is NOT in the three trade readings: it is in the PRICE PATH itself. A coin's own recent oscillation is the only thing that forecasts its future oscillation. ⛔ STILL NO MONEY TEST. The zigzag is hindsight, and 'the ranking persists' is not 'an entry and exit capture it'. That is the next build and it is the one that decides everything. |
| bottom_of_pullback_signature 2026-09-04 · operator: 'why are we buying back in after a spike, wouldn't we look for a buy after the pull back at the bottom?' and 'we should be looking for the signal that it's at the bottom of the pull back.' Two tests: which re-entry trigger actually works, and — the real question — does the textbook bottom signature (a hammer: long lower wick, closing high after rejection) beat the current signature (closing at the LOW)? | ⛔ ONLY THE FULL SPIKE WORKS AS A RE-ENTRY: full spike 3.196% vs banking 3.000%; 'closed at the bottom' alone fires on 99% of trades within 12 bars and returns 2.986% — it is not a signal, it is a common bar shape. Adding 'below the fill price' makes it WORSE (2.670%), the fifth time tonight that demanding a cheaper price selected worse trades. ⛔⭐ THE ABLATION SHRINKS THE SIGNATURE'S CREDIT: full signature +1.132% against EVERY BUYABLE BAR at +0.995% — only +0.137pp, not the +0.435pp the 1,099-bar random-draw control suggested. vol>=2x ALONE is +0.752%, WORSE than no filter at all. range+close without volume gets +1.073% on 68% more signals. ⭐⭐⭐ AND THE TEXTBOOK BOTTOM LOSES OUTRIGHT. Hammer (lower wick >= half the range) +0.731% · hammer closing HIGH +0.729% · hammer + heavy volume +0.400% · hammer + vol + range +0.675% (1.083%/0.202%, unstable) — every variant BELOW the +0.995% of taking any bar, and all fill less often (61-66% vs 79%). ⚠ ⭐⭐⭐ THE COUNTERINTUITIVE RESULT IS THE ROBUST ONE: a bar CLOSING AT ITS LOW is a better bottom detector than a bar that BOUNCED off its low. With a hammer the reversal already happened inside the bar, so you buy after the turn; buying the weak close means buying while it is still being sold and taking the reversion in the following bars. That is the project's 'buy weakness' doctrine confirmed for roughly the tenth time, now against the single alternative most likely to beat it. ⭐⭐ AND IT PROPERLY SIZES THE ENTRY'S EDGE: +0.137pp of expectancy over every buyable bar. The strong capital-simulation result (12/12 against random draws at every slot count) therefore comes substantially from SLOT DYNAMICS — a 79% fill rate turning capital over faster — not from per-bar expectancy. Both are true and the earlier framing overstated the second. ⛔ volume alone being WORSE than no filter means the conjunction, not any single component, is what works. |
| why_clustered_signals_are_worse 2026-09-04 · operator: 'let's find out why the clustered signals are worse.' 100 slots captured 58% of signals and returned +5.1% while 50 slots captured 41% and returned +10.0% — the marginal signals are bad. Why? | ⛔ IT IS NOT THE CLUSTERING. Cluster size is NOT monotone in outcome: alone -0.111% · 2-3 -0.669% · 4-6 -0.865% (worst) · 7-11 -0.418% · 12+ -0.356%. As a live filter it is worthless: skipping bars with >2/>4/>6/>10/>20 simultaneous signals moves the mean by -0.099pp to +0.045pp. ⭐⭐⭐ THE CAUSE IS UNDERNEATH IT AND IT IS MONOTONE — BTC'S OWN 4h BAR. BTC bar < -2%: 120 signals, median cluster 19, fill 38%, net -1.191% · -2% to -0.5%: 454, cluster 7, fill 35%, -1.014% · FLAT: 389, cluster 2, fill 49%, +0.037% · BTC > +0.5%: 134, cluster 3, fill 44%, +0.197%. Clustering is a SYMPTOM: when BTC drops, dozens of coins capitulate at once and none of them bounce because the whole market is falling. ⛔⛔ AND FILTERING ON IT DOES NOT HELP THE PORTFOLIO. Per signal the filter is monotone and clears both halves at three thresholds (-0.513% -> +0.078% at the -0.5% cut, keeping 48%), but under real capital at 8 slots: NO FILTER 89 trades +10.6% maxDD 3.6% (halves +7.7%/+7.7%) vs skip-below--0.5% 75 trades +8.1%, skip-below-0.0% 123 trades +7.4% maxDD 7.7%, skip-below-+0.5% 107 trades +7.0% maxDD 1.8%. ⚠ ⭐⭐⭐ THIS EXPLAINS THE CAPACITY CEILING. With 8 slots the strategy takes only ~8% of signals, so SLOT SCARCITY IS ALREADY THE FILTER and it is a better one — at 100 slots you reach down into the marginal signals, which are exactly the ones firing while BTC falls (-1.0 to -1.2%). Scarcity was doing the filter's job better than the filter does. ⭐⭐ AND IT IS THE OPERATOR'S OWN FACTOR #1 ARRIVING FROM A NEW DIRECTION — 'BTC's day is the regime gate', except at the BAR level and INVERTED: not 'only trade hot days' but 'do not buy while BTC is falling'. Consistent with the standing crash-day law. ⭐⭐⭐ THE DEEPER LESSON, and it is the fourth time today: A FILTER THAT IMPROVES THE AVERAGE SIGNAL CAN HURT THE PORTFOLIO, because a slot-constrained portfolio never takes the average signal — it takes whatever is available when a slot frees. Per-signal and capital-constrained scoring disagree again, and only the second is money. ⭐ ONE REAL OPTION FOR THE OPERATOR: skipping when BTC's bar is below +0.5% HALVES the drawdown (3.6% -> 1.8%) for 3.6pp less return — a risk choice, not an edge improvement. ⛔ NOT added to the registration; the frozen rule takes every signal a slot can hold. |
| what_the_winners_do_inside_the_window 2026-09-04 · operator: 'do the strong buys do anything in that 60 minute window that would trigger a buy during that time since it did something that made it stick out?' Six candidate EVENTS inside the window, each scored on how much more often winners trigger than losers, and on the money if you buy at the trigger. | ⭐⭐ THEY DO — THEY TICK UP THROUGH THE SPIKE CLOSE, AND IT IS THE CLEANEST SEPARATION OF THE SESSION. First tick +1.0% over the close: 48% of winners vs 22% of losers, LIFT 2.18. First tick +0.5%: 70% vs 44%, lift 1.57. ⛔ Everything else is noise: recovers-after-a-dip 30% vs 26% (lift 1.17), two rising 10-min lows 100% vs 96% (1.04 — fires on everything), buy share >0.55 94% vs 93% (1.01), buy share >0.60 91% vs 93% (0.98, slightly INVERTED). ⛔⛔ AND BUYING ON THE EVENT IS WORSE THAN WAITING: +1.0% trigger +0.180% per signal, +0.5% trigger +0.256%, against Candidate G (wait the full 60min, buy if above -0.5%) at +0.731% and a baseline of +0.167%. By the time it has ticked +1% you are paying up and there is less room to +3.681% — and you also buy the 22% of losers that ticked and then died. ⚠ ⭐⭐⭐ THIS SHARPENS THE LAW #3 CASE INSTEAD OF WEAKENING IT. Two facts now exist that NO SINGLE THRESHOLD CAN COMBINE: the winners tick above the close (informative, lift 2.18) AND buying on the tick costs more than it is worth (worse price). A rule must choose one. A judgement can hold both — 'it ticked +1.2% twenty minutes ago and has settled to +0.1% with thirty minutes left' is a state a model can price and a threshold cannot even represent. That is exactly the operator's framing: 'the buy could be anytime within this window of 60 minutes.' ⭐ Buy share is dead here for the fourth time (lift 1.01 and 0.98) — it fires on essentially every trade regardless of outcome. ⛔ 14th BROKEN CHECK: my 'median minutes to trigger' column printed MILLISECONDS (1,020,538) — the subtraction was never divided by 60,000. The lift and money columns are unaffected; the timing column was wrong and is not quoted. ⛔ n=60. |
| indicators_rotate_through_the_window 2026-09-04 · operator: 'look at what the other indicators are doing, this could be a mix of indications' and then 'see how those measures are doing different things earlier and later in the decision process. it's the same indicator but it's telling us different things at different times.' | ⭐ THE WHOLE-WINDOW SWEEP: dollar_share AUC 0.694 · vs_own_vwap 0.682 · price_end 0.650 · price_max 0.633 · PEOPLE_share 0.633 · people_minus_dollars 0.366 (INVERTED — in winners the DOLLARS are more bullish than the PEOPLE, i.e. bigger buyers; in losers more people at smaller size) · tradesize_vs_pre 0.375 (inverted) · realised_vol 0.590 · pace/gap/large-share/building all noise. Best is INSIDE a day-blocked luckiest-of-14 null (0.194 vs p95 0.208) — 14 readings on n=60 cannot clear — but FOUR cluster at 0.63-0.69, which is not one lucky column. ⭐⭐⭐ AND THE SEGMENT SPLIT CONFIRMS HIS HYPOTHESIS WITH TWO CLEANLY MONOTONE, OPPOSITE SHAPES. PRICE RISES 0.497 -> 0.597 -> 0.637 -> 0.650 across the four 15-min segments — DEAD CHANCE in the first quarter hour. PACE FALLS 0.583 -> 0.531 -> 0.480 -> 0.473 — the BEST reading on the board in the first quarter hour and useless by the last. vs_vwap and people_share peak in the MIDDLE (0.657 and 0.633 at 30-45min). dollar_share is erratic (0.582/0.421/0.636/0.463) and should be read as noise. ⚠ ⭐⭐⭐ PRICE AND PACE SWAP PLACES INSIDE THE WINDOW. At minute 10 price carries NO information (0.497) and pace carries the most (0.583); by minute 55 pace is dead (0.473) and price is strongest (0.650). They are complementary IN TIME, and they are the two readings with clean monotone shapes across all four segments — far more convincing than the erratic ones. ⭐⭐⭐ THIS IS THE SHARPEST ARGUMENT YET FOR THE LAW #3 AI DECISION AND IT IS NOW EMPIRICAL RATHER THAN A HUNCH: a fixed formula must weight pace and price identically at minute 10 and minute 55; the data says minute 10 should be almost entirely PACE and minute 55 almost entirely PRICE, with vwap and people-share carrying the middle. NO THRESHOLD CAN EXPRESS A WEIGHTING THAT ROTATES THROUGH A WINDOW. ⛔ n=60 and the whole-window best is inside its null. This locates the STRUCTURE of the decision, not its size. |
| the_simulator_was_wrong 2026-09-04 · operator: 'fix the simulator and re-run everything through it.' Two capital simulations written hours apart reported +31.7% and +4.9% for the SAME rule on the SAME data, differing only in how often position exits were evaluated. | ⛔⛔⛔ EVERY POSITIVE PORTFOLIO NUMBER FROM THE SESSION WAS A SIMULATOR ARTEFACT. Corrected: all 159 symbols no stop -11.6% · CRYPTO ONLY NO STOP (as registered) -15.3% maxDD 22.7% halves -19.5%/+4.6% · crypto +-7% stop -15.2% · +-10% stop -13.4% · osc_prior>=median -6.7% · OSC_PRIOR>=75TH PCT +5.1% maxDD 9.2% halves -2.6%/+7.7% · 8 slots -17.5% · target +6% -8.6% · cut at 24h -6.5%. RETRACTED IN FULL: +11.9% (the registered expectation), +12.0%, +22.8%, +24.4%, +27.4%, +31.7%, +38.6%, +43.1%, the 12-of-12 random-draw comparison, and the entire slot sweep. ⭐ THE CORRECTED ENGINE PASSES A CONSISTENCY CHECK THE OLD ONE FAILED: crypto-all per-signal -0.488% -> predicted -8.2% -> engine -15.3%; osc>=75th +0.181% -> +2.0% -> +5.1%. Same sign, right magnitude. ⚠ ⛔⛔⛔ THE TELL WAS AVAILABLE FOR HOURS AND WAS NEVER CHECKED. This rule's per-signal expectancy measured -0.488% repeatedly during the same session. A rule losing half a percent per trade CANNOT return +12% over 235 trades unless the portfolio SELECTS a better subset — and it selects nothing, it just runs. The portfolio number must track the per-signal number and it never did. ⭐⭐⭐ STANDING CHECK, now in the pre-registration's pass line: ANY reported portfolio return must be reconciled against trades x ticket x per-signal-mean / capital BEFORE it is quoted. ⭐⭐ THE ROOT CAUSE WAS ARCHITECTURAL, NOT ARITHMETIC: both broken versions carried mutable scan state (pos['i']=j) advanced inside a loop over a timeline that was not the position's own. Precomputing each exit removes the possibility. ⭐ The rule as registered (D/E/F) is now EXPECTED TO LOSE and is left registered unchanged — a pre-registration with a negative expectation is still a valid test, and rewriting it to chase the one surviving cell is exactly the fitting it exists to prevent. Candidate H (osc>=75th) registered as a LEAD: +5.1%, negative first half, and the prior is substantially a volatility sort (top quartile 80% fill but 58% hit -7%, max down -11.75%). |
| every_decision_redecided 2026-09-04 · operator: 'do we need to go back and figure everything since all of the numbers moved?' Yes — every decision that used the broken capital simulator, re-decided on ops/sim_engine.py. | ⛔ MOST OF THE SESSION'S DECISIONS REVERSE. (1) PEGGED EXCLUSION REVERSED: all 159 symbols -11.6% BEATS crypto-only -15.3% — the FX/stablecoins were real ballast. (2) THE 48h CUT REVERSED: no cut -4.8% is BEST, cut-at-48h -15.3% is WORST (12h -12.8%, 24h -6.5%, 96h -12.2%). (3) THE TARGET: +3% is the WORST cell (-15.3%); +5% -11.2%, +1.5% -11.5%, +2% -11.8%, +8% -12.7%. 'Turnover beats size' is NOT supported. (4) SLOTS: MONOTONE, fewer is better — 1 slot -7.2%, 2 -11.8%, 3 -15.3%, 5 -16.1%, 8 -17.5%, 12 -21.8%. (5) THE STOP barely matters: -10% -13.4% best, no stop -15.3%, and EVERY stop variant has a NEGATIVE second half. (6) ⭐⭐⭐ THE BTC FILTER REVERSED AND IS THE LARGEST EFFECT: no filter -15.3% · skip if BTC bar < -1.0% -2.7% · < -0.5% +2.5% · < 0.0% +1.8% with maxDD 5.3% and BOTH HALVES POSITIVE (+0.3%/+1.6%). A 17.8-point swing on a filter dismissed hours earlier on the broken simulator's say-so. (7) THE PRIOR HOLDS: osc>=75th +5.1%. ⭐ BEST SURVIVING CONFIGURATION: capitulation spike + BTC bar >= 0.0% + osc_prior >= median, 3 slots, $100, +3% target, 48h cut = +5.2% over 120d, maxDD 4.6%, halves +2.3%/+2.9% — the most balanced pair produced all session. ⚠ ⭐⭐⭐ THE DEEPEST CORRECTION IS METHODOLOGICAL. I claimed FIVE separate times that per-signal and capital-constrained scoring 'disagree, and only the second is money.' THAT WAS ENTIRELY THE BUG. The per-signal numbers said this rule loses (-0.488%) and that the BTC filter helps (-0.513% -> +0.078%) — both now confirmed by the corrected engine. THE SIMPLE MEASUREMENT WAS RIGHT THE WHOLE TIME AND I OVERRODE IT REPEATEDLY WITH THE BROKEN ONE. ⭐⭐ Both surviving filters have INDEPENDENT per-signal support and were not invented to fit this table: BTC's bar was monotone against outcome in the clustering analysis, and osc_prior cleared its own day-blocked null at AUC 0.726. ⛔ +5.2% over 120 days is ~16% annualised, not the 38% quoted an hour earlier, and it remains in-sample and best-of-many. ⛔ DELIBERATELY NOT AMENDED INTO THE REGISTRATION A FOURTH TIME: the first three amendments were forced by a broken instrument, which is legitimate; a fourth, now the instrument works, would be fitting the registration to the answer just found. D/E/F stay registered EXPECTING TO LOSE. |
| execution_grid_excess_over_noise 2026-09-04 · operator: 'rebuild the execution layer as a grid on the corrected engine.' 11,760 configurations (universe x BTC gate x osc prior x target x stop x cut x slots x ticket) on ops/sim_engine.py, each one ALSO run on day-matched random entries. | THE CORRECTED ENGINE WAS STILL BEING READ AGAINST ZERO, AND THAT REVERSED A DECISION. Ranking cells on raw return in a window whose 2nd half is a bull market hands you the longest hold on the grid -- which is exactly how the 09-04 handoff reached 'no cut is best, the 48h cut is worst'. Against the SAME cell measured on day-matched random entries the entry effect is MONOTONE IN HOLD LENGTH: 12h +19.2pp, 24h +16.6, 48h +10.3, 120h +9.4, 15d +2.5, 30d +0.7, no cut +1.8. The spike's whole advantage lives in SHORT holds and is gone by two weeks. Also settled by the DIFF column: osc_prior ADDS NOTHING (lifts real -12.5->-2.6 and noise -22.8->-11.6, the same ~10pp -- a volatility/coin-quality sort that helps ANY entry); the BTC gate is MOSTLY MARKET TIMING (>=+0.5 has the SMALLEST entry effect, +3.4pp, because it helps the random entries just as much); +3% is the BEST target on entry effect (+14.5pp), restoring the choice the raw-return read had called 'the worst cell'; no stop wins on both readings (8th confirmation). ⚠ TWO DEFECTS IN THIS TOOL'S OWN FIRST RUN, both caught before reporting. (1) THE SLOTS AXIS WAS CONFOUNDED BY POSITION SIZE: at a fixed $100 ticket 1 slot deploys $100 of $1,404 and 8 deploys $800, so in a rule whose average trade loses, deploying less is automatically better -- the grid printed a clean monotone 'fewer slots is better' (-2.83% -> -6.35%) that was nothing but size. THE TELL WAS IN THE SAME TABLE: mean net PER TRADE moves the OTHER way, -0.851% at 1 slot to -0.537% at 8. Fixed by crossing slots with a ticket axis (fixed100 vs fund/slots). (2) THE NULL WAS TOO GENEROUS: 5 controls, best-real +4.47% against a luckiest-of-5 +4.20%, printed 'CLEARS its null' on a 0.27pp margin. Now 20-40 controls, a percentile, and a leave-one-control-out null on the excess grid. --prove-cache asserts the perf caches added to sim_engine change no field over 180 configurations. |
| hold_length_vs_matched_control 2026-09-04 · How long should the trade be held -- measured PER SIGNAL (no slots, no capital, no allocator) against a matched control at every length, because the grid's reading came from a portfolio and the standing rule is that a complex instrument contradicting a simple one is the suspect. | THE SPIKE BEATS A MATCHED RANDOM ENTRY AT EVERY HOLD FROM 4h TO 120h, AT PERCENTILE 100 OF 40 CONTROLS. Per-signal net, no gate: 4h -0.454 (noise -0.711), 8h -0.433 (-0.764), 12h -0.362 (-0.790), 24h -0.139 (-0.793, EDGE +0.654pp), 48h -0.488 (-0.935), 120h -0.913 (-1.214); 15d/30d/no-cut fall to percentile 72-80. THE 24h PEAK REPLICATES at targets of +2%, +3% and +5% and with the gate on. WITH THE GATE the rule is NET POSITIVE: BTC >= -0.5%, 24h, +3% -> +0.537% a trade after the full 0.681% round trip, edge +0.582pp, both halves positive (+0.636 / +0.534), n=378. Portfolio 3 slots x $100: +7.44%/121d, maxDD 5.3%, edge +13.77pp, percentile 100. ⚠ The 'no cut' row is the only one whose absolute return is positive and it has the SMALLEST edge -- it is the drift, not the rule, which is the same trap the Vault already recorded when Candidate D's headline collapsed under a hard 48h bound. The tool prints an explicit agreement check between the per-signal and portfolio instruments and says which to trust when they differ. |
| break_even_cost_and_maker_entry 2026-09-04 · The edge is worth most of the fee -- so state the BREAK-EVEN COST rather than a return, and test the one lever that would close the gap: a maker entry, since the exit is already a resting limit. | BREAK-EVEN ROUND TRIP 1.218%. Gross +1.218% a trade (1st half +0.782%, 2nd +1.502%), day-matched random +0.664%, EDGE +0.554pp BEATING 40 OF 40. Net +0.537% at the 0.681% taker cost. NOT OUTLIER-DRIVEN: median +3.681%, 5% trimmed mean +1.785%, dropping the five biggest winners leaves +1.185%, win rate 71.2%, positive on 74 of 99 days. THE DECOMPOSITION: random coin no gate -0.112% gross; + the BTC gate +0.664% (net -0.017%); + the spike +1.218% (net +0.537%) -- NEITHER COMPONENT REACHES THE 0.681% ROUND TRIP ALONE. ⛔ THE MAKER ENTRY FAILS DECISIVELY: every resting-bid arm is worse than buying the close (bid at the close 96% fill but gross/fill collapses 1.218% -> 0.341%; -1.0% bid, 65% fill, -0.158%; -2.0% bid, 45% fill, -0.666%) -- textbook adverse selection, monotone in bid depth. AND THE CAUSE IS TIMING, NOT PRICE: buying the close k bars later, same hold, same cost -- on time +1.317%, 4h late +0.923% (-0.394pp = 30% of the whole edge), 8h late +0.357%, 12h late +0.120%, 16h late -0.312%. THE EDGE IS GONE IN 16 HOURS. ⚠ ⛔ THE DAY-BLOCKED 95% CI ON THE GROSS MEAN IS +0.521%..+1.830%, i.e. NET -0.160%..+1.149% at the taker cost: THE GROSS EDGE IS ESTABLISHED, THE NET IS NOT, AT 95%, ON 99 DAYS. Two verdict lines in this tool's first run were wrong and were fixed: the CI was compared against the MAKER round trip while the headline was a TAKER rule (letting an interval containing zero print as a pass), and the maker verdict tested net > 0 instead of net > taker, printing 'A MAKER ENTRY TURNS THE RULE POSITIVE' for an arm 0.5pp WORSE than buying the close. ⭐ The timing result is 'the first leg IS the move' measured a further way, and it is why a maker entry cannot work here. |
| where_is_the_prize_entry_vs_exit 2026-09-04 · Operator: 'we now need an AI making the entry decision for certain and one eventually watching the exit... but the 3% profit exit does cover it so it's not as urgent as the AI entry bot, am I correct?' A sequencing question with a measurable answer: both AI layers are bounded above by a hindsight ceiling and both ceilings can be computed exactly. | THE EXIT HAS ~2x THE ROOM, AND THE PREMISE IS WRONG IN AN IMPORTANT WAY. Exit ceiling +4.943pp a trade on SELLABLE CLOSES (+8.784pp on bar highs -- wicks inflate it 43%, so closes are the honest number). Entry ceiling +2.364pp. ⭐ AND THE ENTRY'S CEILING IS STRUCTURALLY CAPPED BY THE EXIT: with a +3% limit the best any trade can do is +3.0%, so perfect entry selection tops out at exactly +3.0%/trade. AN ENTRY AI CAN NEVER BE WORTH MORE THAN +2.36pp WHILE THE EXIT CAPS AT +3%. Realistic sub-prizes: dropping only the 33 never-profitable signals adds +0.589pp (entry), while the 88 losers that WERE in profit (mean best +1.76%) and gave it back are worth +1.011pp (exit) -- the unambitious exit fix beats the entire realistic entry fix. 59.5% of trades fill; on sellable closes they then run a further +2.48pp median / +7.55pp mean, and 27% never close above the limit at all. THE TARGET SWEEP IS FLAT EITHER SIDE OF +3% (excess +10.2/+17.5/+16.0/+10.1 at +1.5/+3/+5/+8), so NO FIXED TARGET CAPTURES THE RUN -- the cleanest LAW #3 case the project has: a decision a filter provably cannot make, with a measured prize behind it. ⚠ ⭐⭐⭐ THE MINUTE TAPE CONFIRMS IT INDEPENDENTLY AND ADDS THE HARD PART. 36 filled cases with >=60min observed after the fill: peak after the fill median +4.23% (mean +8.45%), BUT where it ENDED median +0.26% (mean +1.50%), worst after the fill median -2.84% (p10 -14.28%), and 47% ENDED BELOW THE FILL PRICE. Minutes to the peak: median 335, p25 52, p75 1054 -- the peak has no home. ⇒ THE ROOM ABOVE +3% IS REAL AND NO FIXED LATER EXIT CAPTURES IT; holding is a coin flip that costs a -2.84% median drawdown. That is why the +3% limit is not broken and must not be loosened mechanically -- it converts a 47% coin flip into a certain +3% and is the reason the rule wins 71%. ⛔ ROOM IS NOT REACHABILITY: the 09-04 exit test measured a +4.54% ceiling on a different entry and NONE of nine live mechanisms reached any of it. |
| method | what it scored |
|---|---|
| net_flow_over_turnover 2026-08-30 · same, normalised by turnover | real, but it is a THINNESS TILT not a flow signal, and 5x underwater on cost (+0.118pp vs 0.681%) ⚠ the number was right and the story was wrong |
| btc_negative_stand_aside 2026-08-28 · stand aside when BTC is losing? | cleared -0.49% vs declined -0.06% at 4h - the gate's claim REVERSED, but INSIDE noise (block-level p=0.184) ⚠ ARMED on operator judgement. Unanswerable three times now. |
| strategy_lab_8 2026-08-28 · 8 rule sets scored on the same data | observed spread 1.51pp vs a 5.40pp noise scale - NO WINNER |
| holding_period 2026-08-25 · how long to hold | the edge got monotonically WORSE the longer held (-1.54% at 1h to -5.78% at 6h) ⚠ one day is one sample; this is why every decision is marked at 1/4/24/72h |
| mover_history_30d 2026-09-01 · the history of the big movers - what happens before and as they take off? | THE PATTERN IS REAL: movers show range 2.85x their own weekly normal (vs 1.61x), trade count 2.14x (vs 1.49x), turnover 2.23x (vs 1.78x) - while price is DOWN 1.03% over the prior 6h (vs +0.43%) and closes LOW in its bar. Activity doubles, price does not rise. AUC 0.58-0.60 on three independent activity measures. BUT on unconditioned buyable coin-hours the best is count_x at +0.88pp against a luckiest-of-8 null of 1.26pp - INSIDE. ⚠ the clearest pre-move description the project has found, and the first study run across a range of BTC regimes (3 days >= +3%, best +7.30%). It separates; it does not yet pay. ⚠️ The enriched case-control table reads +9.12pp and OUTSIDE the null - that is the enrichment artifact, caught for the second time in one day, and both tables are printed so it cannot be quoted alone. |
| vwap_gap_inverted 2026-09-01 · does a coin closing BELOW its period VWAP pay? | YES, INVERTED and it clears its null. Bottom decile (closed furthest BELOW VWAP) makes +1.00% over the next 4h vs -0.27% for the top decile - spread -1.27pp against a day-shuffled luckiest-of-8 null of 1.00pp, OUTSIDE. Offset-stable (0.28pp). Net of the 0.681% round trip: +0.32%. ⚠ ⭐ THE FIRST POSITIVE NET THAT CLEARS ITS NULL. But it is best-of-8 and it clears by 0.27pp, so it is a LEAD, not a finding - it needs its own pre-registered confirmation on data it was not found in. Direction is consistent with everything else: buy weakness. ⚠️ Its cousins turnover_x/count_x/range_x all FLIP on the offset guard despite AUC 0.60-0.62 - they separate the label and their money spread is not stable. |
| pullback_entry_sequence 2026-09-01 · THE OPERATOR'S OWN METHOD: run-up, pullback, buy the pullback with a resting limit, exit when it moves again | ALL 9 CELLS POSITIVE at the long horizon in the 30d window; best cell run>=10% / retrace 0.25 gives +2.04% NET at +12h (hourly) and +0.85% NET at +8h (4h bars, 2.6 trades/day) against a random-buyable control of +0.19%/+0.01%. ⭐ HOLD-OUT: split by time, 4h window gives +0.731% net (n=110) and +0.978% net (n=100) — HOLDS IN BOTH HALVES and stably. Offset-stable in both windows. ⚠ ⭐⭐ THE STRONGEST CANDIDATE THE PROJECT HAS. Still INSIDE a luckiest-of-9 null (+1.445% vs observed +0.849%), so it is not declared — but it survives a time-split hold-out, which nothing else has. Needs a PRE-REGISTERED forward test on data it was not found in. |
| pullback_backtest_outcome 2026-09-02 · if we run the three live candidates over the 120-day backtest, what is the outcome? | ⛔ SUPERSEDED IN PART — see note. Under ONE-POSITION-AT-A-TIME (a bound I invented, not a plan anyone proposed): A -59.0%, B +38.8%, C +309%. Under the OPERATOR'S ACTUAL MODEL (risk a % of equity per trade, capped by what the coin can absorb, as many concurrent as qualify) at 10% risk: A +22.9%, B +18.0%, C +36.7%. ⭐ A's SIGN REVERSES -- the capital model was doing the damage, not the signal. ⛔ BUT THE TAIL DEPENDENCE IS UNCHANGED: minus the best 3 trades A is +0.4%, B +0.0%, C +11.9%. A and B are entirely their outliers; only C survives, and C is also the only one whose MEDIAN trade is positive (+2.14% vs A's -0.48%). ⚠ ⭐⭐⭐ OPERATOR CORRECTION 2026-09-02: 'we shouldn't be trading one at a time. we trade as many as should be done that day. we don't trade the entire fund, we decide what the percentage of the total is worth risking then cap it at what the coin can handle.' Both numbers previously reported were wrong models: unconstrained assumes infinite capital, one-at-a-time is the worst possible pairing with an outlier-driven return. ⭐ AND COIN CAPACITY IS NOT BINDING AT THIS FUND SIZE -- median capacity $3,206 against a 10% ticket of $140. Cash only binds above ~33% risk per trade. |
| winners_vs_losers 2026-09-03 · compare winners vs LOSERS instead of winners vs average — given a big move is coming, can we tell which way? | ⭐ IT SEPARATES, AND IT IS THE FIRST DIRECTIONAL SIGNAL IN THIS PROJECT THAT IS NOT A VARIANCE TELL. Every price reading agrees: THE LOSERS HAD ALREADY GONE UP MORE. ret_4h AUC 0.294 (winners -0.10% vs losers +3.72%) · ret_12h 0.353 · ret_24h 0.350 (+0.94% vs +7.07%) · ret_48h 0.400 · ret_72h 0.394 · close_in_bar 0.324 (losers closed at 0.69 of their bar, winners 0.39) · pos_in_range 0.414. Best |AUC-0.5| 0.206 against a day-shuffled luckiest-of-15 null of 0.122 — OUTSIDE. Seven readings agreeing in one direction is a family, not a lucky column. ⛔ AND IT DOES NOT COMPOSE. The separation is conditioned on a big move HAVING happened, which is hindsight. As two live stages: stage 1 (top-variance moments) does work — top decile has 17% of bars gaining >=10% and 18% losing >=10% against 5%/6% baseline — but stage 2 (buy what has not run) makes it WORSE in 3 of 4 configurations: top-25% -0.17% -> half not-run -1.08% -> quarter -1.55%; top-10% +0.72% -> half +0.82% -> quarter -1.29%. ⚠ ⭐⭐⭐ THE SHARPEST STATEMENT THE PROJECT HAS: THE DIRECTION IS READABLE GIVEN A BIG MOVE; WE CANNOT GET TO 'GIVEN A BIG MOVE' WELL ENOUGH FOR IT TO PAY. Stage 1 tops out at ~17% big-move rate and the tilt does not help on the other 65%. ⭐ It also confirms 'buy weakness' for the seventh time, now in its cleanest form — variance held fixed, same coins on both sides, so it cannot be beta. ⛔ Do not act on it: it is in-sample and conditioned on hindsight. |
| capitulation_spike_entry 2026-09-03 · stay inside the entry/exit windows of the biggest tradeable wins — what is consistently happening there? | ⭐ THE PERFECT ENTRY IS ONE BAR AND IT IS A SPIKE. Aligned on it: volume 4.27x its own trailing normal, range 2.23x, closing in the bottom third (0.32), a slightly DOWN bar — while every neighbouring hour is ordinary (0.7-1.1x). The hour after reverses immediately (+0.64%, closing at 0.78 of its bar). ⭐ AND THE WINDOWS ARE WIDE: at 75% capture the ENTRY window is 36h median (2-50h) while the EXIT window is only 10h (1-37h) — entry precision is not the problem, exit precision is, which independently confirms the perfect-play decomposition. LIVE TEST: the signature (vol>=2x, range>=1.5x, close in the low 40%) fires on 6.54% of buyable bars; held 48h it returns +1.61% mean net against +0.94% for every buyable bar (+0.67pp), win 42% vs 37%, n=1,079 — and it BEATS A DAY-MATCHED RANDOM DRAW (null 95th pct +1.46% vs observed +1.61%). Lift on big moves grows with size: 1.21x at >=+15%, 1.46x at >=+30%, 2.19x at >=+50%. ⚠ ⭐⭐⭐ THE FIRST LIVE-COMPUTABLE SIGNAL IN THIS PROJECT TO BEAT PICKING AT RANDOM ON THE SAME DAYS — which is the control that killed every previous candidate. ⛔ But it clears by only 0.15pp, its MEDIAN is still -0.54% (the edge is entirely in the tail), and the full signature appears at only 37% of the 30 perfect entries (volume>=2x alone appears at 77%). PRE-REGISTERED, not acted on: ops/PREREG-capitulation-entry-2026-09-03.md, clock starts 2026-09-04T00:00Z, judged at >=300 bars/40 days. ⚠️ close_in_bar is partly definitional; vol 4.27x is not. |
| leg_start_signature_three_readings 2026-09-03 · operator: 'now we have a real target, it's the jumps in small slices... go through each data point and comb through all of the key points... do that with 3 of the data points and do it for the entire history.' What is true in the minutes BEFORE an up-leg that turns out to be BIG (>=3%), against a trough that produces nothing (<1.5%)? | ⭐ PACE CLEARS ITS NULL AND NOTHING ELSE DOES. PACE (trades/min vs the coin's own normal): AUC 0.530/0.562/0.569/0.568 at 5/15/30/60min — best PACE_30 at |0.069| against a day-blocked luckiest-of-12 null of 0.040, OUTSIDE. Big legs run at 1.77x normal pace beforehand, dead legs 1.39x. Money translation monotone across all five cells: mean leg 2.32/2.37/2.43/2.55/2.91%, big-leg rate 20.7/20.5/22.1/22.2/28.6%. Stable: positive on 9 of 10 days, median AUC 0.597, range 0.499..0.642. PEOPLE (buy_count share) 0.533-0.557 and SIZE (large_trade_share) 0.539-0.553 lean the same way but weaker. ⛔⛔ AND IT IS THE TENTH VARIANCE TELL. Before big DOWN legs pace reads 1.728 vs 1.460 dead, AUC 0.542 — nearly as strong as up. Asked directly, among ALL turning points does high pace pick UP over DOWN: 48.7% / 52.2% / 50.4% / 53.3% by pace quartile — a 4.6pp tilt that is not even monotone. PACE FORECASTS THE SIZE OF THE NEXT MOVE AND NOT ITS DIRECTION. ⚠ ⭐⭐⭐ THE TENTH TIME THIS PROJECT HAS FOUND A VARIANCE TELL — but for the FIRST time the target may not need direction. The old target was 'will this coin go up' and magnitude-without-direction was useless. The new target is an oscillator you buy after a drop and sell after a rise: there, MAGNITUDE IS THE EDGE and direction comes from the oscillation. That reframing is untested and must not be assumed. ⛔ The leg sizes here are the ZIGZAG's — hindsight troughs. A real entry does not get the trough, and no money test has been run. ⚠️ The two most recent days are the weakest (09-02 AUC 0.499, 09-03 0.516) — watch for decay rather than assuming noise. |
| candidate_parameter_holdout 2026-09-04 · The rebuild's parameters were chosen by looking at all 121 days. Does choosing them on the past carry into the future? | ⛔ NO. Ranked on first-half EXCESS over the matched controls, only 2 OF 10 top cells kept a positive excess in the second half: real -0.285pp against a noise band of -0.013pp, PERCENTILE 20 -- inside the null. The two halves disagree about hold length. ⭐ AND YET THE WHOLE SURFACE IS POSITIVE: 144 OF 144 BOUNDED-CUT CELLS have positive excess (+0.19pp to +1.54pp), smooth and monotone in every direction, with the only non-positive cells being the unbounded holds. ⇒ THE PLATEAU IS REAL; ITS MAXIMUM IS NOISE. That is why Candidate K registers the CENTRE of the plateau and not the peak (BTC>=0.0/+3%/30d, +1.54pp, deliberately not registered). ⚠ ⛔ THE FIRST RUN OF THIS FILE RANKED ON RAW RETURN AND IT CHOSE 'no cut' FOR NINE OF ITS TEN PICKS -- an unbounded hold in a rising sample is leverage on the drift, the controls did the identical thing and BEAT the real set (+2.858% vs +2.479%), and the whole test read 'inside the noise band' while measuring nothing but who held longest. Ranking on the SAME cell's excess over its day-matched controls removes the drift from both the choosing and the scoring. It then still failed, which is the honest answer. ⚠️ The 144 cells are nested subsets of the same signals, not 144 independent tests -- the count describes the SHAPE; the p-value-like statement is the 40/40. |
| entry_decider_ai_built 2026-09-04 · Operator: 'yes build the AI for the entry point.' LAW #3's decider: at the capitulation spike's CLOSE, TAKE / SKIP / UNSURE plus a 0-1 conviction, recorded with reasoning and stamped with the mechanical outcome afterwards. It decides nothing real. | ⛔ NO EDGE IN SAMPLE, AND THE INSTRUMENT SAYS SO CLEANLY. AI-TAKE +2.323% vs the INCUMBENT (take every signal) +2.034% = +0.288pp -- but against a MATCHED RANDOM DECIDER with the same accept rate on the same days (4,000 draws) it sits at PERCENTILE 68.8, inside the null, on only 19 TAKEs (the verdict line is suppressed below 40). The CONVICTION ranking, split at the model's OWN median: above +2.164% vs below +1.846%, gap +0.317pp, PERCENTILE 75.8 of a shuffled null -- also inside. And the decision categories are not ordered: SKIP (+2.095%) beats UNSURE (+1.745%). ⭐ The reasoning is substantive and independently rediscovers the project's own fake-vs-real finding ('not actually a capitulation -- coin is +10.4% over the prior 24h', 'parabolic prior run-up', 'failed blow-off top after a rally') -- but good reasoning that does not order outcomes is still no edge. ⚠ ⛔⛔ THE PROMPT IS FROZEN HERE AND MUST NOT BE TUNED. The outcomes have now been seen; any prompt change from this point is fitting, not building. A prompt fingerprint (sha256 of system+user+model) is recorded on EVERY decision and --score refuses to pool versions unless they are DECLARED equivalent in code with the reason -- e44be46e0102 and 8758291f9b08 differ only in a JSON-shape instruction the CLI transport needs, which is output plumbing and not a changed question. ⛔ DESIGN DEFECT CAUGHT AND FIXED MID-BUILD: the first prompt returned 5 SKIP / 1 UNSURE / 0 TAKE on its first six signals -- an unscoreable degenerate arm, because with no TAKEs there is nothing to compare against the incumbent. Fixed by making CONVICTION (0-1) the primary output, which cannot degenerate, and by making the decision framing symmetric ('the average signal is mildly PROFITABLE, so ordinary is not a reason to skip'). ⛔ AND shadow_decider.py, which has claimed since 2026-09-04 to be 'an AI call on every signal', CONTAINS NO MODEL CALL -- grep for anthropic/claude/api_key returns nothing. It assembles context and records it and nothing asks. 9th instance of the present-but-unwritten pattern and the most expensive kind: a file everyone believes is an AI layer, running hourly on cron, producing a growing file of unanswered questions. |
| method | what it scored |
|---|---|
| leverage_dial 2026-08-31 · is coin choice a beta dial turned by the timing call? | S1 PASS r=+0.650; S2 13.6%; S3 FAILS p=1.000 - nothing forecasts the board ⚠ the dial has nothing to multiply |
| book_intent 2026-09-02 · does the RESTING book lead price? | not yet run - the recorder fix lands 2026-09-02 00:05 UTC ⚠ the only column we record that is intent rather than executed history |
| hot_btc_regime — · does ANY current method pay on a BTC >= +3% day? | cannot be scored - the regime has never been in the tape ⚠ ⭐ THE OPERATOR'S STANDING HYPOTHESIS. 1 day in 13; most recent 2026-08-21, four days before recording began. Every row in this registry inherits this limitation. |
| forward_suite 2026-09-01 · run a whole suite of variants live, or just the 2 registered candidates? | BOTH, in two tiers that never mix. Only the 2 registered candidates are judged at the pass line; the family is scored ONLY against 4 predictions fixed in advance (short holds beat hold12; plain holds beat targets; run-up 6-18 beats 3-4; deep retrace beats shallow). A family prediction cannot be met by one lucky cell, so MORE variants make that evidence stronger — the opposite of the selection problem that makes picking a winner from N worthless. ⚠ ⭐ the operator asked for the suite; the two-tier split is what makes it buy knowledge rather than noise. ⛔ Caught before the clock started: candidate B used 0.60 for BOTH the setup and the entry, which is not what the matrix measured (676 setups vs 248) and scored -0.111% in-sample. Corrected with zero forward bars recorded. |
| jump_entry_needs_a_hot_day 2026-09-03 · enter EVERY +5% jump immediately, portfolio-sized — does it pay, and does it depend on BTC's day? | ⭐⭐⭐ THE OPERATOR'S STANDING THESIS, CONFIRMED ACROSS TWO INDEPENDENT WINDOWS, AND HE CALLED IT BEFORE THE TEST EXISTED. 30d hourly, hold 48h: all days +60% — but EXCLUDING 2026-08-19/20/21 it is +2% (mean -0.19%, median -4.44%, win 33%), while those three days alone give mean +10.50%, median +5.99%, WIN 78%. The top 3 days contribute 104% of all positive return-points. 120d 4h REPLICATION, monotone across four regimes: BTC>=+3% n=55 mean +6.47% median +1.95% win 60% · BTC +1..+3% mean -1.51% · BTC -1..+1% mean -3.88% win 30% · BTC<-1% mean -3.82% win 29%. Portfolio at 10% risk over 120 days: EVERY signal $582 (-59%) · HOT DAYS ONLY $1,561 (+11%) · everything else $509 (-64%). ⚠ ⭐⭐⭐ THIS RECONCILES THE WHOLE PROJECT. Earlier work found big movers exist on EVERY day (>=+5% hit rate flat at 95-100% across regimes) and I reported that as evidence against the operator. Both are true and they are different questions: THE OPPORTUNITIES EXIST EVERY DAY; THE STRATEGY ONLY PAYS ON HOT DAYS. ⭐ The hot-day condition is knowable in real time (BTC's day-so-far) and the hot-day median is POSITIVE (+1.95%) — almost nothing else in this project has a positive median. ⭐ And it was predicted a priori: the operator stated it repeatedly before any of this analysis existed, which makes it a real prediction rather than a data-mined split. ⛔ 6 hot days / 55 signals is small, +11% over 120 days is modest, and it is in-sample. Needs pre-registration. ⭐ PRE-REGISTERED 2026-09-03: ops/PREREG-hotday-jump-2026-09-03.md, clock starts 2026-09-04T00:00Z, judged at >=60 signals across >=4 hot days (~2-3 months). ⭐ THE GATE WAS MADE TRADEABLE FIRST: the backtest used BTC's FULL-DAY move (hindsight); the live gate is BTC's move from the UTC day open TO NOW, and it survives almost intact (68 signals, +9.65% mean, +4.80% median, 76% win against the hindsight version's 74/+10.50%/+5.99%/78%). The +3% level is registered because it is the number the operator named, NOT because it scored best. |
| board_pressure_gate 2026-09-03 · the gate is supposed to be 'is the WHOLE BOARD being bought' and BTC>=+3% was only ever a proxy for it. Measured directly from our own buy_vol/sell_vol tape, is board buying (a) broader on a hot day, (b) the same thing as the BTC gate, (c) worth money as a gate on the +2%/1h jump entry? | (a) DESCRIPTIVE ONLY, n=1 day: 09-03 is the broadest-bought day of the ten (breadth 0.538 vs a cold range 0.459-0.511). At the BUCKET level, with the gate ON at that moment: dollar-weighted buy share 0.597 vs 0.488 off, while breadth moves only 0.537 vs 0.492 -- the MONEY moves much more than the count of coins. 7 ON buckets. (b) NO: corr(breadth, BTC-from-open) = +0.233 pooled, +0.203 within-day; dollar-weighted +0.200 / +0.163. Board buying is NOT the BTC gate under another name at 15min resolution. (c) UNANSWERABLE TODAY AND THE TOOL SAYS SO: every gate cell is negative net (breadth quartiles -0.93 to -1.06%, dollar-wt -0.25 to -1.47%, BTC quartiles -0.50 to -1.31%), all three top-quartile spreads INSIDE a day-blocked luckiest-of-3 null of +1.136pp. **AT THE OPERATOR'S OWN BAR: BTC >= +3% -> n = 0.** No scored event in the history of our tape has ever fired with the gate ON; the gate crossed at 14:30Z and a jump needs 4h of forward tape. ⚠ ⭐⭐⭐ THE QUARTILE HID THE QUESTION. The top BTC quartile on this tape starts at +0.919% -- nowhere near 'an active day' -- so a quartile split reports on a gate that was never on. Fixed bars at the operator's own numbers now print alongside. ⭐⭐ AND THE DAY IS THE WRONG UNIT: 09-03 was under +1% until 12:00Z and did not cross +3% until 14:30Z, so averaging its cold morning into its hot afternoon repeats the read-the-day-by-its-close error. The gate is a MOMENT and it is knowable live. hot_day_watch now captures gate_on{first_cross_at, minutes_at_or_above} + operator_bar_windows per day (backfilled 10/10, origin wired). ⏱ First scoreable hot-gated event completes ~18:30Z 2026-09-03; cron runs hourly. ⛔ DISCOVERY ONLY -- the pre-registered hot-day clock starts 09-04T00:00Z and none of this may retune it. |
| gate_within_day 2026-09-03 · the hot-day claim has only ever been tested BETWEEN days (6 hot vs ~90 cold, of which 2 lost and 3 were one consecutive run). 2026-09-03 was cold until 12:00Z and crossed +3% at 14:30Z — so does the SAME POP behave differently before and after the gate turned, inside ONE day, with coins/liquidity/book/tape all held constant? | PARTIAL AND IT TURNED AGAINST THE GATE AS THE SAMPLE FILLED. Tracked live as events completed: 1h hot-half edge went +0.40pp (n=29) -> +0.26 (n=33) -> +0.06 (n=42) -> +0.22 (n=45) — it moves by more than its own size. At 2h the sign REVERSES: hot +0.20% net 41% win (n=41) vs cold +0.50% 54% (n=63) = -0.30pp. 4h hot half still n=0 (completes ~18:30Z). ⛔⭐⭐⭐ THE CONFOUND, AND IT IS THE REAL RESULT: the gate is read at ENTRY but the return accrues over the next 4h, so a 'cold' pop at 11:00Z exits INSIDE the run. Splitting the cold half by where it EXITED: into the hot window +1.49% (1h, n=9) / +3.06% (2h, n=12) / +4.56% (4h, n=15); fully before it -0.12% / -0.10% / -0.46% (n=54/51/42). ⭐⭐⭐ ON THIS DAY THE MONEY WAS IN BEING POSITIONED BEFORE THE GATE TURNED AND HOLDING THROUGH IT (+4.56%), NOT IN ENTERING ONCE IT WAS ON (+0.20% at 2h). ⛔ Buy share in the hour AFTER the pop is LOWER in the hot half and FELL as the sample grew: 0.450 -> 0.442 -> 0.431 vs a steady cold 0.476, both under 0.5. ⚠ ⭐⭐⭐ THE FIRST WITHIN-DAY TEST THE PROJECT HAS EVER RUN, and its first real output is a challenge to the finding it was built to confirm. 'A coin that pops on a hot day has a following wind' is measured with the gate at ENTRY; today says the following wind belongs to positions ALREADY OPEN when the gate turns. That is the project's own standing lesson one level up — the first leg IS the move (run_anatomy), reacting to the leg loses (leg_latency) — now appearing at the VENUE level rather than the coin level. ⚠️ RECONCILE BEFORE BELIEVING EITHER: the 120d OHLC hot-day result (+6.47%) labels a day by its CLOSE, so many of its winning entries were cold-at-the-time trades that ran into the afternoon — possibly the same effect wearing the gate's clothes. ⚠️ The exit-window split is partly definitional (anything held through a board-wide run gains) — but the HOT-ENTRY events were also inside that run and earned far less, and that is the comparison that is not definitional. ⛔ ONE DAY, n=15 in the load-bearing cell. An observation, not a test. DISCOVERY — the pre-registered hot-day clock starts 09-04T00:00Z and none of this may retune it. |
| exit_mechanism_test 2026-09-03 · the money test said the exit was the leak (66% time-stopped at -2.01%). So which exit MECHANISM — not which parameter — recovers it? Entry held completely fixed; nine different KINDS of exit decision plus a hindsight ceiling. | ⛔⭐⭐⭐ EVERY LIVE MECHANISM LOSES AND THE BEST ONE IS THE CLOCK I HAD CALLED THE LEAK. CLOCK_90 -0.204% · TARGET_3 -0.227% · CLOCK_240 -0.346% · PACE_FADE -0.389% · EOD -0.439% · FLOW_FLIP -0.434% · TURN -0.538% · TRAIL_3 -0.699% · TRAIL_2 -0.702%. Best live is INSIDE a day-blocked luckiest-of-9 null of +0.433%. ⭐⭐⭐ AND THE CEILING IS ENORMOUS: PERFECT exit = +4.536% mean, 86% win, positive on 10 of 10 days. THE ENTRY IS NOT THE PROBLEM — 86% of these positions are in profit at some moment and average +4.54% at their peak. ⭐⭐⭐ WHY NO RULE FINDS IT: the peak is spread across the ENTIRE window. p10 8min · p25 28 · MEDIAN 119 · p75 300 · p90 399 · cap 480. Within 30min 25% · 30-90min 18% · 90-240min 24% · AFTER 240min 33%. There is no time at which the peak tends to sit, so no clock and no trail can locate it. ⚠ ⛔⭐⭐⭐ THIS CORRECTS MY OWN EARLIER CLAIM. I reported 'the exit is the leak, not the entry' from the money test's 66%-time-stopped diagnostic. Held against eight alternatives, the 90-minute clock is the BEST of them — the loss is not caused by choosing a bad exit rule, it is that NO RULE OF THIS KIND CAN LOCATE THE PEAK. The leak is real; the diagnosis of it was wrong. ⭐⭐ TRAILING STOPS ARE THE WORST TWO ARMS (-0.70%), confirming 'stops lose' for the sixth independent time. ⭐⭐ AND THE OPERATOR'S OWN CONSTANT, used exactly as the buy/sell study said it should be (watched during the trade, not used to select), exits at -0.434% — no better than doing nothing. Tested rather than assumed. ⭐⭐⭐ THIS IS THE CLEANEST LAW-#3 CASE THE PROJECT HAS PRODUCED: a decision a filter provably cannot make, with a measured 4.7pp prize sitting behind it, and the operator predicted it — 'I still think there is going to be an intelligent way to make the final decision on the entry and exit.' ⛔ UNTESTED AND THE OBVIOUS GAP: every mechanism here is ALL-OR-NOTHING. Scaling out — taking profit in pieces, which is what the operator actually described — is a different mechanism class and has never been run. |
| armed_flow_exit 2026-09-04 · operator: 'the reversal is an exit trigger if it gets past the fade. so it's the same signal but means something different at different times.' Every prior test listened from bar 1, so during the fade window it heard dying trades as exit signals. Arm the reversal exit ONLY after the trade proves it survived — by time, by profit, or by sustained buy pressure. | ⭐⭐⭐ THE DISTINCTION IS REAL AND IT IS PROFIT, NOT TIME. Arming on TIME makes it WORSE than unarmed in all four variants: TIME_2 -0.911 · TIME_4 -0.872 · TIME_6 -0.898 · TIME_12 -0.935, against UNARMED -0.717. Arming on PROFIT transforms it: UP_2 -0.375% (53% win, median +0.310%) · UP_3 -0.085% (52% win, median +0.269%) · UP2_AND_TIME4 -0.488% (54% win, median +0.384%). WIN RATE GOES 19% -> 52% AND THE MEDIAN FLIPS FROM -0.716% TO +0.269%. All three profit-armed variants agree, which is stronger than any single cell. HELD_3 (buy share held 3 consecutive hours) -0.582%, in between. ⛔ BUT THE MEAN STILL DOES NOT CLEAR: best is UP_3 at -0.085%, only +0.06pp better than doing nothing, INSIDE a day-matched null of -0.075%, and it is best-of-8. ⭐ THE REASON IS IN THE LAST COLUMN: UP_3 ARMS ON ONLY 48% OF TRADES. The other half never reach +3%, stay deaf, and bleed to the 48h cap. ⚠ ⭐⭐⭐ THE OPERATOR'S FRAMING WAS RIGHT AND MORE SPECIFIC THAN HE STATED IT: the same reading means different things at different points in a trade, and the thing that separates the two regimes is PROFIT, not elapsed time. A trade up 3% has proven it is not a fade; a trade that has merely survived four hours has proven nothing — and time-arming is WORSE than no arming at all, which is a real result in its own right. ⭐⭐ The distribution genuinely changed (win 19%->52%, median -0.716%->+0.269%); the expectancy did not. ⇒ THE ARMED AND NEVER-ARMED TRADES ARE TWO DIFFERENT POPULATIONS AND ONE RULE IS STILL BEING APPLIED TO BOTH. That is exactly trade_taxonomy (5 kinds, best hold 1 to 12 bars, wanting opposite handling) reappearing at the exit. The next build is a SPLIT rule, not a better threshold. 🔭 Still parked: 'distance from buyer to seller' (operator 09-04) — spread or imbalance magnitude, ASK WHICH. |
| split_rule_profit_window_decay 2026-09-04 · operator: 'build the split rule... figure out the profit window and what says it is over... I would assume the amount of potential profit is how long it lasts. a coin only does this for so long then it moves to another coin so once we get to doing multiple legs we will see less and less profit after the first few big hits.' Three tests: the armed/unarmed split rule, the profit window, and the decay hypothesis. | A. THE SPLIT RULE FAILS AND ITS FAILURE IS THE FINDING. Cutting the never-armed trades makes it monotonically WORSE: cut 2h -0.827% · 4h -0.853% · 6h -0.834% · 8h -0.747% · 12h -0.664% · NO CUT -0.085%. B. THE REASON: TIME TO ARM (+3%) IS A MEDIAN OF 13 HOURS (p25 5h, p75 27h). Cutting at 2-12h does not remove dead trades, it kills trades before they can prove themselves. ⭐⭐⭐ AND THE SPLIT ITSELF IS THE BIGGEST SEPARATION THIS PROJECT HAS PRODUCED: ARMED 296 (48%) mean +3.446%, WIN 93% · UNARMED 324 (52%) mean -3.311%, win 14%. Time to peak median 14.5h; the reversal fires median 16h and lands AFTER the peak in 71% of armed trades — IT IS A CONFIRMATION, NOT A FORECAST — capturing a median 85% of the peak. C. ⛔ THE DECAY HYPOTHESIS IS UNSUPPORTED IN EITHER DIRECTION. The raw table appeared to show later legs BIGGER (leg 1 4.88% rising to leg 8+ 8.85%, gaps shrinking 215m -> 10m) but that is pure SELECTION — only the hottest coins ever reach leg 8. Within the same coin-day, paired: mean(first 2 legs) - mean(rest) = +0.63pp, median +0.52pp, first-two bigger in 20 of 35 coin-days (57%), against a two-sided sign-flip null of [-1.65, +1.54] — INSIDE. ⚠ ⭐⭐⭐ THE ARMED/UNARMED SPLIT IS THE STRONGEST SEPARATION IN 70 METHODS (+3.446% at 93% win vs -3.311% at 14%) AND IT IS UNUSABLE AS BUILT, because it declares itself at a median of 13 hours. The whole question now collapses to one thing: IS THERE ANYTHING AT ENTRY, OR IN THE FIRST FEW HOURS, THAT PREDICTS WHICH TRADES WILL ARM? That is the fake-vs-real question again but with a far cleaner label and on the one entry that clears its null. ⭐ 'What says it is over' is ANSWERED: the buyer-to-seller reversal, and it is a confirmation that arrives after the peak 71% of the time while still capturing 85% of it — good enough to use, useless as a forecast. ⛔ 14th BROKEN CHECK: my decay null was ONE-SIDED and printed 'INSIDE the null' for an observed -2.90pp against a p95 of +0.77pp — a result strongly significant in the OTHER direction. A one-sided test cannot see a reversal, and it reported the wrong verdict confidently. Replaced with a two-sided paired sign-flip test. 🔭 Still parked: 'distance from buyer to seller' — spread or imbalance magnitude, ASK WHICH. |
| after_the_3pct_still_running_or_dead 2026-09-04 · operator: 'so what we really need is a standard we are in for 3% then we have to figure out what is the signal after 3% that it is still running or dead.' At the moment the +3% limit fills, is there anything left, and does anything at that moment say whether to ride or bank? | ⭐⭐⭐ THERE IS A GREAT DEAL LEFT AND IT IS NOT REACHABLE BY HOLDING. After the fill, the PEAK over the next 48h averages +11.26% (median +5.21%): 62.0% go on to another +3%, 50.8% to another +5%, 36.1% to another +10%. But fixed holds capture almost none of it — +12h mean +0.87% (median -0.25%), +24h +1.24% (-0.39%), +48h +1.92% (-0.33%), all with NEGATIVE medians. ⭐⭐⭐ AND THE SIGNAL IS WHERE THE FILL BAR CLOSED. Running vs dead at the fill moment: close_in_bar 0.737 vs 0.356 (AUC 0.762) · run past the trigger 4.095% vs 2.400% (AUC 0.786) · time-to-fill 7 bars vs 21 (AUC 0.417). ⭐⭐ AND IT PAYS, WHILE RIDING BLINDLY DOES NOT: bank every time +3.000% · ALWAYS ride +2.315% (second leg -0.685%) · ride if close_in >= 0.5 +3.438% (second leg +0.637%, 69% of fills) · >= 0.6 +3.404% · >= 0.7 +3.393%. Three thresholds on a plateau, +0.438pp over banking. Day-blocked null on the second leg: p95 -0.047% vs observed +0.637% — OUTSIDE. ⛔ TIME-TO-FILL FAILS ON MONEY DESPITE ITS AUC: riding the fast fills returns -0.994% and total 2.562%, WORSE than banking. ⛔ AND THE HOLD-OUT FLIPS SIGN: first half -0.421% (n=227), second half +1.315% (n=354). ⚠ ⭐⭐⭐ THE OPERATOR'S FRAMING WAS EXACTLY RIGHT AND IT DECOMPOSES THE PROBLEM PROPERLY: bank a standard +3%, then ask a SECOND, separate question about whether to re-enter the same move. That converts an unanswerable 'when is the top' into an answerable 'is this bar strong or weak', and the answer is the simplest reading on the bar — WHERE IT CLOSED IN ITS OWN RANGE. ⭐⭐ THE LOAD-BEARING NUMBER IS THAT ALWAYS-RIDING LOSES (-0.685%) WHILE FILTERED RIDING GAINS (+0.637%): the filter is doing the work, not the second leg existing. ⛔ NOT ESTABLISHED — the hold-out flips sign, so this is a lead on 844 in-sample fills. It must be pre-registered alongside the base rule, not bolted on afterwards. ⛔ time-to-fill is the 13th reading to separate a label (AUC 0.417) and lose money (-0.994%). |
| stop_loss_under_capital_and_the_ordering_bias 2026-09-04 · operator: 'we have to figure out how to do a proper stop loss.' Every prior stop test in this project was PER SIGNAL and lost (six ways). Tonight established that the 48h clock-cut LOSES per signal and WINS under capital constraints — so a drawdown stop had never been tested the way that matters. | ⭐⭐⭐ UNDER CAPITAL CONSTRAINTS A STOP LOOKED TRANSFORMATIVE: no stop 82 trades +12.0% maxDD 4.8% · -3% stop 349 trades +43.1% maxDD 1.5% (ret/DD 29.3, halves 25.4%/22.5%) · -5% +30.2% · -7% +23.7% · -10% +18.1%. Monotone to -1% (+54.0%), and it survived 3% slippage (-3% still +38.0%). The mechanism is TURNOVER — 349 trades vs 82 — the stop's real job is refusing to let a dead position squat in a slot. ⛔⛔ AND IT WAS LARGELY AN ARTEFACT OF INTRA-BAR ORDERING. The code checked the TARGET before the STOP on bars containing both. Assume the opposite: -1% goes +48.7% -> -27.7% (195 ambiguous bars), -2% +45.4% -> -6.4%, -3% +40.2% -> -0.4% (76 bars), -5% +28.4% -> +12.1%, -7% +22.8% -> +12.9% (12 bars), -10% +17.1% -> +10.9% (6 bars). THE TIGHTER THE STOP THE MORE THE RESULT RESTS ON AN UNJUSTIFIABLE ASSUMPTION. ⭐⭐⭐ RESOLVED ON THE MINUTE TAPE, NO ASSUMPTION: at a -3% stop the STOP FIRES FIRST 25 TIMES AGAINST THE TARGET'S 24 — a coin flip. TRUE per-signal net: -3% -0.399% · -5% -0.422% · -7% -0.113% · -10% -0.042%, sitting at 44-50% of the way from the pessimistic to the optimistic bound at every level. Applying that midpoint calibration to the 120-day sim: -7% ~= +18%, -10% ~= +14%, against +12.0% with no stop. ⚠ ⭐⭐⭐ A STOP DOES HELP, BY ROUGHLY +6pp, NOT THE +31pp FIRST REPORTED. A -7% stop is the defensible version: only 12 ambiguous bars, both bounds positive, and the true path measured rather than assumed. The tight stops (-1% to -3%) are unmeasurable on 4h data and must not be used. ⛔⭐⭐ THE METHODOLOGICAL FINDING IS BIGGER THAN THE TRADING ONE: any backtest on 4h bars that uses BOTH a target and a stop is silently choosing which one fires, and the choice is worth up to 76pp. Every such result in this project's history is suspect unless the ordering was resolved. ⭐ It also inverts the standing 'stops lose' doctrine correctly: stops lose PER SIGNAL and help under CAPITAL CONSTRAINTS, for the same reason the clock-cut does — a freed slot is worth more than a recovery. That is now the fifth time per-signal and capital-constrained scoring have disagreed in one session. ⛔ The minute tape starts 2026-08-25, so only 61 signals could be resolved. It settles the ORDERING question, not the size of the effect; that accumulates forward. |
| where_the_false_runs_first_dip 2026-09-04 · operator: 'I bet the false runs are dipping and the real runners are staying strong. where do the false runners start to dip in that time frame? where do they first show they are not going to hold? we might be able to dump them earlier and not use a stop loss per say.' | ⭐⭐⭐ HIS HYPOTHESIS IS CONFIRMED AND THE TIMING IS SPECIFIC. Winners vs losers, median price and separation: 15min +0.09%/-0.05%, AUC 0.497 — EXACT CHANCE, indistinguishable · 30min +0.37%/+0.21%, AUC 0.597 · 60min AUC 0.650 · 90min +0.28%/-0.02%, AUC 0.690 · 240min AUC 0.700 · 480min +1.23%/-0.10%, AUC 0.706. Losers already below -2%: 4% at 30min, 22% at 90min, 33% at 180min, 48% at 480min. ⭐⭐ THE WINDOW MATCHES buy_sell_pressure's INDEPENDENT FINDING — 'the split only opens 15-60 minutes after entry' — measured there on BUY SHARE, here on PRICE. Two different readings, same clock. ⭐ ACTING ON IT: dump at 60min if below -0.5% gives +0.319% vs a +0.155% baseline (+0.164pp), -1.0% +0.279%, -2.0% +0.303% — all three thresholds at 60min improve. 90min is mixed. EVERY 120min VARIANT IS MATERIALLY WORSE (-0.407 to -1.203pp) — by two hours the information is priced and you sell after the damage. ⛔ 'Dump below 0.0%' is the worst cell at every check time (-0.413pp at 60min): it throws out winners that are merely resting. ⛔ A -7% stop makes it WORSE here (-0.079% vs +0.155%), the opposite of the 4h capital-constrained result — different sample, per-signal framing, read nothing into either. ⚠ ⭐⭐⭐ THE DIAGNOSIS IS SOLID AND THE TREATMENT IS NOT YET MEASURABLE. n=60 signals over 10 days; a +0.16pp improvement on sixty trades cannot be established. What IS established is the WINDOW: they are identical at 15 minutes (AUC 0.497) and clearly parted by 60-90 (0.650-0.690), confirmed by two independent measurements. ⭐ THAT IS THE FIRST TIME THIS PROJECT HAS LOCATED A DIVERGENCE IN TIME rather than just failing to find one — six other angles today (entry depth, survival, drawdown, buy-share confirmation, wait-and-hold, danger-window survivorship) all came back invariant. The difference is that those asked a BINARY question at the end; this traced the PATH. ⭐ 60 MINUTES IS THE MOMENT TO LOOK, AND THE THRESHOLD MUST SIT BELOW THE NOISE (-0.5% to -2%, never 0.0%). ⛔ Needs more minute tape — it starts 2026-08-25 and accumulates from here. |
| candidate_k_registered 2026-09-04 · Register the rebuild's result as one more candidate -- NOT as an amendment. D/E/F/G/H/I/J stay exactly as they are and keep expecting to lose. | CANDIDATE K: capitulation spike (vol >=2x AND range >=1.5x AND close in the lowest 40%), non-pegged, bar traded >= $224,616, AND BTC's OWN 240min bar >= -0.5% at the signal's close; buy that close as a TAKER; resting limit at +3.681% gross (+3.0% net); otherwise market at the close of the 6th bar (24h); NO STOP; 3 slots x $100. In-sample +0.537% a trade. THE REGISTRATION STATES ITS EXPECTATION IN ADVANCE: roughly HALF the in-sample number forward (+0.2% to +0.3%) counts as the rule working, anything at or above the in-sample number should be treated as suspicious, and a result inside +/-0.15% of zero is the single most likely outcome -- which would mean the edge is real but does not clear the fee, making the COST the next thing to work on. ⚠ WIRED, NOT JUST WRITTEN: collect_candidate_k() in prereg_scorer.py scores it on the 240min cache (its entry, gate and cut are all defined on 4h bars, so scoring it hourly would be a different rule wearing its name), with the coverage and not-advancing guards, and settleable reported separately from fired. ⛔ AND WIRING IT UP FOUND A PRE-EXISTING CRASH THAT WOULD HAVE KILLED THE WHOLE LANE: portfolio_sim.simulate() reads t['in'] and prereg_scorer.collect() never wrote it, so the HOT-DAY JUMP arm -- the only one with portfolio=True -- raises KeyError the moment it has enough settled signals to be judged, and that crash lands BEFORE candidates D/E/K are scored. One hot day after the clock started would have taken the registration-scoring lane down while the log looked like a normal run. 8th instance of the present-but-unwritten field pattern; fixed at the capture point. ⛔ No order is placed by any of this. |
| candidate_kf_registered 2026-09-04 · Operator ruling 2026-09-04: 'run K and K-filtered side by side from the same clock.' This resolves the conflict left open in §8 -- K enters at the spike's close with no wait, while his standing ruling of the same day says 'at this point we always use the flash-vs-real signal for entry.' Neither candidate is amended into the other. | DIAGNOSTIC ONLY, n=10 TAKEN OF 33, DECIDES NOTHING. K-M -0.659% (58% win), K-W -0.514%, K-S +1.176% (lookahead), K-F +1.070% (70% win). K-F - K-M = +1.729pp, decomposed as SELECTION +1.835pp (lookahead) / TIMING -0.106pp / waiting alone +0.146pp. Entry slip from waiting only +0.18% mean, against the +1.53% measured on run-up triggers -- a capitulation spike closes at its LOW and does not keep running in the next 45 minutes, so the wait is far cheaper on THIS entry than on the one the filter was measured against. THE FILTER TAKES ONLY 30% of K's signals here (vs 56% on run-up triggers), so K-F needs roughly 3x as long as K to reach 250 trades: K ~2026-11-24, K-F ~2027-02. That is a real cost of the filter, stated now rather than discovered later. ⚠ ⛔⛔ THE TAIL CLAIM DOES NOT HOLD YET, AND THE CHECK THAT CAUGHT IT IS THE POINT. Loss rate 42.4% -> 30.0%, mean loser -5.05% -> -2.34%, worst -26.77% -> -3.28% -- and the first version of this scorer printed 'the filter cut the tail' off exactly that. THE WORST OF n DRAWS IS NOT COMPARABLE ACROSS DIFFERENT n: a filter taking 10 of 33 trades has fewer chances to hit the tail even if it selects nothing at all. Size-matched (4,000 draws of 10 from K-M): worst is -7.66% median, 5-95% range -26.77%..-3.18%, and K-F's -3.28% sits at PERCENTILE 94.7 -- INSIDE what random selection of the same size gives. On this tape the tail benefit is NOT established; it may appear at n=250, which is what the clock is for. TWO MORE REPORTING DEFECTS FIXED IN THE SAME PASS: report() printed '23 still need 48h of tape' about 23 signals the filter had DECLINED (settled, scored, deliberately not taken -- a reader would conclude the arm was starved of data when it was doing its job), and the pre-clock run printed NOTHING for K-F at all, so a wired candidate was indistinguishable from an unwired one. |
| method | what it scored |
|---|---|
| crash_stand_aside 2026-08-27 · stand aside on a -10% BTC day? | fires correctly, holds through the bottom and the dead-cat bounce, releases 13h later ⚠ BUILT and red-proofed but it GATES NOTHING - nothing consumes the posture |
| maker_side_cost — · does entering as a maker put the surface above the cost line? | at 0.681% taker: 0 of 256 cells positive. At 0.320% maker: 8 positive, best +0.299%. At 0.160%: 42 positive. ⚠ a resting maker order may never fill, and 'waiting loses to the trades it skips' is already measured. A sensitivity, not a plan. |
| swing_anatomy_persistence 2026-09-03 · operator: 'if useless is up 65% there were probably 10 different times we could have bought and sold that coin and made money.' Treat a coin-day as a PATH, not one trade: how much up-movement actually existed, in how many legs — and does a coin that oscillates in the morning keep oscillating in the afternoon? | ⭐⭐⭐ HE UNDERCALLED IT. USELESS on 09-03: 13 up-legs totalling +116.3% against a buy-and-hold of +52.9% — the move was available MORE THAN TWICE OVER, and all 13 legs cleared the 0.681% round trip. ⭐⭐⭐ AND THE DAY'S % GAIN DOES NOT IDENTIFY THE TRADEABLE COINS: AKE finished +2.8% (peak +7.0%) with 22 up-legs totalling +141.2% — the MOST harvestable movement on the board — while ZEC +18.0%, ENA +17.8%, ADA +10.3% were ONE leg each, pure drift with nothing to trade repeatedly. ⭐⭐⭐ PERSISTENCE: rank coins by morning sum-of-up-legs and the afternoon ranking holds — spearman +0.764 on 09-03, and POSITIVE ON 10 OF 10 DAYS (median +0.757, range +0.455..+0.907), including days the board fell -4.99% and -3.04%. Survives the outlier check: dropping AKE+USELESS moves pearson 0.953->0.644 but spearman only 0.764->0.725; dropping the top 5, spearman +0.660. Median afternoon movement: top-half-by-morning +8.60% vs bottom-half +1.90% (excl. the two outliers, +6.40% vs +2.48%). ⭐ CONCENTRATION: top 1 coin carries 22.5% of the board's harvestable movement, top 5 = 56.1%, top 10 = 70.5%. ⚠ ⛔⛔ HINDSIGHT ZIGZAG — the legs turn at exactly the right minute and nobody can do that. This is a CEILING and a SHAPE, not an achievable return, and NO strategy has been tested against money here. What is NOT hindsight is the PERSISTENCE: the morning ranking is knowable at noon and the afternoon is what it predicts. ⭐⭐⭐ WHY THIS MATTERS MORE THAN ANY GATE: it works on 10 of 10 days regardless of BTC, so it does not require waiting for the 2.3 hot days a month the operator's bar implies. ⚠️ MUST BE CHECKED AGAINST THE 8/31 TRAP before being believed: realised vol looked this strong and turned out to be BETA (beta 2.76). Sum-of-UP-legs is direction-free path length rather than return, which is a real difference — but the check has not been run. ⚠️ Also unrun: whether a real entry/exit can capture any of it, and the true spread on the thin coins that dominate the top of the list. |
| spike_entry_plus_3pct_limit 2026-09-04 · operator: 'why don't we just exit now with this signal at +3 and round trip. see what that does then we will build from there if we get a positive response.' The win-only limit exit had been tested on a momentum detector I invented and on random bars — never on the capitulation spike, the one entry that clears its own null. | ⭐⭐⭐ POSITIVE, AND IT SURVIVES EVERY CONTROL RUN AGAINST IT. 120d/4h: fill 79.4%, median 36h, expectancy +1.132% vs a RANDOM-BAR control of +0.697% (+0.435pp). 30d/hourly: fill 74.7%, median 27h, expectancy +1.526% vs random +0.934% (+0.592pp). HOLD-OUT KEEPS THE SIGN IN BOTH HALVES ON BOTH DATASETS: 4h +0.467%/+1.796%, hourly +2.376%/+0.676%. With >=120 bars of room the fill rate is 79-82%. ⭐⭐⭐ THE CAPITAL SIMULATION — the exact test that killed the +5% version — IS POSITIVE IN ALL EIGHT CONFIGURATIONS: 120d/4h 3 slots +24.4% (25 opened, 22 filled) · 5 slots +17.7% · 8 slots +20.0% · 12 slots +17.0%; 30d/hourly +8.1% / +10.6% / +12.3% / +14.4%. The +5% version by contrast opened only 7-31 trades in 120 days for -1.6%/+15.6%/-1.2%. ⭐⭐⭐ AND IT IS NOT THE MARKET'S DRIFT: against 12 random-entry draws at each slot count the spike wins 12/12 EVERY TIME (48 of 48), and its return exceeds the MAXIMUM of the random range in all four — random medians are NEGATIVE (-5.2 / -8.4 / -3.7 / -2.6%). ⚠ ⭐⭐⭐ THE FIRST COMPLETE RULE IN THIS PROJECT THAT BEATS ITS CONTROLS ON MONEY. Entry: the capitulation spike (clears its null twice, independently). Exit: a resting +3%-over-cost sell, no stop, no clock. Both halves of both datasets keep the sign; the capital sim is positive in 8 of 8; random entries with the identical exit LOSE money. ⛔ WHAT IS NOT ESTABLISHED, and it is not small: (1) THIS IS IN-SAMPLE on 120 days — the spike parameters (vol>=2x, range>=1.5x, close<=0.40) were themselves chosen on 09-03 from this same family of data, so a pre-registered FORWARD test is the only thing that can settle it; (2) THE BAG — every configuration ends with 100% of capital committed and zero cash, and the worst single unfilled position is -86.0%, so the headline return includes unrealised marks; (3) the 30d hourly window is the tail of the 120d window, so the two datasets are NOT independent; (4) only 12 random seeds per cell; (5) entry slippage beyond the round-trip cost is not modelled. ⇒ NEXT: PRE-REGISTER IT before anything else, exactly as the pullback candidates were on 09-01. |
| the_reread_after_3pct 2026-09-04 · operator: 'the 3% is solid for us and then we do a re-read of the coin to see if it's time to jump or time to hold, right? what did you come up with for that factor and what happens when we test it?' Nine candidate readings at the fill moment, judged by HOLD-OUT CONSISTENCY rather than in-sample AUC. | ⭐⭐⭐ ONLY TWO OF NINE SURVIVE BOTH HALVES, AND THEY ARE THE SAME IDEA: bar_ret (the fill bar's own open-to-close move) 1st half +3.014% / 2nd +4.050% / all +3.613%, and run past the trigger +3.059% / +3.968% / +3.584%. Everything else fails a half: close_in 2.854%/3.761% (⛔ THE FACTOR I REPORTED AN HOUR EARLIER — it fails the first half), vwap_gap 2.824%, range_x 2.214%, vol_x 2.089%, prev_close_in 2.169%, two_strong 1.085%, ttf 1.576%. Banking every time = 3.000%. ⭐⭐⭐ AND THE FILL-BAR RETURN IS PERFECTLY MONOTONE ACROSS FIVE QUINTILES, IN THE FULL SAMPLE AND IN EACH HALF INDEPENDENTLY. Second leg by quintile: -4.170% / -1.932% / -0.650% / +0.976% / +2.287%. First half: -5.272 / -2.678 / -2.440 / +0.048 / +1.466. Second half: -3.347 / -1.211 / +0.597 / +1.735 / +2.690. Ride only the top 40%: second leg +1.634%, total +3.656% vs +3.000% banking (+0.656pp), against a day-blocked null of p95 +0.261% — OUTSIDE. ⚠ ⭐⭐⭐ THE RE-READ IS: DID THE BAR THAT PAID YOU CLOSE STRONG, OR DID IT SPIKE TO YOUR LIMIT AND GET SOLD BACK? A bar closing up >3% keeps running; one closing under +1.6% is finished, and riding it costs -4.17%. That is the operator's 'still running or dead' answered in its own terms, and it is readable at the instant the limit fills. ⭐⭐ MONOTONE IN FIVE QUINTILES AND IN BOTH HALVES SEPARATELY is the strongest structural evidence this project has produced — earlier candidates were monotone in one dimension and died on the hold-out. ⭐ JUDGING BY HOLD-OUT INSTEAD OF IN-SAMPLE AUC IS WHAT FOUND IT, and it immediately killed my own close_in result from an hour before. ⛔ STILL IN-SAMPLE ON 120 DAYS — the base rule and this re-read must be PRE-REGISTERED TOGETHER; bolting the re-read on afterwards is what makes a clean result unfalsifiable. ⛔ bar_ret and run are near-duplicates (both measure bar strength) so this is ONE finding, not two. |
| move_the_buy_to_the_divergence_point 2026-09-04 · operator: 'would we move the buy decision to that point and only buy the runners?' The winners and losers are indistinguishable at 15min (AUC 0.497) and parted by 60-90min — so move the ENTRY to the moment they separate instead of buying the spike close and dumping the weak later. | ⭐⭐⭐ 60 MINUTES IS UNIQUELY GOOD AND IT IS EXACTLY THE DIVERGENCE POINT. Baseline (buy at the spike close) +0.167% per signal, 63% win. WAIT 60min: above -0.5% buys 48 of 60 and returns +0.914% per bought / +0.731% PER SIGNAL at 67% win — 4x the baseline while keeping 80% of signals; above -2.0% +0.690%, -1.0% +0.637%, 0.0% +0.317%, +0.5% +0.364%. ALL FIVE THRESHOLDS AT 60MIN BEAT THE BASELINE. WAIT 30min: +0.039 to +0.472%, mixed and weaker. WAIT 90min: ALL FIVE NEGATIVE, -0.028% to -0.292%. ⚠ ⭐⭐⭐ IT PEAKS PRECISELY WHERE TWO INDEPENDENT MEASUREMENTS SAID THE POPULATIONS PART — price divergence (AUC 0.497 at 15min rising to 0.690 at 90min) and buy_sell_pressure's finding on BUY SHARE ('the split only opens 15-60 minutes after entry'). That coherence is real evidence. ⛔⛔ AND THE SHARPNESS IS ALSO WHAT NOISE LOOKS LIKE: a clean flip from uniformly-better at 60min to uniformly-worse at 90min on n=60. REGISTERED AS CANDIDATE G, NOT SWAPPED INTO THE FROZEN ENTRY — changing the entry on sixty samples is the exact error this project keeps documenting, and running it alongside costs nothing. ⚠️ G NEEDS MINUTE DATA, which the 4h OHLC cache cannot supply, so it is scored on the recorder tape only and its sample grows far more slowly than D/E/F. ⭐ If it holds it is the largest single improvement found: +0.731% vs +0.167% per signal. |
| pre_spike_oscillation_prior 2026-09-04 · operator: 'let's do an investigation into the pre-spike data. is there something that happened there that will be something that AI can use?' Fourteen readings of the 24 bars (4 days) BEFORE the capitulation spike, winners vs losers. | ⭐⭐⭐ THE STRONGEST READING OF THE ENTIRE SESSION AND IT CLEARS ITS NULL BY MORE THAN DOUBLE. osc_prior — the sum of the coin's up-legs (3% zigzag) over the 24 bars BEFORE the spike — reads WINNERS 15.44% vs LOSERS 5.04%, AUC 0.726, against a day-blocked luckiest-of-14 null of 0.102: |0.226| OUTSIDE. The whole pre-spike family is coherent: ret_pre_24h 0.648 (+4.08% vs +0.17%), ret_pre_48h 0.648 (+5.01% vs -0.39%), vol_trend 0.647 (1.77x vs 1.27x), ret_pre_12h 0.644, ret_pre_96h 0.642, range_trend 0.619, pos_in_24bar_range 0.612 (0.676 vs 0.452), down_bars_in_a_row 0.419 INVERTED (winners 0, losers 1). ⭐⭐⭐ AND AS A LIVE FILTER UNDER CAPITAL IT IS MONOTONE: no filter 82 trades +12.0% maxDD 4.8% (halves 6.7/10.9) · 25th pct 97 trades +13.7% · MEDIAN 200 trades +31.7% maxDD 4.2% (9.6/20.6) · 75th PCT 190 trades +38.6% maxDD 0.4% (halves 21.7%/18.3%). ⚠ ⭐⭐⭐ IT IS THE OSCILLATION PERSISTENCE ARRIVING AS A PRE-SPIKE PRIOR. The 09-04 finding (a coin's recent oscillation predicts its future oscillation, pooled spearman +0.728, positive 10 of 10 days, stronger in the liquid half) now explains WHICH capitulation spikes bounce: A SPIKE IN A COIN THAT HAS BEEN OSCILLATING BOUNCES; THE SAME SPIKE IN A DEAD COIN DOES NOT. Winners had 3x the prior up-leg movement. ⭐⭐ THE WHOLE FAMILY SAYS THE SAME THING: winners were RISING before the spike (+4.1% over 24h, +5.0% over 48h) while losers were flat or falling with a run of down bars. It is THE DIP IN A STRONG COIN, not the dip in a dying one — which refines 'buy weakness' rather than contradicting it. ⭐⭐⭐ AND THE TRADE COUNT RISES AS THE FILTER TIGHTENS (82 -> 200 -> 190): filtering to oscillating coins makes trades FILL FASTER, which frees slots faster, which produces MORE trades. Quality drives turnover — the same mechanism as the 48h cut and the stop, now on the entry side. ⛔ In-sample on 120 days, best-of-4 (but monotone across all four), and the zigzag logic was found earlier the same day on the same data — a pre-registered forward test is required before this is believed. |
| method | what it scored |
|---|---|
| rank_climb_trigger 2026-08-29 · does a rank climb predict a run? | every cell negative; trigger-as-run -0.47% vs random -0.18% ⚠ NEGATIVE selection value - the harder it selects the more it destroys |
| old_trigger_climb30_plus5 2026-09-01 · the ORIGINAL system's entry: climb >=30 places AND coin up >=5% | +1h -2.72% (25% win) · +4h -3.16% (21% win) · +24h -2.93%; money-floor only -1.83% ⚠ the worst entry measured. Demanding MORE of the move before acting makes it worse - the same monotone the latency sweep found. ⚠️ The operator's standing objection applies in full: this has never been scored on a hot BTC day. |
| dip_entry_prejump_base 2026-08-27 · wait for the price to return to its pre-jump base? | base-0.5% -> +1.06%/65% win; but LOSES to the trades it skips in 4 of 5 BTC regimes ⚠ the coins that never come back down are the best ones; every waiting rule declines them |
| runner_detector 2026-08-28 · can a never-return runner be detected up front? | run_up_pct AUC 0.709 vs a LABEL, 0.548 vs money; separation vanished out of sample (27% vs 28%) ⚠ the definitional trap, first instance |
| thermometers 2026-08-27 · is the tape hot or average? | AUC 0.46-0.49 |
| buildup_before_moves 2026-08-29 · does participation build before a move? | best AUC 0.497; no column builds; the one difference found was definitional and flat on money ⚠ our data records EXECUTED trades. Build-up in intent would be invisible in it. |
| net_flow_selection_dollars 2026-08-30 · does the largest net inflow coin run? | top decile negative in 12/12; dollar rank is a SIZE rank |
| realised_vol_selection 2026-08-31 · does realised vol pick the coin? | REVERSES to -0.365pp; it is BETA (2.76). Market-neutral is not beta-neutral. |
| btc_direction_readings 2026-09-01 · does any BTC reading forecast the board? | best of six -0.250pp against a luckiest-of-six null of 1.203pp. corr(BTC 1h, board NOW) +0.503, with breadth +0.78 - it IS the board under another name. |
| leg_latency_detector 2026-09-01 · if we stop predicting and just REACT, is anything left? | ALL 16 NEGATIVE NET. Best +0.052% gross = -0.629% net, adding +0.061pp over a random buyable pick. Monotone: the more of the move you require, the more you lose. ⚠ the constraint is COST, not selection and not detector speed |
| flush_watchlist 2026-09-01 · is the pre-run flush a watchlist? | a run follows 0.49% of flushes vs 1.45% of non-flushes - LIFT 0.34x. 1 turned, 100 kept falling. ⚠ P(flush|runner) high, P(runner|flush) BELOW base rate. A base-rate error I had recommended acting on. |
| reversal_sweep_256 2026-09-01 · reverse every part of the entry, 1 at a time and in every mixture | ZERO of 256 net positive. Best (-0.062%) INSIDE the luckiest-of-256 null (+0.320%). No single reversal helps. Reversing the TAPE knob costs 1.30pp. Gross spans 1.701pp - WIDER than the 0.681% round trip. ⚠ the entry choice matters; the whole surface sits under the cost line |
| exit_rules 2026-08-27 · stops and trailing exits | EVERY stop/trailing rule loses; mean drawdown -3.29%; best exit tested is simply hold 24h |
| big_movers_need_a_hot_btc_day 2026-09-01 · do the big movers only show up on a hot BTC day? | NO, but PARTLY. 120 days / 365 buyable movers / 6 BTC days >= +3%. Reaching >=+5% in 4h by BTC quartile: 95% / 99% / 97% / 100% from BTC -6.51% to +7.30% - the opportunity is present on essentially EVERY day. But size does tilt: movers on BTC >= +3% days made +25.61% against ~+17.2% overall. So 'it does not work without an active day' is NOT supported; 'it works BIGGER on an active day' has weak support (24 events). ⚠ ⭐ the 30-day run said the biggest movers were on the WORST BTC days; at 120 days that reverses to a mild tilt the other way. The 30-day read was too small. What survives both: the >=+5% hit rate is flat and high in every regime. |
| five_percent_target_exit 2026-09-01 · 'if we win by 5% at first we can expand the win' — does taking +5% beat holding? | NO. Holding beats the +5% target in both windows: 4h bars +1.53% (hold 8h) vs +0.02% (target); hourly +2.73% (hold 12h) vs +1.65% (target). The trailing exit is worst of the three. Beyond ~8-12h the hold decays and goes negative. ⚠ ⭐ the +5% target CAPS the winners, and the winners are where the money is. Consistent with 'every stop/trailing rule loses'. The answer to the exit question so far is: hold roughly 8-12 hours, take neither a target nor a trail. |
| setup_density 2026-09-03 · is 'an active day' really about HOW MANY setups fire, rather than what BTC did? | LOOKED STRIKING, DIED ON INSPECTION. Bucketed by setups-per-day the means are monotone — 1 setup +0.051%, 2-3 +0.404%, 4-9 +1.329%, 10+ +4.004% (62% win, median +1.49%, the only bucket with a positive median). ⛔ BUT the 10+ bucket is TWO DAYS and one of them (08-21, BTC +7.30%) carries it at +8.03% while the other (06-15) is -0.42%. And INSIDE a flat BTC band (-1.5%..+1.5%) the effect REVERSES: 1-3 setups +0.463%, 4+ setups -0.134%. ⚠ ⭐ corr(setups per day, BTC's day) is only +0.282, so density is NOT simply a BTC proxy — but it is not a signal either. The monotone was one day. |
| board_breadth_selection 2026-09-03 · among concurrent setups, prefer the ones firing when board breadth is LOW? | NO. Mean spread -2.69pp (lowest-breadth quartile +2.654% vs highest -0.037%) looked like the 'buy the weak tape' finding again — but the MEDIAN GOES THE OTHER WAY (-0.53% vs +0.33%) and the win rate is HIGHER in high breadth (52% vs 44%). Day-shuffled null +/-4.75pp against an observed 2.69pp: INSIDE. ⚠ ⭐ a mean-only reading in this dataset is almost always the tail talking. Check the median first. |
| wider_stops 2026-09-03 · our stops were 3% and 5% and the median LOSER's MAE is -7.19% — does a wider stop work? | NO, AND MONOTONE. no stop +0.990% · stop 15% -0.744% · stop 10% -1.164% · stop 8% -1.172% · stop 5% -1.498% · stop 3% -1.043%. Every stop loses to no stop, and the wider the stop the less damage it does — i.e. the best stop is no stop. A 3% stop fires on 137 of 210 trades. ⚠ sixth independent confirmation that stops lose here. |
| breadth_separates_but_does_not_pay 2026-09-03 · board breadth separates SPIKE-FADE from BUILDER at AUC 0.724 — does that convert into a better exit? | THE SEPARATION IS REAL AND SURVIVES EVERY GUARD: AUC 0.724 (BUILDER median breadth 66.8 vs SPIKE-FADE 42.5), day-shuffled 95% band [0.351,0.660] so OUTSIDE the null, and corr(day's mean breadth, day's mean return) is only +0.058 so it is NOT a day/beta effect. ⛔ AND IT DOES NOT CONVERT. Routing low-breadth trades to hold1 and high-breadth to hold8 LOSES at every split tested: best +0.76% against +0.99% for a plain fixed hold2, i.e. -0.23pp, INSIDE a day-shuffled null of +0.78pp. ⭐ WHY: the 2-class separation is on 82 of 213 trades (39%); the rule also routes DIP-RUN (25%, wants hold 12) and DEAD (33%) and wrecks them. A 5-class problem is not solved by a 2-class separator. ⛔ And breadth does NOT predict death: 30% dead in both the low and high thirds. ⚠ ⭐⭐⭐ THE PRE-BUILD ANSWER, AND IT IS A GATE-A FINDING. The trades are genuinely different kinds, but NOTHING WE CURRENTLY RECORD AT ENTRY converts that into a better exit — the single best separator we have loses money as a rule. An AI layer reading today's columns would be guessing. The blocker is not the decision layer, it is the INPUT: we do not capture enough at decision time. ⭐ The obvious new input exists as of 2026-09-02 — depth_imbalance is valid for the first time and has never been on a journal row. |
| perfect_trade_tells 2026-09-03 · work backwards — find the perfect trades, then ask what was a tell BEFORE they happened | STRONG TELLS FOUND, AND THEY ARE ALL VARIANCE TELLS. Against time-matched controls the perfect trades were already running: ret_48h AUC 0.792 (median +11.79% vs +0.04%), count_x 0.776 (2.76x normal), ret_72h 0.762, vol_x_4h 0.734, range_x 0.717, and bars_quiet 0.152 (median 0 bars since its last 10% move, against 59 for controls). Day-shuffled luckiest-of-15 null reaches 0.121; observed 0.348 — OUTSIDE the null. ⛔⛔ THEN THE DECISIVE CHECK: the SAME readings mark the 40 WORST trades equally or better — ret_48h 0.761 best vs 0.764 worst · vol_x_4h 0.750 vs 0.775 · count_x 0.775 vs 0.770 · range_x 0.720 vs 0.766 · bars_quiet 0.237 vs 0.153. And the worst trades had run up MORE beforehand than the best (median ret_48h +34.22% against +11.79%). Same coins too: best-40 is ESPORTS x8 / AKE x6, worst-40 is ESPORTS x10 / AKE x3. ⚠ ⭐⭐⭐ EVERY TELL THAT PRECEDES A PERFECT TRADE EQUALLY PRECEDES A DISASTER. These readings select VARIANCE, not profit — the 8/31 beta lesson in a new form, and the strongest version of it yet because the AUCs are high AND outside their null. ⭐ It also explains the rest of the project at a stroke: why the strategy is outlier-driven, why the median trade loses while the mean is positive, and why targets and stops BOTH fail — we are selecting high-variance moments, and looking backwards only shows the up tail. ⛔ Working backwards from winners cannot distinguish a good entry from a violent one; the comparison must always be winners vs LOSERS, not winners vs average. |
| before_the_jump 2026-09-03 · what happened BEFORE the jump — and does it separate the jumps that ran from the ones that died? | ⛔ NOTHING SEPARATES THEM. A jump continues only 34% of the time. Every reading of the prior 24h leans the same way but weakly: squeeze AUC 0.610 · count_x 0.602 · vol_x_24h 0.600 · range_x 0.593 · ret_24h 0.591 · vol_x_6h 0.577 · pos_in_24h 0.570 · ret_6h 0.496. Best |AUC-0.5| is 0.110 against a shuffled luckiest-of-9 null of 0.129 — INSIDE. The pre-jump state of a runner and of a failure are the same state. ⚠ ⭐⭐ THE ANSWER TO 'WHAT HAPPENED BEFORE THAT JUMP' IS: THE SAME THING THAT HAPPENED BEFORE THE JUMPS THAT DIED. All nine readings lean toward the runners (more volume, more trades, wider range, more prior gain) which is the variance signature again — it precedes both outcomes. ⚠️ 52 vs 103 is a small sample and every reading leans the SAME way, so this is 'not separable at this sample' rather than 'proven identical'; it would be worth re-running when the forward window has widened it. ⭐ The load-bearing number is the base rate: TWO THIRDS OF +5% JUMPS DIE. |
| dispersion_and_bigmover_gates 2026-09-03 · is the trigger the winner-loser SPREAD (money rotating from losers to winners), or simply that a big mover exists that day? | ⛔ BOTH FAIL, AND BOTH FAIL IN THE SAME DIRECTION. Winner-loser spread as a gate: no gate -1.28% · >=26pp (median) -0.97% · >=40pp (p75) -1.75% · >=49pp (p90) -5.42% — A WIDER SPREAD IS MONOTONICALLY WORSE. 'A big mover exists today' as a gate: no gate -1.94% · a coin already +25% today -2.36% (WORSE than no gate) · +25% mover AND BTC cold -4.88% win 31% · +25% mover AND BTC hot +5.45% · BTC hot with NO mover condition +9.17% win 65%. Adding the mover condition to the BTC gate REDUCES it. ⭐ And corr(winner-loser spread, BTC so far) = -0.040, so the spread is genuinely an independent reading — it just does not help. 38 of 103 days have a +25% buyable mover; only 6 have BTC >= +3%. ⚠ ⭐⭐⭐ READ AS PEOPLE, THIS IS THE MOST USEFUL NEGATIVE YET. A wide winner-loser spread, or one coin flying while the rest bleed, is a FRAGMENTED market -- one crowd piling into one thing. That is not money rotating from losers into winners; it is an isolated pump, and chasing into it LOSES (-4.88%, 31% win). BTC >= +3% is a proxy for the opposite: EVERYONE IS BUYING EVERYTHING. ⭐ So the hot-day finding is not 'more opportunity on hot days' -- it is 'a coin that pops on a hot day has a following wind, and the same pop on a cold day gets sold into.' That also explains why a big mover EXISTING is worthless: what matters is whether the rest of the board is being bought alongside it. |
| oscillation_money_test 2026-09-03 · THE decisive test: does a real entry and exit — no hindsight anywhere — capture any of the oscillation persistence (+0.728 spearman, 10/10 days)? | ⛔ IT FAILS. TOP by oscillation: 500 trades, mean net -0.423%, 43% win, POSITIVE ON 1 OF 10 DAYS, INSIDE the resampled null (p95 -0.232%). THE TWO-COMPARISON RULE: beats BOTTOM-ranked by +0.388pp (so the ranking does point the right way) but LOSES TO A RANDOM BUYABLE COIN by -0.201pp. And it loses to doing nothing: the same positions simply held over the identical window returned -0.350% against the rule's -0.423% — the trading made it worse. Random's held version was +0.138%. ⭐⭐⭐ THE DIAGNOSTIC IS THE VALUE: target hit 147 trades (29%) mean +3.250% · TIME-STOPPED 331 trades (66%) mean -2.010% · end-of-day 22 (4%) -1.073%. When the entry finds a leg it pays +3.25%; two thirds of the time the 90-minute clock realises a -2% loss on a position that was still open. ⚠ ⭐⭐⭐ THE PERSISTENCE IS REAL AS A DESCRIPTION AND DOES NOT SURVIVE CONTACT WITH AN EXIT. That is the same wall every candidate in this project has hit, and the diagnostic says WHERE: the entry works 29% of the time at +3.25%, and the exit converts the rest into -2.01%. It independently reproduces the 09-03 morning finding from three other measurements — THE EXIT IS THE LEAK, NOT THE ENTRY. ⛔ This is ONE rule with parameters fixed in advance. A grid search over hold/target would find a positive cell and it would be the luckiest-of-N trap the reversal sweep already documented (0 of 256 cells positive at the real round trip). The honest next move is a different EXIT MECHANISM, not a better hold number. ⛔ A random buyable coin beating the ranked one is the load-bearing failure — the ranking is not adding tradeable value at this entry. |
| new_factor_hunt_six_unread 2026-09-03 · operator: 'create tests against new factors that we haven't looked at yet... looking at the real change places and find the factors that are indicating the coming run or turn.' Six readings the project has NEVER used, at every turning point, for both directions. | ⛔ ALL SIX FAIL. At a TROUGH, big bounce vs dead: FLOW_SPLIT 0.466 · DEPTH_TILT 0.467 · TRADE_GAP 0.477 · SIZE_SHIFT 0.516 · VWAP_GAP 0.516 · DEPTH_PULL 0.503 — best |AUC-0.5| = 0.034 against a day-blocked luckiest-of-6 null of 0.037. NOTHING CLEARS. At a PEAK, big drop vs dead: best 0.024, nothing clears. Inside PACE terciles none works in more than 1 of 3 and the signs are inconsistent — no incremental information over the one reading that did clear. ⛔⭐ THE ONE BIG NUMBER IS AN ARTEFACT: VWAP_GAP separates trough from peak at AUC 0.211, but price sitting below its own recent vwap IS what a trough is. Matching on the size of the preceding leg does not fix it because the preceding leg's DIRECTION sets the sign. Reported as definitional, not as a direction signal. ⚠ ⭐⭐ A CLEAN, WIDE NEGATIVE AND IT IS WORTH HAVING: the microstructure we record — order-book size, resting tilt, trade-size shift, the silence between trades, and the people-vs-dollars split nobody had ever computed — does NOT mark a turn. Combined with the pace result, that is 3 + 6 = nine distinct readings tested at the turning points, of which exactly one clears its null and that one forecasts SIZE not DIRECTION. ⭐ The people-vs-dollars split (FLOW_SPLIT) was the most promising idea on paper — retail absorbing distribution vs size accumulating quietly — and it reads 0.466/0.510. It is now tested rather than assumed. ⛔ DEPTH_TILT had only 959 valid troughs (book valid from 09-02 only); it deserves a re-run once the valid-book window is weeks rather than days. |
| round_trip_sequential_test 2026-09-03 · operator caught a design error: 'we are getting in and out at different points to capture the run ups and leave before the run down. is that how you did the test?' It was NOT — exit_mechanism_test locked each coin out for 480min after one entry, so it scored ONE round trip per coin per 8h against a coin (USELESS) that produced THIRTEEN legs in eighteen hours. Redone as a sequential trader: every exit re-arms the entry immediately. | ⛔⭐⭐⭐ THE ROUND TRIPS HAVE NO GROSS EDGE AT ALL, AND COST IS THE ENTIRE LOSS. GROSS per coin-day across eight very different mechanisms: -0.371 / -0.482 / -0.317 / -0.532 / -1.104 / +0.386 / -0.463 / +0.383 %. Essentially zero, in both signs. COST per coin-day: 1.18% to 8.78%. NET: TARGET_3 -1.664 (1.7 trips) · TARGET_2 -1.833 (2.1) · CLOCK_90 -3.941 (5.1) · TRAIL_2 -5.604 (6.6) · CLOCK_30 -6.157 (9.6) · TARGET_OR_TRAIL -6.740 (9.4) · FLOW_FLIP -8.397 (12.9) · TRAIL_1 -8.871 (12.2). ⭐⭐⭐ THE LOSS IS ALMOST PERFECTLY MONOTONE IN THE NUMBER OF TRIPS. Buy-and-hold the same coin-day: -1.463%, BETTER than every trading mechanism. Hindsight zigzag ceiling: +11.786% per coin-day. Best mechanism reaches -14% of the ceiling. CLOCK_30 and FLOW_FLIP have POSITIVE gross (+0.386, +0.383) and still lose 6-8% a day to 9.6 and 12.9 round trips. ⚠ ⭐⭐⭐ THIS IS THE SHARPEST STATEMENT OF THE PROJECT'S CENTRAL WALL. It is not that the exit is wrong or the entry is wrong — the dip entry does not locate leg bottoms AT ALL, so gross is ~0 and every additional round trip is 0.681% of pure subtraction. 'Trade the swings' is right about the movement existing (+11.79% a coin-day of real up-legs) and wrong about our ability to find its turns. ⛔ IT ALSO INVALIDATES A CONCLUSION I PUBLISHED HOURS EARLIER: 'the peak has no home, 33% land after 240min' was measuring the high of an 8-hour hold — a DRIFT, not a swing — because of the same lockout bug. The operator caught the design error; I did not. ⭐ The direction problem is now the whole problem, and nine readings have been tested against it: only pace clears its null and it gives SIZE, not direction. ⛔ Untested still: a MAKER entry (0.32% round trip vs 0.681%) would halve the cost column but not create gross edge — at 12.9 trips it is still -4.1%/day. |
| one_win_per_run_up 2026-09-03 · operator: 'we need a strategy that simply works to capture at least one win out of each of the run ups that we find.' Do not try to buy the bottom — detect a run-up already under way, take ONE trade in it, and take a small fast profit. Momentum entry + tight target is the one combination never tested here. | ⛔ EVERY CELL LOSES AND THE PARAMETERS DO NOT MATTER, WHICH IS THE POINT. Net per trade: -0.769 / -0.796 / -0.825 / -0.826 / -0.767 / -0.819 / -0.769 %. ⭐⭐⭐ GROSS, backing out the 0.681% cost: -0.088 / -0.115 / -0.144 / -0.145 / -0.086 / -0.138 / -0.088 % — a range of 0.06pp across seven very different target/stop combinations. THE ENTRY HAS NO GROSS EDGE, AND NO EXIT PARAMETER CAN CREATE ONE. Coverage: the live detector reaches only 824 of 3,483 run-ups (24%). Exit split at the best cell: target 17% (+3.42%) · stop 16% (-4.70%) · TIMED OUT 67% (-0.894%). Positive on 1 of 10 days. ⚠ ⭐⭐⭐ TAKEN WITH round_trip_test THIS CLOSES BOTH DIRECTIONS AT MINUTE SCALE. Buying the DIP: gross ~0 (-1.10% to +0.39% across eight mechanisms). Buying the MOMENTUM: gross ~0 (-0.09% to -0.15% across seven). Two opposite entry philosophies, fifteen mechanisms, and every one lands at zero gross — the loss is the 0.681% round trip, every time. ⭐ It is also the NINTH confirmation that 'it has already moved' entries are backwards, and it reproduces leg_latency_test (16 cells, all negative) with a completely different exit. ⭐⭐⭐ THE STRUCTURAL OBSERVATION THAT MATTERS MOST: every positive NET this project has ever produced was measured on 4-HOUR BARS over 120 DAYS of OHLC (vwap_gap inverted +0.32% net; the pullback entry +0.85% net at +8h with a hold-out that held; the hot-day jump). EVERY minute-scale test on our own 10-day tape has come back at or below zero gross. The edge, if it exists, may simply not live at the minute scale — or 10 days of tape is too little to find it. That is a testable claim and it is the cheapest next move: run this same one-win-per-run-up rule on the 240min/120day cache. |
| limit_only_gear_to_win 2026-09-03 · operator: 'gear it to win. just place a sell order for 5% plus the in round trip fee and see what that does.' No stop, no clock, no end-of-day flatten — the ONLY exit is a resting limit at entry x (1 + 5% + 0.681%), so every completed trade is exactly +5.0% net. Positions carried across days. | ⭐ THE FILL RATE IS GENUINELY HIGH: 65.0% on the minute tape (median 748min = 12.5h to fill), 78.1% over 120 days. Expectancy per entry is POSITIVE: +0.944% raw, +0.579% after censoring is fixed, +1.728% over 120 days. ⛔ BUT EVERY SUPPORTING NUMBER FAILS. (1) THE DETECTOR IS WORTHLESS: a RANDOM buyable minute fills BETTER (67.6% vs 65.0%) and has HIGHER expectancy (+1.180% vs +0.944%). (2) THE FILL RATE IS NOT PATIENCE, IT IS THE RALLY THE SAMPLE ENDS ON: entries opened 08-25..08-29 with 6-10 days to fill reached 61.4%, while entries opened 08-31..09-03 with 1-4 days reached 68.2% — MORE TIME, FEWER FILLS. Same inversion over 120 days: 1st fifth (120 days available) 72.7% vs 4th fifth (~48 days) 87.8%. (3) THE TAIL: the unfilled bag has mean MTM -9.92% and median -1.96% but a WORST OF -98.83% — one such position erases twenty +5% wins. (4) ⭐⭐⭐ THE CAPITAL SIMULATION IS THE REAL ANSWER: at 3/5/10 concurrent positions the strategy opened only 7/19/31 trades in 120 DAYS and ended with 100% OF CAPITAL IN UNSOLD POSITIONS AND ZERO CASH IN EVERY CONFIGURATION. Returns -1.6% / +15.6% / -1.2% on n=7/19/31 — noise. ⚠ ⭐⭐⭐ WHAT IT ACTUALLY DOES: it wins most of the time and then it STOPS. Two thirds to three quarters of positions reach +5%, and the quarter that never do permanently occupy a slot, so the strategy converts its capital into a bag of losers and grinds to a halt — 7 to 31 trades in four months, every slot full, no cash. That is the classic failure mode of never taking a loss, and it is invisible in per-trade expectancy, which is exactly why per-trade expectancy was the wrong number to look at. ⭐ The +5% target itself is not the problem and the fill rate is a real, usable property of crypto volatility. The problem is that nothing ever closes a loser, so capacity goes to zero. ⛔ ANOTHER BROKEN CHECK, caught before reporting: the first censoring table banded FILLED trades by time-to-FILL instead of time-AVAILABLE, printing a 92.5% fill rate for the entries with the LEAST time — backwards and confident. Time-to-fill and time-available are different quantities. |
| target_ladder_and_confirmed_entry 2026-09-04 · operator: 'fix the entry as you had it before... even if you have go down to 3% at first or 1% plus the round trip.' And then: 'we found a way to determine if it was a flash run up or a real one. apply that so we only enter those trades.' Two tests: (a) the target ladder at +1/2/3/5% net across three entries, (b) the buy-share confirmation from buy_sell_pressure.py used as an ENTRY FILTER for the first time. | ⭐ THE LADDER FOUND THE FIRST ENTRY THAT BEATS RANDOM AT EVERY RUNG. DIP expectancy +0.147 / +0.561 / +0.987 / +1.563% at +1/2/3/5%, against RANDOM -0.384 / -0.126 / -0.042 / +0.273%. Lower targets fill faster (DIP@1%: 89% fill, median 3.2h; DIP@5%: 64%, 22.9h). ⛔ BUT THE CONFIRMATION FILTER FAILS AND IS INVERTED AT ITS STRONGEST SETTING. Filter >=0.50: entered +0.958% vs REJECTED +1.041% = -0.083pp. >=0.55: +0.976 vs +1.008 = -0.031pp. >=0.60: +0.713 vs +1.115 = -0.401pp — the tighter the filter, the worse it does. Across buy-share bands the expectancy is non-monotone (+0.733 / +1.768 / +0.901 / +1.594 / +0.713%), which is noise, not a signal. ⛔⛔ AND WAITING IS EXPENSIVE: the same trades taken IMMEDIATELY at the trigger return +1.949% against +0.958% after a 45min wait — waiting costs about -1.0pp, with +1.53% of entry slip. ⛔ Every capital simulation still ends with ZERO CASH. Stops raise turnover ~10x (10 -> 112 trades opened) and make returns worse (-4.0% no-stop vs -4.9% at -5%; the -10% stop with the 0.55 filter is -16.6%). ⚠ ⭐⭐⭐ THE CONFIRMATION SIGNAL IS REAL AND MEASURES THE WRONG THING. buy_sell_pressure found buy share separating RAN from FADED at +15-60min (0.575 vs 0.480) and that finding stands — but RAN/FADED is a LABEL defined by a 4h forward return, and against MONEY at a +3% limit the same reading pays nothing. That is the eleventh time in this project that a reading has separated a constructed label and not the money (run_up_pct 0.709 vs label / 0.548 vs money; the activity measures at AUC 0.60-0.62 flipping on the offset guard). ⭐⭐ THE OPERATOR'S INSTINCT ABOUT THE FLASH IS NOT REFUTED — waiting DOES avoid fades — it is that the fades you avoid cost less than the runners you buy 1.53% higher. That exact trade-off was measured on 09-03 (avoids 64% of faders, still loses, runners are 3.1x the damage) and it reproduces here with a completely different exit. ⛔ Per-entry expectancy here still counts an unsold bag at market on a tape that ends on a +5% BTC day, so the +1.949% headline carries the same rally artifact as the +5% test. CAPITAL, NOT SIGNAL, IS THE BINDING CONSTRAINT: 10 days can only complete ~2 slot cycles, so the 120-day simulation remains the real capacity answer. |
| spike_entry_flow_exit 2026-09-04 · operator: 'you even had the exit trigger. it was showing the reversal of buyers to sellers.' The flip had been tested 3x and lost — but always on entries with gross ~0. Test it on the CAPITULATION SPIKE, the only entry that clears its own null, using hourly bars rebuilt from our own minute tape so both readings exist on the same rows. | ⛔ IT FAILS IN EVERY IMPLEMENTATION, AND THE FAITHFUL ONES FAIL WORST. PERFECT (hindsight) +4.447%, 73% win. Best LIVE exit is HOLD_48 at -0.140% — DOING NOTHING. Then TRAIL_3 -0.478% · FLOW_FLIP (per bar) -0.717% at 19% WIN · HOLD_12 -0.864% · FLOW_FLIP_2 -0.877% · FLOW_MEAN_4 -0.943% · FLOW_MEAN_3 -0.966% · FLOW_MEAN_4_45 -0.977%. Best live is INSIDE a day-matched null (-0.016%). ⭐ THE DIAGNOSIS: 73% of these entries touch profit and average +4.447% at their peak, and every flow-based exit fires before that. FLOW_FLIP wins only 19% of its trades because 42% of the hours inside a winning climb close DOWN (daily_winner_path) — a 0.50 line per bar is a hair trigger on a statistical tendency. ⚠ ⭐⭐⭐ THIS IS NOW THE MOST-TESTED SINGLE IDEA IN THE PROJECT: four prior tests (exit_mechanism_test -0.434%, round_trip_test -8.397%, buy_sell_pressure's own -0.96%) plus five implementations here. THE REVERSAL IS A REAL DESCRIPTION OF WHAT FADES DO AND IT IS NOT A USABLE EXIT TRIGGER. ⭐⭐ AND THE FAITHFUL IMPLEMENTATION BEING WORSE IS THE INFORMATIVE PART — I first implemented it per-bar and assumed the windowed version would rescue it; the 3h and 4h means are WORSE than the per-bar version, so the failure is not a tuning artefact. ⛔ Note the window: this 10-day own-tape sample has a NEGATIVE baseline (-0.746%) against the 30d OHLC baseline of +0.969%, so absolute numbers here are not comparable to the replication. 🔭 PARKED AT THE OPERATOR'S INSTRUCTION (2026-09-04): 'I bet one of the factors is the distance from buyer to seller, that will be a factor we add later, not now, just keep it in the back of your head.' |
| what_predicts_arming_and_bidask 2026-09-04 · the armed/unarmed split looked like the biggest separation in the project (+3.446%/93% win vs -3.311%/14%) but declares itself at 13h. Is there anything at entry or in the first hours that predicts it — and does the operator's bid/ask spread ('I'm talking about the bid / ask') do it? | ⛔⛔⭐ THE LABEL ITSELF IS LARGELY DEFINITIONAL AND THAT IS THE FINDING. realised_vol predicts ARMING at AUC 0.669 (outside a luckiest-of-28 null of 0.067) and predicts FINAL MONEY at AUC 0.491 — nothing. By realised_vol quartile: armed 31%/38%/55%/67% BUT 'fell -3%' 22%/45%/44%/61% and final net -0.010/-0.342/+0.036/-0.222% — NO PATTERN. High-volatility coins touch +3% more often AND fall -3% more often in equal measure. 'Armed' is substantially a VOLATILITY SORT, and since ARMED is DEFINED as 'reached +3%', the +3.446% vs -3.311% split is close to saying trades that went up made money. ⛔ THE BID/ASK SPREAD SEPARATES BUT IS NOT INDEPENDENT: raw SPREAD_bidask before entry reads ARMED 0.0014 vs UNARMED 0.0005, AUC 0.669 (and 0.625/0.612/0.607 at +1/2/4h), with SPREAD_narrowing@1h at 0.318 — best |0.182| against a luckiest-of-16 null of 0.168, clearing by only 0.014 on 98 entries. INSIDE realised_vol terciles it collapses: mid 0.550, high 0.495, low too few. It was measuring the same thing volatility measures. ⭐ ONE GENUINE ODDITY WORTH KEEPING: price_move@1h AUC 0.364 — the trades that go on to arm FALL FURTHER in the first hour (-1.81% vs -1.02%). ⚠ ⛔⭐⭐⭐ THIS CORRECTS MY OWN CLAIM FROM AN HOUR EARLIER. I called the armed/unarmed split 'the biggest separation in 71 methods' and it is mostly an artefact of a threshold label: ARMED means the position touched +3%, so armed trades having positive returns is nearly a restatement. The enrichment trap (already caught twice in this project, in run_anatomy and mover_history) reappeared in a new costume — as a LABEL rather than a sample. ⭐ THE RULE THAT SHOULD HAVE CAUGHT IT: score against MONEY, never against a constructed label. realised_vol vs arming 0.669, vs money 0.491, is that rule stated in one line. ⭐ THE OPERATOR'S FACTOR WAS A REASONABLE HYPOTHESIS AND IS NOW TESTED RATHER THAN ASSUMED: a wider bid/ask before entry does mark the trades that arm, but only because wide spreads and high volatility are the same coin. ⛔ It deserves ONE re-run when the book-valid window is weeks rather than 2 days and 98 entries. |
| lower_entry_on_the_spike 2026-09-04 · operator: 'why don't we go in at a lower number than on the trades... what you are saying is there is a better entry. am I right?' Two readings pointed that way: the trades that arm FALL FURTHER in the first hour (-1.81% vs -1.02%, AUC 0.364), and retrace depth was monotone on 09-01 (limit_75 positive against 22 of 22 exits). So rest a limit UNDER the capitulation spike's close instead of taking it. | ⛔ THERE IS A BETTER PRICE AND NOT A BETTER ENTRY. Per-signal by depth, averaged over the three windows: market -0.140% · -0.5% -0.372% · -1% -0.227% · -1.5% -0.077% · -2% +0.028% · -3% -0.157%. NOT MONOTONE — an inverted U peaking at -2%, best single cell -2%/4h at +0.065% (+0.205pp over market) out of 18 tried. ⛔⛔ AND THE DECIDING COLUMN: THE LIMIT SYSTEMATICALLY CATCHES THE WORSE TRADES. At -2%/4h the 33% it filled made +0.199% per fill while THE 67% IT MISSED WOULD HAVE MADE +0.760% at market. That holds at every depth and every window and GROWS with how long the limit stays live — at 12h the missed signals are worth +1.6% to +2.1%. ⚠ ⭐⭐⭐ THIRD INDEPENDENT CONFIRMATION OF THE SAME EFFECT: 09-01 missed_fills_are_the_runners found the 39 setups a limit never caught went +11.149% while the 211 it caught went -0.555%; the per-setup rescoring of the pullback candidate found the same; and now it reproduces on a completely different entry. A RESTING LIMIT DECLINES THE RUNNERS, and the deeper or longer it rests the more of them it declines. ⭐ THE OPERATOR'S REASONING WAS SOUND AND THE DATA THAT MOTIVATED IT IS REAL — the trades that arm really do fall further first. The problem is that you cannot separate 'fell further and came back' from 'fell further and kept falling' at the moment the limit would rest. ⭐ WHAT THIS SETTLES: take the capitulation spike's own close. The +0.2pp available from a better price is smaller than the cost of the signals a limit declines, and it is best-of-18 on top of that. |
| cancel_and_reenter 2026-09-04 · operator corrected the previous test's design: 'if a limit doesn't catch then that means we have to just change the entry. we cancel that bid, create a new one to catch that wave. it's not a failure just a different buying signal.' The lower_entry test scored a no-fill as 0.0 — the trade never happening. His model is cancel-and-re-enter: the signal is still valid, so take the wave at the new price. | ⭐ HIS CORRECTION IS RIGHT AND IMPROVES 11 OF 15 CELLS over limit-only, most where the limit was throwing away the most trades: -0.5%/2h goes -0.460% -> -0.291%, -0.5%/4h -0.429% -> -0.227%, -3%/12h -0.247% -> -0.041%. ⛔ BUT THE FAMILY STILL LANDS AT BREAKEVEN. Best cell -2%/12h at -0.001% against a market entry of -0.140% = +0.139pp, best-of-15. Shallow limits with the chase are still WORSE than market (-0.291 / -0.227 / -0.235 vs -0.140); only the -2% and -3% cells edge past it. The 'chase leg' — what the missed trades returned when bought late — swings from +0.629% to -0.347% with no stable pattern, so there is no story about when chasing pays. ⚠ ⭐⭐ THE OPERATOR CAUGHT A REAL MODELLING ERROR: treating a no-fill as a lost opportunity rather than as a re-entry is not how anyone trades, and it made the limit family look worse than it is. Corrected, the limit-declines-the-runners effect is smaller than lower_entry_test reported — but it does not reverse. ⭐ WHAT SURVIVES ALL OF 09-04's ENTRY WORK: TAKE THE CAPITULATION SPIKE'S CLOSE. Resting lower, resting longer, and chasing after a miss all land within a rounding error of it and most land worse. ⛔ ALL OF THIS IS ON THE 10-DAY OWN-TAPE WINDOW where the baseline is -0.746% and market entry is -0.140%; the same entry on 30d OHLC reads +1.776% against a +0.969% baseline. THE ENTRY COMPARISON NEEDS NO BUY/SELL DATA — only the armed exit does — so this family should be re-run on 120d of OHLC for ~10x the sample before -2% is believed or dismissed. |
| entry_timing_sweep_and_wait_filter 2026-09-04 · operator: 'if there are runners going for 20%, 30% or more, we are not entering and exiting at the right time... are we getting in early enough or did we need to get in before your entry point. we should test the entry point. move it up and down.' Then: 'why don't we just enter the winners a bit later? do we lose all the profit by passing the false run point?' | ⭐ EARLIER IS MUCH WORSE, LATER IS FREE. Entry at -1h is the worst cell on the board (-2.040%), -2h -1.602%, -3h -1.385%; the spike close -0.140% with the best win rate (42%); +3h +0.040%, +6h +0.170%, +12h -0.081%. ⭐⭐⭐ AND THE BIG-RUNNER COLUMNS ARE FLAT ACROSS EVERY ENTRY POINT: >=+20% peak stays 3.6-4.8% and >=+30% stays 1.6-2.7% whether you enter three hours early or twelve hours late. WE ARE ALREADY GETTING INTO THE RUNNERS — entry timing is not what costs them. Mean PEAK is +5.63% while we capture -0.140%: the gap is the exit, the same 4.7pp the ceiling test found. ⛔ THE DRAWDOWN SIGNAL IS REAL INFORMATION AND NOT ACTIONABLE AS A CUT: drawdown in the first 4h vs FINAL MONEY reads AUC 0.653 (never dropped 1% +1.315% vs dropped 1%+ -1.541%, spread +2.856pp, OUTSIDE its null) — but cutting at the checkpoint realises -2.1% to -4.2% and lands WORSE than hold-all in 10 of 12 cells. ⛔ AND THE COMBINATION FAILS ITS HOLD-OUT: wait 6h + skip anything that dropped 0.5% gives +0.703% on 27% of signals, MONOTONE in both dimensions (tighter filter better, longer wait better) — but INSIDE a day-blocked null of +1.015%, positive on only 4 of 9 days, and the hold-out FLIPS SIGN: first half -1.894% (n=52), second half +1.878% (n=115), with 09-01 alone contributing +6.759% on 19 trades. ⚠ ⭐⭐⭐ THE OPERATOR'S QUESTION WAS THE RIGHT ONE AND ITS ANSWER RELOCATES THE PROBLEM: he suspected we were entering at the wrong time and the sweep says we are not. Runner availability is FLAT from -3h to +12h, so no entry timing captures more 20-30% moves than any other. What the sweep proves is that the +5.63% mean peak is reachable from anywhere and we hold none of it. ⭐⭐ 'Do we lose the profit by passing the false run point' — NO. Waiting 3-6h past the spike is free and slightly better (+0.170% vs -0.140%). ⭐ MONOTONE STRUCTURE IS NOT ENOUGH. The wait-and-filter surface was monotone in BOTH dimensions, which is much stronger evidence than a lucky cell, and it still failed the hold-out with a sign flip. 167 trades over 9 days cannot carry a claim. ⛔ 12th and 13th time a during-trade reading has been real information that no rule can act on, because acting on it realises the loss. |
| two_round_trips_vs_conditional_ride 2026-09-04 · operator: 'what if we exit then come back in after 3%. there is a second to the coin, right? let's see if instead of holding for a big percentage we can get 2 small percentages that cover the round trips.' | ⛔ TWO SMALL ROUND TRIPS LOSE TO ONE. Bank once and stop = 3.000%. Best arbitrary re-entry is buy-back-at-market at 2.833% (-0.167pp); every pullback version is far worse (1.765% at -1%, 2.015% at -2%, 2.269% at -3%, 2.375% at -5%) — waiting for a better price selects worse trades, for the fourth time tonight. ⭐ RE-ENTERING ON A NEW SPIKE IS DIFFERENT AND DECAYS MONOTONICALLY WITH DELAY: a fresh spike within 12 bars (2 days) arrives for 34% of trades, median wait 16h, second leg +0.563%, total 3.190%; within 4 days +0.272% / 3.113%; within 8 days -0.372% / 2.787%; within 15 days -0.173% / 2.870%. THE SECOND TO THE COIN IS REAL BUT ONLY INSIDE ABOUT TWO DAYS. ⭐⭐⭐ AND IT STILL LOSES TO SIMPLY RIDING THE STRONG BARS: ride-if-strong 3.656% (1st half 3.241%, 2nd 3.959%, BOTH POSITIVE) vs re-enter-on-new-spike 3.207% (1st half 2.865% — FAILS). Composing them reaches 3.698% overall but DROPS THE FIRST HALF TO 3.013% from 3.241% — the re-entry leg dilutes, adding +0.042pp overall while costing 0.228pp in the half that matters. ⚠ ⭐⭐⭐ THE OPERATOR'S ECONOMICS WERE RIGHT AND THE SELECTION WAS NOT. Two +3% round trips only cost an extra 0.681%, so the arithmetic favours them — but the second entry arrives WITHOUT the signal that made the first work, and entries without the spike lose money (random median -5.2% in the capital sim). The edge is in the SIGNAL, not in the coin. ⭐⭐ THE DECISIVE COMPARISON IS FIRST-HALF, NOT OVERALL: the combined rule has the best headline (3.698%) and the worse first half (3.013% vs 3.241%). Choosing on the overall number would have shipped the weaker rule. ⭐ 34% of coins do throw a fresh spike within two days and those pay +0.563% — the persistence is real, it is just worth less than reading the bar in front of you. ⇒ THE RULE STANDS AS: bank +3%, and ride only when the bar that filled you closed strong. |
| buy_at_the_stop_level 2026-09-04 · operator: 'where the stop is firing, could we move the buy to that point in the run? and then we had the signal not to buy the others because they would be falling?' Two questions: is the stop level a better ENTRY, and does entry depth distinguish the coins that keep falling? | ⭐ DEPTH RAISES THE FILL RATE A LOT: at the spike close 52% fill · -3% 60% · -5% 64% · -7% 70% · -10% 77% · -15% 82%. But the level is only REACHED 62/44/30/17/7% of the time, so most signals produce no trade. Per fill it gets WORSE at the deep end: -7% -1.030%, -15% -2.251%, -20% -4.138%. ⛔ AND UNDER CAPITAL IT LOSES TO SIMPLY USING A STOP: spike close + -7% stop 209 trades +22.8% maxDD 3.5% (halves 13.6%/13.6% — the most stable cell of the session) vs -5% below 75 trades +13.4%, -3% below 45 trades +4.1%, -7% below 46 trades +1.9%, -10% below 50 trades +4.4%, -7% below WITH a -7% stop 138 trades +11.8% (halves 0.8%/10.0%, unstable). ⛔⛔ AND THE SECOND QUESTION IS A HARD NO: 'kept falling' (filled then still lost >5%) is 18% AT EVERY DEPTH — 0%, -3%, -5%, -7%, -10%, -15% — and 21% at -20%. ENTRY DEPTH CARRIES NO INFORMATION ABOUT WHICH COINS KEEP FALLING. ⚠ ⭐⭐ THE STOP IS BETTER USED AS A STOP. Waiting to buy at the stop level means only reaching it on 30% of signals and skipping the 70% that never fall that far — including the winners. That is the LIMIT-DECLINES-THE-RUNNERS effect for the FOURTH time tonight, now at depth rather than at the margin. ⭐⭐⭐ AND THE INVARIANT 18% IS THE USEFUL PART: the proportion that keeps falling does not move with entry depth, which means PRICE DEPTH CARRIES NO SIGNAL ABOUT WHICH ONES DIE. Combined with everything else ruled out today — buy-share confirmation, pace, volume, spread, trade size, depth tilt, the people-vs-dollars split — the false-run signal is not in the post-entry price path and not in the microstructure. The one place not yet looked at against THIS entry is how the coin behaved BEFORE the spike (run_anatomy, 09-01, never tested against the capitulation entry). ⭐ Best configuration found in the whole session: spike close + -7% stop, halves 13.6%/13.6%. |
| wait_the_danger_window_buy_survivors 2026-09-04 · operator, clarifying: 'instead of buying the coins earlier, we wait to the time where the stop would have fired to buy only the ones that didn't fall.' A survivorship filter — let the danger window pass, then buy whatever held up. | ⛔ EVERY CELL IS WORSE THAN BUYING AT THE SPIKE CLOSE (-0.488% per signal), INCLUDING 'PER TAKEN' WHICH IS SURVIVORS ONLY. 4h/-7%: 94% survive, fill 52%, per taken -0.526%, per signal -0.497% · 4h/-5%: 92%, -0.618% · 4h/-3%: 81%, -0.795% · 8h/-7%: 92%, -0.750% · 12h/-3%: 64%, -0.903%. ⭐ THE FILTER BARELY FILTERS — 94% survive a -7% test in four hours — so it removes 6% of signals that were not the problem, and charges three ways for it: you buy the bounce at a higher price, the fill rate falls (53% -> 47-52%), and the survivors have less room left to reach the target. ⚠ ⭐⭐⭐ THIRD 'WAIT AND CONFIRM' VARIANT TESTED IN ONE SESSION AND ALL THREE FAILED THE SAME WAY (the buy-share confirmation, the wait-and-held-up rule, and this). It ties directly to the invariant 18%: THE COINS THAT KEEP FALLING DO NOT ANNOUNCE THEMSELVES IN THE DANGER WINDOW. Not by entry depth, not by survival, not by drawdown. ⭐⭐⭐ THE HONEST CONCLUSION FROM SIX INDEPENDENT ANGLES: the false-run signal is NOT in the price path after the spike. What remains untested against THIS entry is what the coin was doing BEFORE the spike — run_anatomy (09-01) found a real pre-move signature on a different trigger and it has never been applied here. That is the last stone. |
| method | what it scored |
|---|---|
| factor_probe_trade_count_n 2026-09-05 · Operator 2026-09-05, the pre-set first question: 'is there a factor we haven't been tracking or one that we are not testing it with that might give a sharper read.' Three readings verified genuinely untouched by the K pipeline, all FREE (already on disk, no new collection): n_x = the bar's TRADE COUNT vs its own trailing-24 normal (how many PEOPLE acted); size_x = vol/n, the average trade size vs its normal (how BIG each actor was); breadth = how many coins spiked in the same 4h bar (entry_decider computes coins_spiking_this_bar for the AI and no mechanical test has ever used it). | ⭐⭐⭐ n_x IS A REAL READING AND IT IS THE CROWD, NOT THE SIZE. Quintiles of n_x are monotone in the money: +0.083 / +0.376 / +0.724 / +0.977 / +0.878 %, win 53% -> 77%. Top half +1.016% vs bottom +0.202%, gap 95% CI +0.029...+1.668 (clears 0). size_x is DEAD: AUC 0.469, non-monotone, gap CI includes zero, and its hold-out FLIPS SIGN (+0.698 one way, -0.483 the other). The 2x2 decomposition of the volume spike: many-people cells +1.165% / +0.872%, few-people cells +0.174% / +0.231% -- a 4-5x split on WHO, and nothing on HOW BIG. ⭐⭐ IT SURVIVES THE THREE CHECKS THAT KILLED THE LAST THREE CANDIDATES: (1) NOT A THINNESS TILT AND IN FACT THE REVERSE -- LIQUID half AUC 0.664 / gap +1.159pp vs THIN half 0.557 / +0.477pp, i.e. STRONGER where a real order can be filled, only the second reading in this project ever to do that (the other is the swing-persistence); (2) IT PASSES ITS HOLD-OUT IN BOTH DIRECTIONS -- pick the cut on the bear half, score the bull half: gap +0.303; pick on the bull, score the bear: +0.674 -- the exact check K's OWN parameter choice failed at percentile 20; (3) WITHIN-DAY it picks the right COIN among those firing the same day: mean gap +1.377pp over 37 days with >=4 signals, day-bootstrap 95% CI +0.413...+2.444, clears zero, positive 21/37. ⭐⭐⭐ AND IT IS INDEPENDENT OF THE STANDING PANEL FINDING: spearman(n_x, close_in_bar) = -0.015, and the two stack -- K + n_x>=3.14 + cib<=0.25 is +1.426% per signal on n=122/65 days, CI +0.997...+1.825, against K's registered +0.609% (CI +0.041...+1.123). Each reading alone lands ~+0.8...+1.0% with a CI that touches zero; together the CI clears it by a full point. ⛔⛔ THE ONE THAT NEARLY KILLS IT: spearman(n_x, vol_x) = +0.853. n_x is largely the reading the rule ALREADY has under a different name -- vol_x = n_x x size_x by construction. What saves it is the conditional test: inside a vol_x band it still separates (MID gap +0.724 AUC 0.629, HIGH gap +0.826 AUC 0.570) -- BUT IT INVERTS IN THE LOW BAND (vol_x 2.02-3.02: gap -0.269 while AUC reads 0.576, the two disagreeing). ⛔ And on its own AUC (0.612, |dev| 0.112) n_x DOES NOT clear the luckiest-of-4 null (p95 = 0.120). The reading that clears it on AUC is n_x_prev (0.640, percentile 98.2, was the crowd ALREADY arriving in the previous bar) -- but n_x_prev's quintiles are NOT monotone (all the money is in the top fifth) and its own gap CI INCLUDES ZERO. The two instruments pick different winners, which is the honest state: n_x has the money shape, n_x_prev has the rank statistic. ⛔ breadth IS REFUSED: AUC 0.499, non-monotone quintiles. Its spectacular cell -- 10+ coins spiking together, +2.914%, 97% win, n=36 -- is 36 signals across exactly TWO DAYS (08-20 and 08-21). A day-blocked interval cannot even be computed. That is ONE EVENT wearing a factor's clothes, and the instrument caught it before it was reported. ⚠ ⛔⛔ IN SAMPLE. FITTING BY CONSTRUCTION -- this is discovery on the very data K was chosen from, and every number above was found by looking at outcomes. NOTHING HERE MAY AMEND CANDIDATE K, whose clock is running from 2026-09-04T14:00Z; handoff rule 4 says the panel may not nominate a champion and a replacement needs its own pre-registration from zero. The correct destination is forward_panel.py as a free variant. ⭐ THE OPERATOR'S FRAME IS WHAT FOUND IT: 'the trade is the outcome of peoples thoughts and actions.' Dollar volume says how much money; n says how many people; vol/n says how big each one was. Asked that way the answer is unambiguous -- the tradeable half of a capitulation spike is the CROWD, and the size of the individual actor is noise. ⭐ It also converges with the only other reading that ever cleared a null in this project: PACE (trades/min vs own normal, AUC 0.569 on 10 days of minute tape). n IS pace, sitting in 120 days of cached OHLC that every study threw away. |
| capacity_curve_board_ceiling 2026-09-05 · Operator 2026-09-05: a sizing formula that 'peaks according to the most a coin will handle in a trade', names 'the peak of the trading for the amount of money we should have in the account', and 'gives the place where we take money out as profit'; then max out this board and move to a second and third. HOW MUCH MONEY DOES THIS BOARD HOLD, and how fast does it fund the next one? | ⭐⭐⭐ THE CEILING IS REAL AND IT IS LOW. At the registered 10%-of-a-15-min-slice assumption the peak is an $82,500 account ($7,500 ticket x 11 peak concurrent) making $20,831/yr; 10% below it, $65,000 making $18,683/yr at 2.5% maxDD. Beyond that the signal RATE collapses: 10.5/day at a $16k screen -> 1.3/day at $5k tickets -> 0.9/day at $7.5k -> 0.3/day at $25k, where the mean net goes NEGATIVE. ⭐⭐⭐ THE NUMBER THAT MATTERS MOST IS THE FUNDING RATE: one full board throws off ~$359/week, so it funds the NEXT board in ~181 WEEKS. The operator's 'new boards will come up weekly' is off by two orders of magnitude at this edge and this ceiling — weekly cadence needs outside seed capital, a bigger edge, or a much higher A*. ⭐⭐ AND THE ACCOUNT IS SIZED BY A PILE-UP, NOT BY WORKING CAPITAL: peak concurrent 11-13 positions against a MEAN of 0.6-0.8 open. Roughly 90% of the required account is insurance against one crowded bar (the 08-20/08-21 flush put 21 coins in a single bar) and is idle the rest of the time. ⭐ The headroom the operator asked for is cheap and asymmetric: 5% below peak keeps 89.7% of the dollars, 33% below still keeps 72.4%, while being ABOVE the true peak buys capital that earns exactly zero. ⛔ THE WIDEST ERROR BAR IS NOT THE RULE, IT IS THE LIQUIDITY CONSTANT: at 5% / 10% / 20% of a slice the peak account moves $32,500 / $82,500 / $165,000 — a 5x span on an assumption NOBODY HAS EVER MEASURED. Measuring what a coin will really absorb is worth more than any further refinement of the entry. ⚠ ⛔ IN SAMPLE, and every dollar figure is conditional on an edge whose day-blocked 95% interval still includes zero (~180 days from clearing). If the edge is zero the dollars are zero and THE SHAPE IS STILL CORRECT — the shape is a liquidity fact, the level is a bet. ⭐⭐ TWO DEFECTS THIS TOOL FOUND IN ITSELF, both of the project's own signature class: (1) the first grid put its peak on the LARGEST SLOT VALUE IN THE GRID — a maximum sitting on a boundary is not a maximum, it is the boundary; the axis was extended until the curve turned over on its own. (2) It reported 'account = slots x ticket', which at 20 slots and 0.6 mean concurrency was mostly idle cash that never bought anything — replaced by an exact open/close sweep for PEAK SIMULTANEOUS DEPLOYMENT. ⭐⭐⭐ AND ONE IN THE SIMULATOR: sim_engine.run collapsed three different reasons for not taking a signal into one silent `continue`, so a 1-slot account that STOPPED DEAD after 5 of 1,220 signals (it could no longer afford its own ticket after a loss) was read as having simply 'traded 5 times'. slots / cash / liquidity call for OPPOSITE actions — more capital, smaller ticket, nothing helps — and now each is counted and reported. A cash-starved cell can never be quoted as the peak. ⛔ ALSO FOUND: sim_engine's default coin_cap of 0.01 is 1.6x LOOSER than the rule's own liquidity screen (0.00625 of a 4h bar); the tool passes the rule's value explicitly rather than inheriting the divergence. |
| sizing_to_the_bar_raises_the_ceiling 2026-09-05 · Operator 2026-09-05: 'once we get to that peak we can start doing the larger trades as allowed. that would increase the ceiling. see how even this math gets complicated and changes as it moves?' EVERY capacity number this project had produced assumed a CONSTANT ticket, and nothing had ever named that assumption. | ⭐⭐⭐ HE IS RIGHT AND IT ROUGHLY QUADRUPLES THE WORKABLE ACCOUNT. Last account still profitable: CONSTANT ticket $30,000; SIZED TO THE BAR $120,000. THE MECHANISM: a constant ticket caps the whole account at what the SMALLEST acceptable bar can absorb - to put $10,000 on every trade you may only trade bars that hold $10,000, which discards 192 of 264 signals (72 trades left at $30k, TEN at $250k). Sizing to the bar keeps all 264 and only varies the stake. ⭐⭐ AND THE DRAWDOWN FALLS RATHER THAN RISES: at $30,000 constant is 72 trades / +$1,878 / 22.6% maxDD, variable x4 is 258 trades / +$10,882 / 9.4% maxDD - 5.8x the money at LESS THAN HALF the drawdown, because the same dollars are spread over 3.6x as many positions. At $60,000 constant is NEGATIVE (-$3,220) while variable x4 makes +$10,845 at 6.6%. At $120,000: constant -5.9%, variable x4 +8.9% at 3.3% maxDD. ⛔ THE CEILING IS REAL AND STILL THERE: by $250,000 every policy is negative. The average ticket saturates around $10,100 because the bars cannot absorb more, so the account is ~92% idle and a few very large trades on the biggest bars carry the loss. ⛔ IT IS NOT FREE LEVERAGE: at a SMALL account 'variable x2' is pure concentration - at $1,404 it doubles the money (+81.3% vs +40.7%) and doubles the drawdown (43.7% vs 23.4%). The gain only appears once the account is large enough that a constant ticket would have been discarding signals. ⛔ x4 and x10 converge exactly: the BAR cap binds, not the concentration cap. ⚠ ⭐⭐⭐ THE LESSON IS THE SECOND HALF OF HIS SENTENCE, NOT THE FIRST: 'see how even this math gets complicated and changes as it moves?' THE PEAK IS NOT A NUMBER, IT IS A FUNCTION OF THE SIZING POLICY - and every ceiling this tool had printed was conditional on a policy nobody had ever stated out loud. Same failure class as the rest of the project: an unnamed assumption doing load-bearing work. It also vindicates the SHAPE of his original formula - a ticket that moves with the account and with what the market in front of it will take, rather than a constant to be optimised once. ⛔ IN SAMPLE, and conditional on an edge whose day-blocked 95% interval still includes zero. The RANKING of sizing policies is a liquidity fact and robust; the levels are a bet. |
| regime_slot_dial_bear_vs_bull 2026-09-05 · Operator 2026-09-05: 'didn't you say we could switch back and forth between doing 3 during a bear and 8 during a bull market?' I did NOT say that - I said 3 slots was the only config positive in the BEAR HALF, which is a label applied AFTER THE FACT by splitting the window in two. Knowing the regime AT THE TIME is a different problem. Tested anyway, with the hindsight version built as the CEILING. | ⛔ INSIDE NOISE, AND THE CEILING IS WHAT SAYS SO. fixed 3 slots +15.2% (bear +3.3 / bull +12.0) · fixed 8 slots +20.0% (bear -1.4 / bull +21.4) · PERFECT HINDSIGHT switching +24.7% (bear +3.3 / bull +21.4). THE ENTIRE PRIZE FROM PERFECT REGIME KNOWLEDGE IS +4.7pp. The three live dials -- BTC's own trailing 24h / 48h / 7d return, thresholded at zero -- returned +20.5% / +17.9% / +21.1%, i.e. the best captures about a QUARTER of that prize and ONE OF THEM (48h) IS WORSE THAN NOT SWITCHING AT ALL. The dials disagree with each other by 3.2pp while claiming an effect of 1.1pp: the spread exceeds the signal, which is the signature of noise, and picking the 7d one from three would be luckiest-of-3 on the same data everything else was chosen from. ⚠ ⭐⭐⭐ AND IT CORRECTED MY OWN NUMBER FROM THE PREVIOUS ANSWER. I had reported '3 slots makes 3.5x the dollars with 5x the drawdown' from 3 slots x $468 against 8 slots x $175 on a fixed $1,403.85 pot -- CONFOUNDED BY TICKET SIZE, which is the exact error execution_grid.py's own header documents ('THE SLOTS AXIS WAS CONFOUNDED IN THE FIRST RUN OF THIS FILE ... nothing but position size'). With the TICKET HELD CONSTANT the concurrency effect runs the OTHER WAY: 8 slots +20.0% beats 3 slots +15.2%, at 11.1% vs 9.2% drawdown. So 3 slots only 'wins' on a fixed pot because it deploys 3.3x more capital per trade -- that is LEVERAGE, not a concurrency edge, and its 23.4% drawdown is the price. More slots is better per unit of risk AND per dollar deployed. ⛔ Both comparisons are valid; they answer different questions, and only the fixed-ticket one isolates concurrency. ⭐ The bear/bull DIFFERENCE is real (8 slots really is negative in the bear half); what does not exist is a way to know which one you are in, live -- the same wall leverage_dial_test S3 hit when nothing forecast the board. |
| regime_prize_and_its_decay 2026-09-05 · Operator 2026-09-05: 'we know that most of the time the system is in a bear market and not a bull market, so the switch is something that is real and we need a better way to detect it ... it might not be the switch between the two but the LEAD INTO the switch ... I bet we can predict when the markets are going to change. and we can even do it with hours of time before it happens.' Before building any detector: WHAT IS PERFECT REGIME KNOWLEDGE WORTH, and how fast does it decay if the detector is late? | ⭐⭐⭐ HIS PREMISE IS CONFIRMED: the board falls more often than it rises - 65.3% of DAYS (79/121) and 51.5% of 4h bars have the MEDIAN COIN down; median 4h board move -0.029%. ⭐⭐⭐ AND THE PRIZE IS LARGE - 2.4x the slot dial's. Baseline (no knowledge) +20.0% / 11.1% maxDD. SKIP every bear signal: +31.4% at 4.9% maxDD = +11.4pp. SKIP only the WORST-QUARTILE bear: +33.7% at 5.2% = +13.7pp - BETTER THAN SKIPPING ALL BEAR, which means the target is not a clean bear/bull line but the worst 25%, a materially easier thing to detect. Sanity check passes hard: trading ONLY bear returns -11.1% (-31.0pp), so the label is real and correctly signed. Half-size-in-bear captures +8.7pp. ⭐⭐⭐ THE DECAY CURVE IS THE ACTUAL DELIVERABLE AND IT VINDICATES HIS 'HOURS BEFORE' INSTINCT EXACTLY: instant +11.4pp · 4h LATE +7.0pp (61% of it survives) · 8h LATE +1.3pp (11%) · 12h LATE +0.6pp (5%) · 24h LATE -10.6pp (ACTIVELY HARMFUL, worse than no detector at all) · 48h late -11.8pp. THE ENTIRE PRIZE LIVES IN THE FIRST 4-8 HOURS. A detector is not merely better when fast - it is WORTHLESS past ~8h and NEGATIVE past ~24h. That is a hard specification, and it did not exist before this measurement. ⚠ ⛔ IN SAMPLE, ONE PATH, 165 trades in the load-bearing arm, and NO NULL YET - the day-blocked interval on the +11.4pp has not been computed and must be before this is quoted as established. ⛔ AND THE TARGET IS HARD BY CONSTRUCTION: the label is the board's next 24h, so a detector for it is literally forecasting the board's return - the exact thing leverage_dial_test S3 failed at when 'nothing forecasts the board'. What has changed is that we now know what success is WORTH (+11.4pp), what LATENCY it must meet (<=4-8h), and that the easier target (avoid the worst quartile) pays MORE than the harder one (call every bear). ⭐ Judge every proposed detector against +11.4pp and the decay curve, never against zero - the discipline the exit test (+4.54% ceiling, nine mechanisms, none reached it) and the slot dial (+4.7pp) both earned. |
| regime_prize_null_and_first_detector 2026-09-05 · Operator 2026-09-05: 'run the null on it then start on the worst-quartile detector.' | ✅ THE PRIZE IS REAL. +11.4pp against N1 (random trades) at PERCENTILE 100.0, and against N2 (sitting out the same 52 random whole days) at PERCENTILE 99.8 - null median -9.7pp, 95th -0.4pp. Knowing WHICH periods to sit out beats sitting out the same amount of TIME at random. It is not 'trade less'. ⛔⛔ AND ROUND ONE OF THE DETECTOR FAILS FLAT. Eight live readings against the worst-quartile target: board's last 4h 0.499 · last day 0.489 · last week 0.497 · breadth 0.503 · dispersion 0.495 · BOARD-LEVEL n (how many people acted market-wide) 0.525 · dollar volume 0.534 · capitulation breadth 0.476. Best is |0.034| against a day-blocked luckiest-of-8 null whose 95th is 0.094 -> PERCENTILE 13.9, deeply inside. The money test agrees: of 16 filter arms exactly ONE is positive (breadth, skip-bottom-quartile, +4.4pp) and it is not the reading that ranked best (AUC 0.503, i.e. nothing) - one in sixteen is what chance produces. Everything else LOSES, several badly (-18.1pp). ⚠ ⭐⭐⭐ THE TWO RESULTS TOGETHER ARE THE FINDING: THE THING IS REAL AND WE CANNOT SEE IT. That is the 10th confirmation of leverage_dial_test S3's wall ('nothing forecasts the board') - but for the first time with a MEASURED PRIZE (+11.4pp) and a MEASURED LATENCY BUDGET (<=4-8h) behind it, so the next attempt knows what it is aiming at. ⭐ WHAT IS NOT RULED OUT, and it is the honest next move: every reading here is a 4h-OHLC board AGGREGATE. The one input class OHLC cannot supply is the BUY/SELL SPLIT, and board_pressure_gate (registry 56) already measured that board BUYING breadth is a reading BTC cannot see (corr +0.233 with BTC-from-open). It has never been tested against THIS target - because the minute tape is only ~12 days deep, about 70 board bars, ~17 of them in the worst quartile. THAT IS 'CANNOT SEE', NOT 'NOT THERE', and the tape grows a day every day. ⭐ ALSO NOT TESTED: the operator's own framing was the LEAD INTO THE SWITCH - a TRANSITION target ('is the board about to change state') rather than the LEVEL target tested here ('is the next 24h bad'). Different label, genuinely untried, and named rather than silently skipped. |
| regime_transition_target 2026-09-05 · Operator 2026-09-05: 'try the transition target' - his own framing was 'it might not be the switch between the two but THE LEAD INTO the switch and when it is that we decide is the place we change methods.' So: condition on the current state and ask whether it is about to flip, rather than asking the unconditional level question that failed. | ⛔ THE TRANSITION FRAMING FAILS TOO. DETERIORATE (currently OK, about to go worst-quartile - the actionable one): 13 readings, best crowd_shift |0.064| against a day-blocked luckiest-of-13 null whose 95th is 0.116 -> PERCENTILE 44.8, deeply inside. The readings do score slightly better than on the level target (board_vol_x 0.553 vs 0.534, board_n_x 0.541 vs 0.525) but the null grew with them and nothing clears. RECOVER: best |0.121| -> percentile 67.0, inside. CHANGE: board_24h |0.198| -> percentile 100.0, WHICH 'CLEARS' AND IS PURE DEFINITION - verified rather than assumed: P(state changes | currently BAD) = 73.4% vs P(state changes | currently GOOD) = 24.6%, so 'will it change' is mechanically predicted by 'where are you now', and board_24h IS where you are now. Same artefact class as VWAP_GAP's AUC 0.211 and winners_today's 0.238. ⭐⭐⭐ AND THE CHECK THAT EXPOSED IT GAVE THE ROOT CAUSE OF THE WHOLE LINE OF WORK: P(next 24h BAD | currently BAD) = 26.6% vs P(next 24h BAD | currently GOOD) = 24.6%, against a 25.0% base rate. A TWO-POINT DIFFERENCE. THE BOARD'S NEXT 24 HOURS IS STATISTICALLY INDEPENDENT OF ITS LAST 24 HOURS. ⚠ ⭐⭐⭐ THIS IS WHY EVERYTHING FAILED, AND IT IS A BETTER RESULT THAN ANOTHER FAILED READING. If state(t+1) is independent of state(t), then NO framing built on the board's own past - level, transition, deterioration, recovery - has anything to work with, and the 21 reading x target combinations tested across registry 105 and 106 were never going to succeed. It is not that we have not found the right reading; the target has no autocorrelation structure at this horizon at all. ⛔ STOP SPENDING HERE. The prize is real (+11.4pp, percentile 99.8) and the latency budget is known (<=4-8h), but it cannot be reached from 4h-OHLC board aggregates. ⭐ THE ONE INPUT CLASS LEFT is outside the board's own price/volume history: THE BUY/SELL SPLIT, which OHLC cannot supply at all and which board_pressure_gate (registry 56) already showed is a reading BTC cannot see. It needs tape depth - ~12 days today, about 17 worst-quartile bars - so it is a MEASUREMENT TO SCHEDULE, not one to run now. That is 'cannot see yet', not 'not there'. |
| scale_out_test 2026-09-05 · The gap exit_mechanism_test (registry 64) named and could not close: 'every mechanism here is ALL-OR-NOTHING. Scaling out -- taking profit in pieces, which is what he actually described -- is a different mechanism class and has never been run.' Operator 2026-09-03: 'made money the whole time TAKING PROFIT AND KEEPING THE MONEY.' | ⛔ EVERY SCALING ARM LOSES TO K. K (all out at +3%) +0.459%/trade, 20.0% return, 11.1% maxDD, halves -1.4/+21.4. half at 1.5 + half at 3.0: +0.172% (10.6% return). thirds 1.5/3/6: +0.250%. quarters 1.5/3/6/12: +0.220%. third/third/ride: +0.258%. The only arm that edges K on the mean is 'half at +3.0, half RIDES to the cut' at +0.490% (+0.031pp, inside noise) and it pays for it with 27.2% maxDD against K's 11.1% and a BEAR half of -10.3% against K's -1.4% - worse on both risk measures, so it is not better. ⭐⭐⭐ THE MECHANISM IS PLAIN IN THE WIN RATES: banking half at +1.5% RAISES the win rate (66% -> 69%) and LOWERS the mean. K's +3% resting limit already converts an uncertain path into a CERTAIN +3% - the reason it wins 71% in the original measurement - so there is no right tail left for scaling to capture. Selling a tranche below the target only gives away part of a certain gain, and it does nothing at all for the losers, whose side is untouched. ⚠ ⭐⭐ SCALING OUT PAYS WHEN THE PAYOFF HAS A FAT RIGHT TAIL YOU CANNOT TIME. Under a resting limit there IS no right tail - every winner is exactly +3.0%. That is why the mechanism class that was right in principle (the peak has no home, so do not try to find it) is wrong HERE: K does not try to find the peak either, it refuses to look for one. The two ideas solve the same problem and only one of them can be applied at a time. ⛔ This reproduces from a completely different direction the 2026-09-04 finding that 'the +3% limit is not broken and must not be loosened mechanically' and that the target sweep is flat either side of +3.0%. ⛔ IN SAMPLE, six arms, best-of-N; the number that matters is that NO arm's per-trade CI excludes zero and no arm improves BOTH halves. The question is closed, not deferred. |
| short_side_first_test_ever 2026-09-05 · THE FIRST SHORT-SIDE TEST IN THE HISTORY OF THIS PROJECT. Grepped first: of 106 registry entries only 3 mention 'short' and all three mean short HOLD PERIODS, never short SELLING. Every strategy ever tested here has been long-only - in a board that, as measured 2026-09-05, falls on 65.3% of days. Operator: 'do all 3 in the order you suggested.' | ⛔ SHORTING IS DEAD, AND THE CONTROL IS WHAT KILLS IT. A RANDOM SHORT returns -0.543% - the 65% down-days do not become a short edge, because the fee eats the drift whole. Every short entry is WORSE than that random short: capitulation-short -0.860% (so the spike is genuinely a REVERSAL, not continuation - the long thesis confirmed from the other side), blow-off-short -0.822%. Blow-off LONG is -0.359%, better than a random long (-0.720%) but still negative. ⭐⭐⭐ THE DECISIVE ARITHMETIC, IN GROSS: random short gross +0.138%/24h, random long gross -0.039%. THE MARKET'S ENTIRE DIRECTIONAL DRIFT IS THE 0.177pp GAP BETWEEN THEM - and the round trip is 0.681%, THREE POINT EIGHT TIMES LARGER than the drift it would have to beat. Meanwhile Candidate K's long entry is +1.311% GROSS: THE ENTRY IS WORTH 7.4x THE MARKET'S WHOLE DIRECTIONAL DRIFT. Even in the bear half the best short (blow-off) reaches only +0.088% against a random short's -0.110% - a 0.20pp edge against a 0.681% fee. ⚠ ⭐⭐⭐ THE PREMISE WAS RIGHT AND THE CONCLUSION DOES NOT FOLLOW. The board really does fall 65.3% of days; being long-only in it really is a structural handicap. But the handicap is worth 0.177pp and the toll to act on it is 0.681% - so direction is not where the money is, and never was. This retires the whole 'we are long-only in a falling market' concern with a number instead of an argument, and it is the second time today a measurement has shown that a real asymmetry is far smaller than its own transaction cost. ⛔ It also never reached the practical objections (Kraken margin availability on small alts, borrow cost, liquidation) - it did not need to; the gross arithmetic closes it before any of that. ⛔ IN SAMPLE, and the random-bar arms are a 4,000-draw sample of 14,544. |
| switching_in_play_prize_and_accuracy_bar 2026-09-05 · Operator 2026-09-05: 'but if we did trade the change in the market correctly by SWITCHING TRADES IN PLAY it would have a major shift am I correct?' A different mechanism from anything tested: not which signals to TAKE, but flipping a position that is already open when the market turns. | ⭐⭐⭐ HE IS RIGHT, AND IT IS THE LARGEST PRIZE THIS PROJECT HAS EVER MEASURED. hold LONG the whole 24h = +1.213%/trade. PERFECT knowledge of the 24h direction at entry = +5.861% (+4.9pp). PERFECT flipping on every 4h bar WITH FLIP COSTS PAID = +15.443% (+14.2pp) - against Candidate K's entire registered edge of +0.537%. ⭐⭐ AND THE FLIP COST BARELY MATTERS, WHICH INVERTS EVERY PREVIOUS 'MORE TRIPS' RESULT: perfect switching flips a median of 3 times in 24h, costing 1.97%, against a 14.2pp prize (free-flips 17.412% vs paid 15.443%). The in-and-out-all-day test lost monotonically to trip count because it re-paid a whole new ENTRY each time and had no gross edge; here the gross edge is the market's own MAGNITUDE, which is enormous. ⛔⛔ AND THE ACCURACY SWEEP IS THE DELIVERABLE - BREAK-EVEN IS ~62%. Per-4h-bar switching: 50% acc -> -2.664% (i.e. -3.9pp WORSE than simply holding), 55% -> -1.297%, 60% -> +0.546% (still -0.67pp worse), 62-63% -> break-even, 65% -> +2.382% (+1.2pp better), 70% -> +3.905%. The cheaper one-decision-at-entry variant has the SAME ~62% break-even (50% -> -0.655%, 62% -> +0.048pp, 65% -> +0.196pp) on a smaller +4.9pp prize. ⚠ ⭐⭐⭐ THE BAR IS NOW A NUMBER: 62% DIRECTIONAL ACCURACY, OR SWITCHING IS WORSE THAN DOING NOTHING. Context that decides it today: the best directional reading this project has ever produced is PACE at AUC 0.569 (~53-54% accuracy) AND IT FORECASTS SIZE, NOT DIRECTION - asked directly whether high pace picks UP over DOWN among turning points, the quartiles ran 48.7 / 52.2 / 50.4 / 53.3%, a 4.6pp tilt that is not even monotone. Nine readings tested at turning points; one cleared its null and it was the wrong quantity. ⛔ BEING WRONG IS EXPENSIVE, NOT NEUTRAL: at 50% accuracy switching returns -2.664%, not zero, because you pay flip costs AND sit systematically on the wrong side. ⭐ WHY THE PRIZE IS SO LARGE IS ALREADY IN OUR OWN RECORD: the board's harvestable MOVEMENT is +11.79% per coin-day while its whole directional DRIFT is 0.177pp. Switching monetises magnitude; every direction-free strategy we own monetises drift. That is the same magnitude-vs-direction split PACE found, one level up. ⇒ THIS IS THE STRONGEST ARGUMENT YET FOR SPENDING ON DIRECTION - and the first time the required accuracy has been a measured number rather than a hope. |
| flip_detector_round_one 2026-09-05 · Operator 2026-09-05: 'right so we need to figure out the flip and we need to have it before it happens.' Registry 109 measured the prize (+14.2pp) and the bar (62% directional accuracy). Can any reading we own call the next 4h bar's direction on an open position? | ⛔ ROUND ONE FAILS ON BOTH TESTS. ⭐ FIRST, THE BASE RATE MATTERS AND IT IS NOT 50%: 'always long' already scores 52.5% on this population, so a detector must add 9.5pp, not 12pp. There IS free structure in bar position - the up-rate runs 59.1% (bar+1, the bounce) / 52.7 / 56.4 / 50.0 / 48.7 / 48.4% (bar+6) - but as a rule it is worth only +1.0pp. Best reading overall is close-in-bar at 54.4%; then pnl 53.6, to_target 53.6, range_x 53.7, bars_held 53.5, board 53.2, size_x 52.7, rel 52.7, last_ret 52.3, vol_x 52.3, n_x 52.0. THE NULL'S OWN MEDIAN IS 53.1% - so seven of the eleven readings are AT OR BELOW what a shuffled label with the same threshold search produces. Best real 54.4% -> percentile 86.3 against a 95th of 55.5%: INSIDE. ⛔⛔ AND EVEN CLEARING THE NULL WOULD NOT BE ENOUGH: the break-even is 62%, the best reading is 54.4%, SEVEN POINT SIX POINTS SHORT - and at 54.4% switching does not merely fail to help, it LOSES to simply holding, because being wrong is expensive rather than neutral. ⚠ ⭐⭐ THE 62% BAR IS DOING EXACTLY THE JOB A PRE-MEASURED CEILING IS SUPPOSED TO DO. Without it, close-in-bar at 54.4% - 1.9pp over always-long, on 2,244 decisions - would have looked like the beginning of something and justified weeks of refinement. Against the bar it is visibly 7.6pp from useful, and the decision to stop is immediate rather than gradual. ⭐ WHAT IS STILL UNTRIED FOR THIS TARGET, in the order they are worth: (1) THE MINUTE TAPE AT COIN LEVEL - buy/sell split inside each 4h bar of an open position, never tested for direction and the one input class OHLC cannot supply; thin today (~40 in-tape signals x 6 bars) and growing daily. (2) INTERACTIONS - only marginals were tested here. (3) AN AI DECIDER ON THE FLIP: this is now the LARGEST Law #3 case the project has ever had - a decision a filter provably cannot make, with a 14.2pp measured prize behind it, larger than the exit's 4.9pp and the entry's 2.4pp. ⛔ Tempered by the fact that K-AI, built on exactly that logic for the entry, showed NO edge in sample (percentile 68.8). A measured prize justifies the attempt; it does not promise the result. |
| positioning_data_a_new_input_class 2026-09-05 · Operator 2026-09-05: 'think outside the box. if the detector isn't in the 4h ohlc then where is it. find where we can detect it instead of trying to with what we are limited currently to.' | ⭐⭐⭐ AND IT DOES NOT NEED THE TAPE TO GROW - unlike the buy/sell split, this comes with a YEAR of history and can be tested against the whole 120-day window today. FIRST LOOK, 217 signals: OI-change quartiles over the prior 24h give mean net +1.636% / +0.588% / -0.029% / +1.034% (win 76 / 61 / 54 / 75%) against a full-K population mean of +0.630%. THE BOTTOM QUARTILE - open interest COLLAPSING, -52.7% to -5.8% - RETURNS +1.636% AT 76% WIN, 2.6x the whole population. ⭐⭐ THE MECHANISM SPLIT, and it is the operator's own fake-vs-real intuition with a measurable cause behind it for the first time: a capitulation on FALLING open interest is a LIQUIDATION FLUSH (forced sellers being closed out - it exhausts itself and bounces) at +1.112% over 108 signals / 54 days, versus a capitulation on RISING open interest which is NEW SHORTS ARRIVING and can keep going, at +0.507% over 109 signals / 61 days. ⭐ Funding rate alone is weaker and not monotone: quartiles +0.725 / +0.530 / +0.808 / +1.163%, with the most-positive-funding quartile at 76% win. ⛔ THE OI RELATIONSHIP IS NOT MONOTONE EITHER - it is U-SHAPED (Q1 +1.636, Q4 +1.034, middle flat), so what matters looks like the MAGNITUDE of the positioning change rather than its direction. That is the same magnitude-not-direction split PACE found at the turning points, appearing again in a completely different input class. ⚠ ⛔⛔ THIS IS A FIRST LOOK, NOT A RESULT. 217 of 379 signals, 45 coins of 102, IN SAMPLE, NO NULL RUN, and the buckets were read after the outcomes were known. It is registered as a FOUND INPUT CLASS, not as a finding. ⭐⭐⭐ WHY IT IS WORTH MORE THAN ANOTHER PRICE READING, and this is the structural argument nothing price-derived has ever had: PRICE CAN TREND FOREVER, LEVERAGE CANNOT. When funding is extreme somebody is paying a toll to hold, and a toll has a forced end; positioning must mean-revert in a way price need not. That is a MECHANISM rather than a pattern - the first one this project has had. ⭐ And it lands exactly on the operator's own doctrine: 'the trade is the outcome of peoples thoughts and actions and that is what we need to track.' We spent months INFERRING what people were doing from price while the exchange published it hourly. NEXT: fetch the remaining 57 coins, then a pre-registered test with a day-blocked null - and separately against the FLIP target, which is where the 62% bar and the 14.2pp prize are. |
| positioning_test_full_set_with_null 2026-09-05 · Operator 2026-09-05: 'fetch the rest of the coins and run the null.' The first look (registry 111) read its buckets AFTER the outcomes on 45 coins - fitting. This is the SAME six readings, named before the run, on the full set, with a day-blocked luckiest-of-6 null, a money CI and a two-way hold-out. | ⭐⭐ ONE READING SURVIVES ALMOST EVERYTHING: abs_oi_chg, THE MAGNITUDE OF THE 24h OPEN-INTEREST CHANGE. Quartiles are MONOTONE in the money: +0.367 / +0.057 / +0.713 / +1.426%, win 59 / 56 / 63 / 80%, across 41-48 days each. Top quartile +1.446% vs rest +0.376%, gap +1.070pp with a day-blocked 95% CI of +0.337...+1.907 that CLEARS ZERO. ⭐⭐⭐ IT PASSES THE HOLD-OUT IN BOTH DIRECTIONS - the exact check Candidate K's own parameter choice FAILED at percentile 20: cut chosen on the 1st half, scored on the 2nd, gap +0.611; cut chosen on the 2nd, scored on the 1st, gap +1.692. ⭐⭐⭐ AND IT IS GENUINELY NEW INFORMATION, WHICH IS WHAT n_x COULD NOT CLAIM: spearman(abs_oi_chg, vol_x) = +0.127 and spearman with the bar's dollars = +0.073 - essentially uncorrelated with everything we already read, against n_x's +0.853 with vol_x. ⭐ The first look also REPLICATED on 57 new coins: the OI-collapse bucket was +1.636% on 45 coins and is +1.579% (77% win, 44 days) on 97. ⛔⛔ BUT IT DOES NOT CLEAR ITS RANK NULL: |AUC-0.5| = 0.087 against a day-blocked luckiest-of-6 95th of 0.103 - PERCENTILE 85.6, INSIDE. ⛔ AND THE OTHER FIVE READINGS DIE: oi_chg_24h (AUC 0.450, gap CI spans zero, hold-out FLIPS SIGN -0.568 then +0.486), funding_chg (hold-out FLIPS HARD, +1.357 then -1.113), funding, oi_x, oi_chg_4h all with CIs through zero. ⚠ ⭐⭐⭐ THE DIRECTION STORY FROM THE FIRST LOOK DID NOT SURVIVE; THE MAGNITUDE STORY DID. 'A capitulation on FALLING open interest is a liquidation flush' was the first look's headline - on the full set oi_chg_24h has AUC 0.450 and a hold-out that flips sign. What replicates is that ANY LARGE POSITIONING CHANGE, either direction, precedes a better outcome. THAT IS THE THIRD TIME IN ONE DAY MAGNITUDE HAS BEATEN DIRECTION - PACE forecasts size not direction, the switching prize is magnitude while the drift is 0.177pp, and now open interest. It is starting to look like a property of this market rather than three coincidences. ⛔ NOT ESTABLISHED: in sample, 282 signals, and the rank null is not cleared. ⭐ But it is the STRONGEST NON-PRICE READING THE PROJECT HAS PRODUCED - 4 of 5 checks including the two that have killed the most candidates here (the hold-out, and independence from what we already read). ⇒ It belongs in forward_panel as an arm with its own pre-registration, exactly as n_x did, NOT in Candidate K. ⭐ PROCESS NOTE: after the flip detector failed I said the remaining routes 'both need the tape to get deeper, which happens on its own' - and the operator pushed back. Stopping a LINE OF INQUIRY and stopping THE SEARCH are different things, and I blurred them. The answer was one API call away on the venue we already trade. |
| regime_detector_round_three_positioning 2026-09-05 · Operator 2026-09-05: 'we still need to find a way to predict the bear vs bull market.' Rounds 1-2 tested 13 readings, ALL price/volume board aggregates. Positioning had just proved itself a genuinely new input class on the ENTRY, so this is its first test against the REGIME target. | ⛔ ROUND THREE FAILS ON BOTH TARGETS. DETERIORATE: best funding_disp |0.081| against a day-blocked luckiest-of-6 95th of 0.106 -> PERCENTILE 78.3, inside. LEVEL: best funding_disp |0.050| vs a 95th of 0.093 -> PERCENTILE 45.2, inside. board_abs_oi - the exact reading that passed four of five checks on the ENTRY - reads 0.546 and 0.540, i.e. nothing. The money test agrees on every arm. ⭐ THAT ACROSS-TARGET CONTRAST IS ITSELF INFORMATIVE: the same positioning reading that separates outcomes on an INDIVIDUAL COIN AT A CAPITULATION says nothing about THE BOARD'S NEXT DAY. Positioning is a coin-level state, not a market-level forecast. ⚠ ⛔⛔ THREE ROUNDS, 34 READING x TARGET COMBINATIONS, NOTHING HAS EVER CLEARED - and the per-round nulls cannot see across rounds, so anything that ever does clear here needs FORWARD confirmation rather than another in-sample refinement. ⭐⭐⭐ BUT THIS ROUND FINALLY NAMES THE REAL CONSTRAINT, WHICH IS NOT THE READINGS: WE HAVE BEEN TESTING A 24-HOUR TARGET FOR A CONCEPT THAT IS NOT 24 HOURS. 'Bear vs bull' is a multi-day regime; every round has asked 'is the next 24h bad'. The reason nobody tried a 7-day target is arithmetic: 120 days of spot OHLC gives ~17 independent 7-day windows, which cannot support any test at all. ⭐⭐⭐ AND THE FIX WAS FOUND WHILE CHECKING: KRAKEN'S FUTURES OHLC GOES BACK TO 2025-10-07 - 334 DAYS, 2.8x OUR SPOT CACHE'S 120 - because the spot endpoint caps at 720 candles and the futures charts endpoint does not. 334 days gives ~48 independent 7-day windows and spans more than one regime cycle. THAT, NOT A FOURTEENTH READING, IS THE NEXT MOVE ON THIS QUESTION: extend the board history from futures candles, then test the multi-day regime target that has never been testable. |
| futures_history_extension_and_7day_regime 2026-09-05 · Operator 2026-09-05: 'extend the history with futures ohlc and test the 7 day target.' Registry 113 argued the constraint on the bear/bull question was the 24h HORIZON, not the readings, and that 120 days of spot could not support a multi-day target (~17 independent 7-day windows). | ⛔ THE 7-DAY TARGET FAILS: real persistence gap +4.1pp against a block-shuffled null whose 95th is +16.1pp -> PERCENTILE 67.8, INSIDE. So does everything else. 24h +1.5pp (pct 67.8) · 3d +3.8pp (85.5) · 7d +4.1pp (67.8) · 30d -17.3pp (26.2, so the dramatic-looking REVERSAL is noise). ⭐ 14d is the only one that appears to clear: +14.1pp, percentile 95.8 - AND IT IS ONE OF FIVE HORIZONS TESTED, so P(at least one of five exceeds its own 95.8th by luck) = 19.3%. The expected MAXIMUM of five uniform draws is already ~83. IT IS NOT SIGNIFICANT, and it rests on 23 INDEPENDENT WINDOWS. ⭐⭐ THE REAL PRIZE OF THE RUN IS A REPLICATION: the 24h independence result measured on SPOT (26.6% vs 24.6%, +2.0pp) reproduces at +1.5pp on 2.8x the history AND a different venue. That was the load-bearing claim behind 'stop spending here', and it now has an out-of-sample confirmation it did not have this morning. ⚠ ⭐⭐⭐ THE EXTENSION IS KEPT REGARDLESS OF THE NEGATIVE RESULT - it is a durable asset, not a one-test tool. 333 days of validated 4h board history for 97 coins is now on disk and every future question gets it for free; the multi-day regime question is simply not one it could answer. ⛔ AND THE HYPOTHESIS THAT MOTIVATED IT IS NOW REFUTED RATHER THAN UNTESTED: registry 113 said the constraint was the horizon, not the readings. It was not. Lengthening the horizon from 24h to 7 days does not produce structure that a block-preserving null cannot reproduce. FOUR ROUNDS on the regime target - 13 price readings, 5 transition readings, 6 positioning readings, 5 horizons - and nothing has cleared once its own null was built honestly. ⭐ The instrument that keeps earning its place: EVERY dramatic number here (the 14d +14.1pp, the 30d -17.3pp) collapsed the moment the null preserved the autocorrelation the overlapping windows create. Overlapping windows are how a 333-day sample pretends to be 1,832 observations. |
| flip_bar_is_invariant_and_round_two 2026-09-05 · Operator 2026-09-05: 'I think the flip detector is the real thing to find but it is eluding us at the moment.' Two sub-questions: (a) can the 62% break-even be lowered by restructuring WHEN we act, and (b) does the positioning input class - which passed 4 of 5 checks on the ENTRY - work on the FLIP? | ⭐⭐⭐ (a) THE BAR DOES NOT MOVE. Break-even sits at ~63-65% whether the detector speaks on 5% of bars or 100% - selectivity changes the PAYOFF (from +0.32pp to +10.14pp at 90% accuracy) and not the THRESHOLD. And restricting to the biggest bars does not help either: at 60% accuracy the excess over hold-long is -0.50 / -0.24 / -0.17 / +0.02 / -0.08pp across every-bar / top-50% / top-25% / top-10% / top-3% by size. ⭐⭐⭐ THE REASON IS STRUCTURAL AND CLOSES A WHOLE FAMILY OF IDEAS: on a big bar you win big when right AND LOSE BIG WHEN WRONG, so magnitude cancels out of the break-even entirely. Only accuracy survives. ⛔ THIS KILLS THE HOPE THAT THE PROJECT'S ONE FORECASTABLE QUANTITY - MAGNITUDE (PACE, AUC 0.569) - COULD RESCUE THE FLIP. Magnitude tells you what is at stake, not which way, and the stake is symmetric. (b) ROUND 2 OF THE FLIP DETECTOR FAILS LIKE ROUND 1: always-long scores 54.0%; oi_1bar 54.0 · abs_oi_1bar 53.5 · oi_24h 53.5 · funding 53.5 · funding_chg 54.3. Best is funding_chg at PERCENTILE 77.6 against a null whose own median is 53.3%, and 7.7pp SHORT of the bar. ⚠ ⭐⭐⭐ SIXTEEN READINGS ACROSS TWO ROUNDS ALL LAND BETWEEN 53% AND 54.4%, and every null median sits at ~53%. That consistency is itself the information: 4h DIRECTION is not weakly-predictable-but-we-have-not-found-the-reading; it is unpredictable from every input class we currently hold - price, volume, trade count, position-in-hold, unrealised P&L, board state, funding and open interest. ⭐ AND THE PRIZE IS STILL THE LARGEST THE PROJECT HAS MEASURED (+14.2pp against Candidate K's whole edge of +0.537%, a 26x difference), so the operator's judgement that this is the thing to find is correct on the arithmetic. ⇒ WHAT IS LEFT IS EXACTLY TWO THINGS, both named rather than hoped for: (1) THE MINUTE TAPE AT COIN LEVEL - the buy/sell split INSIDE each 4h bar of an open position, the one input class OHLC cannot supply and the only one never tested for direction; it grows a day per day and is already scheduled for the regime question (~2026-10-04). (2) AN AI FLIP DECIDER - the largest Law #3 case the project has, 14.2pp against the exit's 4.9 and the entry's 2.4. Tempered: K-AI was built on exactly this logic for the entry and showed NO in-sample edge (percentile 68.8). ⛔ AND THE HONEST FRAME FOR BOTH: they must clear 62%, not 55%, and being wrong is expensive rather than neutral (-2.664% at a coin flip). |
| loser_detection_reframe 2026-09-05 · Operator 2026-09-05: 'so we need to know how to detect the losers since the signal for the flip is already detected, correct?' and 'we get rid of them and we should have the beginning of a big run, right?' A different question from the flip: not WHICH WAY does the next bar go (per-bar direction) but WILL THIS TRADE LOSE (per-trade, integrating six bars). | ⭐⭐⭐ THE REFRAME IS RIGHT AND THE BAR IS MUCH EASIER: a skip rule breaks even at 41.1% PRECISION among the trades it skips, against a 32.1% loser base rate - i.e. it needs to be 9pp better than random, versus the flip's 62% accuracy. ⛔⛔ BUT THE CEILING IS 6.6x SMALLER. Perfect removal of all 120 losers takes K from +0.609% to +2.774% = +2.16pp, against the flip's +14.2pp - because skipping can only AVOID a loss while flipping CONVERTS it into a gain. ⛔ AND THE RESULT IS THIN: 5 of 48 rules cleared the 41.1% precision bar, which at that threshold on that base rate is close to what chance produces. Best precision: n_x lowest 10% at 59.5% (n=37), n_x lowest 20% at 47.3%, n_x lowest 33% at 44.7%, usd highest 10% at 48.6% (n=37), vol_x lowest 10% at 43.2%. The best MONEY rule - skip the lowest third of n_x - lifts K to +0.877% (+0.259pp) and lands at PERCENTILE 94.8 against dropping the same NUMBER of trades at random: INSIDE, by two tenths of a point. ⭐ AND IT IS n_x, WHICH IS ALREADY THE PANEL ARM K-n_x>=3.14 - the same finding in a new frame, not a second discovery. ⚠ ⭐⭐⭐ THE COMPARISON THAT MATTERS, AND IT IS THE USEFUL OUTPUT OF THE WHOLE EXCHANGE: LOSER-DETECTION is the REACHABLE problem (ceiling +2.16pp, bar 41.1% precision, and an existing reading already sits at 44.7%); THE FLIP is the BIG one (ceiling +14.2pp, bar 62% accuracy, and 16 readings across two rounds all land at 53-54%). Expected value favours pursuing loser-detection while treating the flip as the moonshot. ⭐ AND THE OPERATOR'S IDEA IS ALREADY RUNNING FORWARD WITHOUT ANYONE CALLING IT THAT: the K-n_x>=3.14 and K-absOI>=15 panel arms ARE loser-detectors - 'skip the lowest n_x' and 'keep only high positioning change' are exactly skip rules. The forward test of his reframe started this morning. ⛔ ANSWER TO 'the beginning of a big run': no. Perfect loser removal is a 4.5x improvement on K's per-trade net, which is real but is not the transformation - the transformation is the flip, and the flip is 8pp of accuracy away from anything we can currently see. |
| loser_detection_focused_attempt 2026-09-05 · Operator 2026-09-05: 'focus on the loser detection then, it's reachable.' First characterise the losers (never done), then try to detect them at entry and in flight. | ⭐⭐⭐ THE ONE REAL FINDING IS THE SHAPE OF THE PROBLEM, AND IT WAS NOT KNOWN BEFORE: THE LOSS IS EXTRAORDINARILY CONCENTRATED. The worst 12 losers - 3.2% of all 374 trades - carry 38% of the total loss (mean -14.99% each); the bottom five deciles of losers together carry only 16.5%. Removing just those 12 takes K from +0.609% to +1.126%; the worst 24 takes it to +1.383%; the worst 39 to +1.655%. ⇒ THE TARGET IS NOT '120 LOSERS', IT IS '12 DISASTERS'. ⭐ TAXONOMY: 75% of losers ROSE BUT NEVER REACHED THE TARGET, only 25% were DEAD ON ARRIVAL - so most losers are near-misses, not obvious failures. ⭐⭐ AND THE DISASTERS ARE VISIBLE IN THE FIRST BAR: bar+1 low averages -2.79% for winners, -2.53% for all losers, and -9.17% FOR THE WORST 10% - a 3.3x separation, in flight. ⛔⛔ BUT IT IS NOT ACTIONABLE. Bailing when bar+1's low breaches a threshold: -5% -0.108pp · -7% -0.143 · -9% -0.017 · -12% +0.025 · -15% +0.025 · -20% +0.009. EIGHTH CONFIRMATION THAT STOPS LOSE, now on the tail where it looked most promising - among the 10 trades that dip <=-9% at bar+1, as many recover as die. ⛔ AND NOTHING AT ENTRY SEES THEM: against 'the worst 24' every reading scores 84.8-86.4% where always-say-no scores 93.6%, best at PERCENTILE 52.8; against 'dead on arrival' best is percentile 65.6. Both deeply inside their nulls. ⚠ ⭐⭐⭐ THE CONVERGENCE THAT IS THE REAL OUTPUT OF THE WHOLE SESSION: LOSER-DETECTION AND THE FLIP NOW BOTTLENECK ON THE SAME MISSING INPUT. Both are blocked not by a lack of ideas but by having exhausted every input class we hold - price, volume, trade count, position-in-hold, unrealised P&L, board state, funding, open interest - and both have exactly two routes left, and they are the SAME two: (1) THE MINUTE TAPE INSIDE THE BAR, which is the only thing that could distinguish a -9% dip that recovers from one that does not (was the selling front-loaded or steady? who was buying it?), and (2) an AI decider. ⇒ THE BUY/SELL TAPE, ALREADY PRE-REGISTERED AND SCHEDULED FOR ~2026-10-04, IS NOW THE BOTTLENECK FOR THREE SEPARATE QUESTIONS - the regime, the flip, and loser-detection. That makes it the single highest-value pending item in the project, and it arrives on its own. ⛔ HONEST CORRECTION TO 'IT'S REACHABLE': the BAR is reachable (41.1% precision vs the flip's 62% accuracy) and that remains true - but no reading we own clears it except n_x, which is ALREADY DEPLOYED as a panel arm. Reachable bar, no new key. |
| was_it_the_real_bottom 2026-09-05 · Operator 2026-09-05: 'if it does drop by 9% and then do a recovery all that meant was it wasn't at the real bottom and it didn't make a real turn.' A reframe from EXIT to ENTRY: can we tell a finished capitulation from an unfinished one? Operationalised three ways: (A) did the spike bar's LOW hold, (B) has this coin already capitulated in the last 24 bars, (C) how far has it already fallen from its 4-day high. | ⭐⭐⭐ THE SEPARATION IS THE STRONGEST THIS PROJECT HAS PRODUCED, AND HIS MECHANISM IS EXACTLY RIGHT AS A DESCRIPTION: low HELD all 24h = +2.337%, 90% win, and ZERO of the 24 disasters; low BROKEN at bar+1 = -0.606%, 55% win, and 20 OF THE 24 DISASTERS; broken later = +0.738%, 4 disasters. EVERY disaster is in a group where the spike's low was taken out. ⛔⛔ BUT IT IS 38% DEFINITIONAL: of the 111 'low held' trades, 42 (37.8%) reached +3% IN BAR+1 - a trade that jumps straight to target CANNOT have broken its low. 'The low held' is substantially a RESTATEMENT OF THE OUTCOME, not a forecast of it. Third definitional artefact of the day (VWAP_GAP AUC 0.211, winners_today 0.238, this). ⛔⛔ AND THE CAUSAL HALF IS KILLED BY THE COST OF WAITING: BAIL returns +0.292% and WAIT +0.379% against K's +0.590%. Against the right control - entering one bar later on ALL signals, which is +0.203% - the low-held filter adds +0.177pp and lands at PERCENTILE 77.6, inside. (B) and (C) are mild and the wrong way round: 2+ prior spikes does BETTER (+0.839%, 79% win) not worse, and the deepest prior falls are not the worst. ⚠ ⭐⭐⭐ THE STRUCTURAL NUMBER THAT CLOSES A WHOLE FAMILY: ENTERING ONE BAR LATER COSTS 66% OF THE EDGE (+0.590% -> +0.203%). The entry edge lives almost entirely in the first bar - which reproduces the standing finding that more than half the gross accrues there, and the 2026-09-04 measurement that being 4h late costs 30%. ⇒ NO RULE THAT REQUIRES WAITING TO CONFIRM CAN WORK HERE, because the thing being confirmed is where the money is. That retires every 'wait for the turn / wait for confirmation' variant at once, including the price-based version of the operator's own flash-vs-real idea. ⭐⭐⭐ AND IT LOCALISES THE ANSWER PRECISELY: 'was this the real bottom' must be answerable WITHOUT WAITING - i.e. from information INSIDE the spike bar itself. That is the minute tape, for the FOURTH time today (regime, flip, loser-detection, and now this). The buy/sell split arrives ~2026-10-04 and is now the bottleneck for four separate questions. |
| where_in_the_bar_was_the_low 2026-09-05 · Operator 2026-09-05: 'how long are the bars you are using. you might be finding the turn then using too long of a period of time to act on it.' Then: 'find it with the 4hour bar then keep reducing the bar till you find the event.' The bars are FOUR HOURS - so 'wait one bar' meant waiting four hours, which is what made every confirmation rule look impossible. | ⭐⭐⭐ FIRST: HIS PREMISE WAS RIGHT AND IT CORRECTED MY OWN CONCLUSION. The decay curve between 0 and 4 hours had never been measured: +0h +1.449% · +1h +1.167% · +2h +1.382% · +3h +1.339% · +4h +0.764%. THE EDGE IS FLAT FOR THREE HOURS AND THEN FALLS OFF A CLIFF AT FOUR. My earlier 'no rule that requires waiting can work' was an artefact of only ever testing a FOUR HOUR wait. ⛔ (The low-held confirmation still fails at 1/2/3h: percentiles 41.9 / 40.3 / 28.4, below the null median.) ⭐⭐ SECOND: WE BUY 2.42% ABOVE THE BAR'S OWN LOW ON AVERAGE (median +1.03%, p90 +5.38%) - 4.1x the entire per-signal edge sitting inside the bar we already have. ⭐⭐ AND THE LOW HAS A HOME: hour 1 26.6% · hour 2 16.2% · hour 3 16.9% · HOUR 4 40.3%. ⭐⭐⭐ THIRD AND THE RESULT: OUTCOME BY WHICH HOUR MADE THE LOW - hour 1 +0.165% (n=37, -1.274pp) · hour 2 +2.274% (n=23) · hour 3 +1.495% (n=25) · hour 4 +1.872% (n=61). SKIPPING THE HOUR-1 SIGNALS - the ones that had already rallied three hours before we bought - keeps 109 of 146 and returns +1.871% vs K's +1.438%, +0.432pp, at PERCENTILE 99.5 against drawing the same number at random. IT CLEARS. ⭐⭐⭐ AND IT IS NOT close_in_bar, WHICH IS ALREADY A PANEL ARM: spearman(lo_hour, close_in_bar) = -0.475, and the effect SURVIVES INSIDE BOTH close_in_bar HALVES - gap +0.803 in the low half, +2.101 in the high half. ⛔⛔ CORRECTED THE SAME DAY, ON 2.8x THE DATA: re-run on 83 days of FUTURES hourly (n=205, 63 days) instead of 30 days of spot hourly (n=146), the rule weakens badly - skip-hour-1 keeps 148/205 and returns +1.029% vs K's +0.841%, only +0.188pp, at PERCENTILE 91.1 - INSIDE. It does NOT survive the expansion. ⭐⭐ BUT THE SPLIT SAYS SOMETHING SHARPER THAN 'IT FAILED': on the 53 days the original test NEVER SAW, the hour-1 penalty is LARGER, not smaller - hour-1 -0.674% vs the rest +0.105%, a gap of +0.779pp, against +0.562pp on the original 30 days. THE HOUR-1 PENALTY REPLICATED OUT OF SAMPLE AND STRENGTHENED. What did NOT replicate is the other half: hour 4 was +1.872% (+0.434pp) on 30 days and is +0.809% (-0.032pp) on 83 - level with K. So 'a fresh bottom is best' is dead; 'a three-hour-old bottom is worst' is alive but not established. ⚠ ⭐⭐⭐ THE PROCESS LESSON, AND IT COST NOTHING BECAUSE THE EXPANSION WAS RUN IMMEDIATELY: percentile 99.5 on 146 signals became 91.1 on 205. The first number was written into memory as CLEARS and had to be corrected within the hour. ⭐ THE OPERATOR'S OWN ARCHITECTURE IS WHAT MADE THE CORRECTION POSSIBLE - he asked for 'a continuing lense that gets us closer to the moment', which forced the question of what depth each rung has, which surfaced that futures gives 2.8x more at EVERY rung (4h 333d vs 120 · 1h 83d vs 30 · 15m 20.8d vs 7.5 · 5m 6.9d vs 2.5), which is what re-tested the finding. ⛔ STANDING RULE THIS EARNS: BEFORE BELIEVING A RESULT AT A NEW RESOLUTION, CHECK WHETHER A DEEPER SOURCE EXISTS AT THAT SAME RESOLUTION. A 30-day answer and an 83-day answer to the same question are not the same test, and the shallower one is the one that will be sitting in the cache. ⭐ WHAT SURVIVES AND IS WORTH CARRYING FORWARD: (1) the 0-to-4h decay curve - FLAT for three hours then a cliff at four - which stands on its own and retired my 'waiting cannot work' conclusion; (2) we buy 2.42% above the bar's own low, 4.1x the per-signal edge, still unexploited; (3) the hour-1 penalty, replicated out of sample, NOT established. ⇒ the next rung (15m, 20.8 days futures) is where it should be re-asked, per his ladder. |
| the_cascade_built 2026-09-05 · Operator 2026-09-05: 'we have to use a continuing lense that gets us closer to the moment till it reaches the time for the AI to make a decision' -> 'build the cascade.' | ⛔ BUILT, RUNNING, AND IT REPORTS THAT NO RUNG EARNS ITS PLACE YET. The --audit mode tests 12 readings across the rungs against day-blocked random draws of the same size: 1h 'low NOT in the first slot' PERCENTILE 92.1 (n=148) · 15m same reading 94.5 (n=76) · 1h 'still falling at the close' 44.4 · 1h 'volume in the back half' 12.7 · 1h 'a single flush' 41.2 · 15m 'volume in the back half' 42.4. NOTHING CLEARS 95, and with 12 readings tested the honest bar is well above 95 anyway. The 5m rung covers only 24 of 374 signals (7 days of data) and is REFUSED as too few to audit rather than reported as a null. ⭐⭐ THE ONE THING WORTH CARRYING: 'THE LOW WAS NOT IN THE FIRST SLOT' IS THE BEST READING AT BOTH THE 1h AND THE 15m RUNG (92.1 and 94.5) - two DIFFERENT TILINGS of the same four hours, not two shots at one dataset. That is the shape of something real and underpowered rather than something absent. ⭐⭐⭐ RE-RUN WITH THE SHAPE RULES INCLUDED (operator: 'run the audit again with the shape rules included'): 10 readings across the rungs. 1h low-not-first-slot 92.1 · 1h still-falling 44.4 · 1h volume-back-half 12.7 · 1h single-flush 41.2 · 15m low-not-first-slot 94.5 · 15m volume-back-half 42.4 · 15m SHAPE A 96.4 · 15m SHAPE B 97.0 · 15m SHAPE A-or-B 99.8 · 15m 3+dips 70.8. THE REGISTERED UNION IS MARKEDLY STRONGER THAN EITHER HALF - 99.8 against A's 96.4 and B's 97.0 - which is what an INDEPENDENT union should look like (their overlap is only 5 of 22 and 27). ⛔⛔ AND A DEFECT IN MY OWN AUDIT MESSAGE, FOUND AND FIXED IN THE SAME RUN: it quoted a FIXED 97th percentile for the multiple-comparison correction and printed '26%', which would have led a reader to DISMISS a reading sitting at 99.8 where the true adjusted figure is 2.5%. A CORRECTION THAT IS TOO BLUNT DISCARDS REAL RESULTS AS SURELY AS NO CORRECTION ADMITS FAKE ONES, AND IT IS THE HARDER ERROR TO NOTICE BECAUSE IT LOOKS LIKE RIGOUR. The line now computes each survivor's OWN adjusted figure: A-or-B 99.8 -> 2.5% SURVIVES; B 97.0 -> 26.1% does not; A 96.4 -> 30.9% does not. ⚠ ⭐ WHAT THE CASCADE MAKES VISIBLE THAT NOTHING ELSE DID: on a live signal it now says, in prose, 'the low came in slot 21 of 48, we are buying +7.05% above it, 29% of the volume landed in the back half' - i.e. the 2.42%-average gap between the bar's low and the price we pay is now ATTRIBUTED to a moment rather than being an aggregate. ⛔ IT DECIDES NOTHING. It assembles context for entry_decider.py, whose prompt is a registered frozen parameter; wiring it into that prompt would restart K-AI's clock at zero and has NOT been done. ⛔ DEPTH IS THE BINDING CONSTRAINT AT THE BOTTOM OF THE LADDER AND IT RESOLVES ITSELF: 4h 333d · 1h 83d · 15m 21d · 5m 7d · our own minute tape 12d. The two bottom rungs deepen daily, exactly like the buy/sell tape already scheduled for ~2026-10-04. Crons added at each rung's own cadence, because a stale sub-bar cache produces an INCOMPLETE TILE over the newest 4h bar - which the tool refuses, so the failure would be visible rather than silent. ⛔ AND THE AUDIT NOW SAYS OUT LOUD THAT ITS OWN COUNT UNDERSTATES THE SEARCH: 'A or B' is DERIVED from the two best of four shape rules, so it is not an eleventh independent draw and its 2.5% is OPTIMISTIC. Its FORWARD CLOCK, not this table, is what settles it. |
| the_shape_of_the_bar 2026-09-05 · Operator 2026-09-05: 'the pattern might not be a straight line it might have a lot of dips but it will be something that is traceable' -> 'test it till you have a means of using it.' EVERY reading this project has extracted from a bar has been a SCALAR - how much volume, where the close sat, where the low was. This is the first reading of the bar's SHAPE. | ⭐⭐ TWO OF FOUR CLEAR IN SAMPLE: A n=22 +2.711% percentile 96.8 · B n=27 +2.661% percentile 97.0, against a population of +1.990%. The two other shape rules do not (choppy alone 29.4, >60% of the bar below mid 85.7). ⭐ A AND B ARE NOT THE SAME RULE - overlap 5 of 22 and 27 - so 'A or B' (n=44, +2.647%) is a real union. ⛔⛔ IT IS NOT SIGNIFICANT: four rules were tested, so P(at least one beats its own 97.0 percentile by luck) = 11.5%, on n=22 and 27 over a 21-DAY window. ⭐⭐⭐ WHAT MAKES IT WORTH A CLOCK IS THE HOLD-OUT PATTERN, WHICH IS COHERENT AND ASYMMETRIC: in the good half of the window (the rest returned +2.6%) the rules add -0.07 / +0.29 / +0.15pp - NOTHING. In the bad half (the rest returned -0.7%) they add +3.18 / +2.28 / +3.21pp. THE SHAPE READING DOES NOTHING WHEN EVERYTHING WORKS AND CARRIES THE PERIOD WHEN NOTHING DOES. That is a DOWNSIDE FILTER - precisely what the whole loser-detection line was looking for - and it is a mechanism rather than a curve fit: a bar still choppy and still turning at its close has not finished capitulating. ⛔ n = 7 / 8 / 12 in that second half. ⛔⛔⛔ WITHDRAWN THE SAME DAY, on the operator's instruction to rebuild the audit so it stops understating the search. A MAX-STATISTIC NULL - the entire search re-run on day-block-shuffled outcomes, including the pick-the-best-two-and-union step - puts the best real reading at +0.720pp against a noise median of +0.744pp: PERCENTILE 47.5, p = 52.4%, BELOW THE NULL'S OWN MEDIAN. TWO ERRORS COMPOUNDED: (1) the excess was measured against the FULL 374-signal mean (+0.609%) when the 15m rung's covered subpopulation returns +1.990%, so every 15m reading carried ~1.38pp of pure COVERAGE DIFFERENCE; (2) the per-reading nulls were correct for a SINGLE PRE-CHOSEN reading and meaningless for one FOUND BY SEARCHING - the search manufactures ~+0.74pp on noise alone. The +2.647% was almost entirely the population, not the shape. ⚠ ⛔ REGISTERED AS A CANDIDATE PRECISELY BECAUSE IN-SAMPLE REFINEMENT ON n=97 IS THE TRAP THAT HAS EATEN THIS PROJECT BEFORE. Parameters frozen (0.6% dip, >=2 dips, low in slot 13-16, close[16]>close[13]); NO sweep of any of them. Pass line fixed at >=120 covered signals across >=30 days, needing >=+0.6pp with a day-blocked interval excluding zero AND clearing a same-size random draw - AND the hold-out claim is scored SEPARATELY: the excess must be LARGER in the worse half than the better half, or it is not the downside filter it is registered as and it FAILS EVEN IF ITS AVERAGE PASSES. ⭐ THE POPULATION HOLE IS STATED: the 15m rung covers 97 of 374 signals (a coin needs a Kraken perpetual AND the signal must sit inside the rung's 21-day depth); a signal outside it is NOT JUDGEABLE, not rejected. The rung deepens a day per day, so coverage rises on its own. ⭐⭐ AND IT VINDICATES THE OPERATOR'S METHODOLOGICAL POINT: I had been testing thresholds on scalars and assuming monotone shapes. 'A lot of dips' was a countable thing nobody had counted. ⭐⭐⭐ THE LESSON, AND IT IS THE THIRD TIME IN ONE DAY THE OPERATOR'S PUSH HAS CORRECTED ME: A CAVEAT IS NOT A CORRECTION. I wrote 'the count understates the search' into the output and let the number stand; he told me to rebuild the instrument instead, and the rebuilt instrument reversed the conclusion. A stated caveat lets a wrong number keep its authority - the reader sees rigour and reads the number anyway. ⭐⭐ AND THE REBUILD FOUND A SECOND ERROR THE OLD AUDIT COULD NOT HAVE SHOWN: comparing a rung's readings against the WHOLE signal set measures the rung's COVERAGE, not its readings. Any per-rung instrument must compare within the rung. ⇒ The rule stays REGISTERED with frozen parameters but is DEMOTED to a mechanism hypothesis with NO in-sample support; only the downside-filter asymmetry (nothing in good periods, +2.3-3.2pp in bad ones) justifies leaving the clock running. |
| the_closing_2h_bar_and_the_grid 2026-09-05 · Operator 2026-09-05: 'we have found a system that works with 4h measures. reducing the measure will give better accuracy. we applied it wrong. do it one step at a time... go from 4hrs to 2hrs.' Then: 'the event is at the end of a 4hr bar... there are more signals happening outside of what we are finding in the 4hr window.' | ⭐⭐⭐ WHAT IS REAL IS A FILTER, NOT A MEASURE: does the CLOSING 2h HALF of the 4h bar independently qualify as a capitulation? KEEP 280 of 580 -> +0.660% vs K's +0.297%, PERCENTILE 99.7 (4 filters tested, adjusted 1.2%). ⭐⭐ AND THE MIRROR IS WHAT MAKES IT CONVINCING: if the OPENING 2h half qualifies instead, -0.079% at PERCENTILE 1.3 - the same mechanism read backwards, not a fifth draw of noise. A flush concentrated in the last two hours is fresh; one concentrated in the first two is a stale bounce we arrive late to. ✅ IT SURVIVES EVERY CHECK: NOT close_in_bar or vol_x (spearman -0.126 and +0.182) and the gap HOLDS INSIDE BOTH HALVES OF EACH (+0.381/+1.014 for cib, +0.518/+0.948 for vol_x); HOLD-OUT POSITIVE IN BOTH DIRECTIONS (+1.566pp first half, +0.203pp second); DAY-BLOCKED 95% CI on the gap +0.232...+1.201, CLEARS ZERO. ⭐⭐⭐ AND THE GRID QUESTION HAS A CLEAN AND COUNTERINTUITIVE ANSWER. A 4h grid shifted by 2 hours sees 674 signals, of which 416 are events the standard grid CANNOT SEE AT ALL. They return -0.304% against the standard grid's +0.297%, at PERCENTILE 0.1. THERE ARE 72% MORE EVENTS OUT THERE AND THEY ARE DECISIVELY WORSE. The grid is not hiding good signals, it is ACCIDENTALLY FILTERING OUT BAD ONES. ⚠ ⭐⭐⭐ THE MECHANISM, AND IT IS THE OPERATOR'S OWN DOCTRINE SHOWING UP IN THE CLOCK: the standard grid is anchored to 00:00 UTC - the same grid every exchange chart, every 8-hourly funding settlement and every other trader's 4h candle uses. A CAPITULATION THAT COMPLETES ON THAT BOUNDARY IS ONE THE WHOLE MARKET SAW AND RESPONDED TO; one that completes at 15:00Z resolves when nobody is collectively looking. THE 4h BOUNDARY IS NOT AN ARTEFACT OF OUR MEASUREMENT, IT IS A SCHELLING POINT. 'The trade is the outcome of peoples thoughts and actions' - including WHEN they are all looking. ⇒ This also strengthens the closing-2h filter rather than competing with it: what matters is the flush concentrating into the final two hours BEFORE a boundary the whole market watches. ⛔ NOT YET REGISTERED: 83 days, in sample, and the filter was selected from four. It needs its own pre-registration before it can become a panel arm - but unlike everything else tested today it has survived independence, a two-way hold-out, a day-blocked interval AND a mirror-image control. |
| the_moment_first_look_at_the_tape 2026-09-05 · Operator 2026-09-05: 'we need to get back to what is happening at that moment. that is how we find it. we are not paying attention to the real signals.' A FAIR CRITICISM OF THE WHOLE DAY: vol_x, range_x, close_in_bar, n_x, size_x, where the low sat, dip counts, open interest, funding, board breadth, grid offsets - every one is A STATISTIC ABOUT A BAR, and not one is what is ACTUALLY HAPPENING. The tape had been treated all day as a scheduled item for the regime question rather than read at the entry moment. | ⭐⭐⭐ THE DIRECTION IS BACKWARDS FROM INTUITION AND FORWARDS FROM OUR OWN DOCTRINE. Split at the median of each reading: buy_share_usd LOW half +1.702% vs HIGH half -2.188% (gap -3.890pp); buy_share_cnt +1.652 vs -2.138 (-3.790); flow_split +0.790 vs -1.275 (-2.065); trades_per_min +0.603 vs -1.088 (-1.691). SIGNALS WHERE PEOPLE WERE STILL SELLING INTO THE CLOSE RETURNED +1.70%; SIGNALS WHERE PEOPLE WERE ALREADY BUYING RETURNED -2.19%. avg_usd, big_share and market_share are flat (+0.467, -0.063, -0.255). ⭐⭐⭐ AND IT IS THE FOURTH INDEPENDENT MEASUREMENT OF ONE MECHANISM, the three others found earlier the same day on OHLC: close_in_bar <=0.25 (closes near its low, still being sold) · the CLOSING 2h bar also spiking (the flush concentrated at the end, percentile 99.7 with a mirror at 1.3) · the hour-1 penalty (already rallied three hours, bad). WE WANT THE CAPITULATION STILL IN PROGRESS, NOT ONE THAT HAS ALREADY TURNED - and the first three are shadows of that cast onto bars, while this is the thing itself. ⚠ ⛔⛔ n=32 AND NO NULL WAS RUN, DELIBERATELY. On 32 signals a null is theatre; the file refuses below 15 and says so. What makes this worth taking seriously is NOT the statistic but that THE DIRECTION WAS PREDICTED IN ADVANCE by four prior findings rather than discovered by searching - the difference between a prior and a fit. ⛔ The tape is the binding constraint and it resolves itself: 12 days today, growing one per day, and the same asset already gates the pre-registered buy/sell regime test (~2026-10-04). ⭐ THE PROCESS POINT THE OPERATOR MADE AND IT SHOULD OUTLIVE THIS ENTRY: a whole day of readings derived from bars had reached the wall, and the input that was never opened was the one describing WHAT PEOPLE WERE ACTUALLY DOING. The project's own doctrine says 'the trade is the outcome of peoples thoughts and actions' - and it spent the day measuring the residue of those actions instead of the actions. |
| wait_for_the_buyer_spike 2026-09-05 · Operator 2026-09-05, three corrections in sequence: (1) 'that point is definitely before a buy but it needs the buying to have a spike if it is going to go on a run - without that there isn't excitement to other buyers to get in'; (2) 'we don't want to buy without the buyer spike, and like we found in the K's it has to hold and build a bit before it will run again'; (3) 'exhaustion isn't safe. it could stay there forever if it doesn't have buyers and our money could be trapped.' | ⭐⭐⭐ THE WINNER/LOSER SPLIT IS THE MECHANISM HE PREDICTED, BEFORE ANY RULE WAS BUILT: winners' peak hourly buy spike is 0.96x the pre-entry rate and ARRIVES IN HOUR 3; losers' is 0.43x and arrives in HOUR 1. BIGGER, AND LATER - which is 'it has to hold and build a bit before it will run again' as a measurement. ⭐ MY FIRST TEST MISSED IT ENTIRELY because it split on spike SIZE alone and ignored TIMING (-0.328% vs -0.158%, nothing). ⭐⭐ THE RULE IS MONOTONE IN THE THRESHOLD: buy at the capitulation close -0.243% (n=32) · wait for a 0.5x spike -0.765% (n=16) · wait for 1.0x +0.627% (n=12) · wait for 1.5x +1.202%, 71% win (n=7). The harder the buying you demand, the better the outcome. ⛔⛔ AND REQUIRING THE PRICE LOW TO HOLD MAKES IT WORSE AT EVERY THRESHOLD (-1.146 / +0.078 / +0.336) - so 'hold and build' is about the BUYING, not about price refusing to dip; a spike that arrives with a further wobble is fine. ⭐⭐⭐ 20 OF 32 SIGNALS (62%) NEVER GET A 1.0x BUY SPIKE IN THREE HOURS. Bought at the close anyway they return -0.101% - not disasters, DEAD CAPITAL. That is the trapped-money population he named, it is the MAJORITY, and the spike requirement is what would exclude it. ⚠ ⛔⛔ n=7 AT THE TOP THRESHOLD AND SIX RULES WERE TRIED. This is far too small to conclude and no null was run, deliberately. What makes it worth pursuing is that the MECHANISM WAS STATED BY THE OPERATOR BEFORE THE MEASUREMENT and the threshold response is MONOTONE - the shape of something real rather than a cell found by searching. ⭐⭐⭐ AND THE COST OF WAITING IS ALREADY MEASURED AND NEARLY FREE IN THIS WINDOW: +1h costs 0.28pp, +2h 0.07pp, +3h 0.11pp, and only +4h falls off a cliff at 0.69pp. THERE IS A THREE-HOUR WINDOW IN WHICH TO WAIT FOR THE SPIKE AT ALMOST NO COST. That is what makes 'do not buy until the buyers show up' affordable, and it is the first time two separate measurements of this session have combined into an executable shape. ⇒ THE BINDING CONSTRAINT IS THE TAPE'S 12 DAYS. The fix is not to wait for calendar time but MORE SIGNALS PER DAY - dropping the ENTRY to a 1h bar inside the same tape window, which is the operator's own next instruction. |
| buyer_spike_at_1h_does_not_replicate 2026-09-05 · Operator 2026-09-05: 'go down to the 1hr reading once you get this to be solid.' The binding constraint on the buyer-spike rule was never the mechanism, it was that 12 days of tape hold only 32 FOUR-HOUR capitulations - so the fix was more signals per day, not more calendar days. | ⛔⛔ IT DOES NOT REPLICATE AND IT INVERTS. Unfiltered 1h population -0.191% (win 58%). wait for a 0.5x buy spike n=168 -0.663% PERCENTILE 5.6 · 1.0x n=82 -0.511% pct 20.8 · 1.5x n=58 -0.899% pct 8.9 · 2.0x n=37 -1.226% pct 6.5. EVERY THRESHOLD LANDS BELOW THE NULL - worse than entering late on a random subset - and there is no gradient; if anything it runs the wrong way. ⭐ THE 4h VERSION'S CLEAN MONOTONE CLIMB (-0.765 -> +0.627 -> +1.202) CAME FROM n=16/12/7, AND A MONOTONE RUN ACROSS FOUR THRESHOLDS ON SAMPLES THAT SMALL IS EASY FOR CHANCE TO PRODUCE. Occam: the 4h result was noise, rather than 'the mechanism is real at 4h and inverted at 1h'. ⭐⭐ WHAT DOES SURVIVE IS THE OPERATOR'S OWN OBSERVATION AND IT GETS STRONGER: 467 of 549 (85%) NEVER GET A 1.0x BUY SPIKE IN THREE HOURS, against 62% at 4h. The dead-capital population he named is not merely real, IT IS THE OVERWHELMING MAJORITY, and it worsens as the measure gets finer. ⚠ ⭐⭐⭐ THIS IS WHY THE 1h STEP WAS THE RIGHT NEXT MOVE EVEN THOUGH IT KILLED THE RESULT: the constraint was never calendar time, it was signals per day, and 17x the sample settled in one run what waiting weeks would have settled slowly. ⛔ THE STANDING LESSON, NOW FOR THE FOURTH TIME TODAY: a monotone response across thresholds LOOKS like mechanism and is not evidence at n<20 per cell - the shape-of-the-bar rules, the hour-of-low finding, and now the buyer spike all produced clean gradients that dissolved at larger n or under an honest null. ⭐ AND THE ONE THING THAT KEEPS SURVIVING EVERY EXPANSION is the CLOSING-2h FILTER (registry 122): 580 signals over 83 days, percentile 99.7 with a mirror at 1.3, independent of close_in_bar and vol_x, hold-out positive both ways, day-blocked CI +0.232...+1.201. It is the only reading of the day that has been tested at scale and held. |
| buying_spike_as_the_confirmation 2026-09-05 · Operator 2026-09-05: 'ok we are looking at the wrong thing. we need to find the buying spike and we already have the tool for that am I correct?' HE IS CORRECT AND IT WAS IN HAND ALL DAY: rule.is_spike() is vol>=2x AND range>=1.5x AND close in the BOTTOM 40%. A BUYING SPIKE IS THE IDENTICAL TEST WITH THE CLOSE IN THE TOP 40% - the mirror built this morning for the shorting test and then never used for this. And because the detector runs on BARS, it needs no tape: 83 days instead of 12. | ⭐⭐ THE RAW NUMBERS ARE THE BEST OF THE DAY: buy at the capitulation close +0.787% (n=206, 65% win) · control enter 1-6h later unfiltered +0.610% (n=1236) · buy the first BUYING SPIKE within 2h +1.376% (n=29, 79% win, pct 89.7) · within 3h +1.549% (n=38, 82% win, pct 96.7) · within 4h +1.445% (pct 96.3) · within 6h +1.357% (pct 96.2). STABLE ACROSS NESTED WINDOWS rather than one lucky cell, and nearly independent of what we already read (spearman with close_in_bar +0.068, with n_x +0.187). ⛔⛔ AND IT STILL DOES NOT CLEAR. The MAX-STATISTIC NULL - the whole four-window search re-run on day-block-shuffled outcomes - puts the real best excess of +0.939pp at PERCENTILE 79.2 against a noise median of +0.359 and a 95th of +1.566. The hold-out is flat-to-negative in the first half (-0.137pp on n=9) and +1.349pp in the second. ⭐ 146 of 206 (71%) never produce a buying spike in six hours, consistent with 62% (4h tape) and 85% (1h tape). ⚠ ⭐⭐⭐ THE WALL IS SAMPLE SIZE, NOT THE MECHANISM, AND THE NULL SAYS SO OUT LOUD: its 95th percentile sits at +1.566pp. A FILTER SELECTING 38 OF 206 SIGNALS HAS SO MUCH ROOM TO LOOK GOOD BY LUCK THAT NO SUCH FILTER CAN BE DISTINGUISHED FROM NOISE AT THIS n, HOWEVER SOUND ITS REASONING. That is a property of the test, not of the idea. ⛔ FIFTH PLAUSIBLE PATTERN TO DIE THE SAME WAY IN ONE SESSION - shape-of-the-bar, hour-of-low, buyer-spike-as-flow, buyer-spike-at-1h, and now buyer-spike-as-structure. EVERY ONE had a stated mechanism, a monotone or stable response, and a small SELECTED SUBSET of a few hundred signals. ⭐ The common shape is the finding: at ~200-600 signals, any rule that keeps a fifth of them is inside its own search's noise. ⇒ WHAT WOULD ACTUALLY SETTLE IT: not another reading, but SIGNALS. The registered forward panel accumulates them at ~3-5/day with nothing to tune, and the closing-2h filter (registry 122, 580 signals over 83 days, pct 99.7 with a mirror at 1.3) remains the only reading tested at a size where the null is narrow enough to answer. |
| anchored_on_the_buying_spike 2026-09-05 · Operator 2026-09-05: 'so we use the buying spike to run the k tests we developed after it has happened.' An INVERSION: anchor on the buying spike and make K's conditions the filter, rather than anchoring on K and looking forward for a spike. | ⭐⭐⭐ HIS SEQUENCE IS REAL AND IT IS THE MOST STATISTICALLY SOLID RESULT OF THE DAY. All buying spikes -0.432% (49% win). A CAPITULATION IN THE PRIOR 3h: n=1155, -0.104%, 55% win, PERCENTILE 100.0. Prior 6h: n=1705, -0.145%, pct 100.0. Prior 12h: -0.305%, 98.7. Prior 24h: -0.326%, 99.7. MONOTONE IN THE LOOKBACK - the fresher the capitulation, the better - with a PERFECT MIRROR: NO capitulation in the prior 24h returns -0.599% at PERCENTILE 0.5. Adding K's BTC gate to a <=6h capitulation gives the best cell, -0.075% on n=1240 at pct 100.0. ⛔⛔ AND EVERY SINGLE CELL IS NEGATIVE. The filter takes a losing population from -0.432% to -0.075% - a real and highly significant +0.357pp - and never reaches profitable. ⭐⭐⭐ THE COMPARISON THAT SETTLES THE DESIGN: entering at the CAPITULATION close returns +0.787% while entering at the BUYING SPIKE that follows it returns -0.075%. ENTERING ON THE SPIKE COSTS ABOUT 0.86pp. You pay for the spike's own move. ⚠ ⭐⭐⭐ THIS IS THE LARGEST-SAMPLE CONFIRMATION THE PROJECT HAS EVER HAD OF ITS MOST-REPEATED DOCTRINE: 'IT HAS ALREADY MOVED' ENTRIES ARE BACKWARDS - now at n=5,509 rather than the dozens the earlier confirmations rested on. The capitulation->spike SEQUENCE is genuinely informative (pct 100.0, monotone, mirrored at 0.5); acting on the SECOND event instead of the first is what destroys it. ⭐⭐ AND IT REPRODUCES THIS MORNING'S SHORTING TEST FROM A COMPLETELY DIFFERENT DIRECTION: blow-off LONG measured -0.359% there on 4h bars; buying spikes here measure -0.432% on 1h bars. Two independent measurements, two resolutions, same answer. ⭐ THE METHODOLOGICAL WIN IS THE ANCHOR CHOICE ITSELF: the same question asked with the rare event as the base population is unanswerable (206 signals, null 95th at +1.566pp) and with the common event as the base population is decisive (5,509 signals, percentile 100.0). WHEN A TEST CANNOT RESOLVE, CHECK WHETHER THE POPULATION CAN BE RE-ANCHORED ON THE COMMONER EVENT BEFORE CONCLUDING ANYTHING. |
| two_engines_flat_vs_hot 2026-09-05 · Operator 2026-09-05: 'so this is the first spike from a flat coin, right. that would be an exhausted coin that is getting ready to run. there is an overlap from a coin that has been hot like useless that we have to be aware of since that is a good buy right now at the right time since it is active.' Then: 'we need to split them like you said. both are valid but for different reasons' and 'maybe that is 2 completely different engines. one following the life of a coin that ran and it's second and 3rd wave and find the first leg of the run.' | ⭐⭐⭐ THE SPLIT IS DECISIVE AND IT INVERTS ONE HALF OF THE FRAMING. THE 'EXHAUSTED FLAT COIN WAKING UP' IS THE WORST POPULATION IN THE DATASET: first spike with none in the prior 24h n=1926 -0.659% at 42% win, PERCENTILE 0.1; FLAT beforehand (heat<1.0) n=1811 -0.748% at 40% win, PERCENTILE 0.0; the conjunction (first spike AND flat) n=1156 -0.760% at 39% win, PERCENTILE 0.1. ⭐⭐ THE 'ALREADY HOT LIKE USELESS' HALF IS CONFIRMED AND IS THE GOOD POPULATION: 3+ prior spikes n=1184 -0.132% at 62% win PERCENTILE 99.9; HOT beforehand (heat>1.5) n=1657 -0.043% at 62% win PERCENTILE 100.0; already UP >5% over the prior 24h n=1392 -0.124% at 67% WIN PERCENTILE 100.0; the conjunction n=924 -0.114% at 66% win PERCENTILE 99.8. ⛔ EVERY CELL IS STILL NEGATIVE - the best is -0.043% - so the split identifies WHERE the edge could live, not an edge. ⚠ ⭐⭐⭐ HIS TWO-ENGINE FRAMING MAPS ONTO THIS EXACTLY, WITH ONE SUBSTITUTION. ENGINE 1 (the first leg of a run) IS NOT THE FIRST BUYING SPIKE - that is the -0.760% population. THE FIRST LEG IS THE CAPITULATION ENTRY, CANDIDATE K, AT +0.787%: the engine already exists and is already on a clock. ENGINE 2 (the life of a coin that ran - its 2nd and 3rd waves) IS NEW: buying spikes in an already-hot coin, n=924, 66% WIN, percentile 99.8. ⭐⭐ AND ENGINE 2 HAS STRONG PRIOR SUPPORT ALREADY IN THE RECORD: the swing-persistence finding (morning sum-of-up-legs predicts the afternoon ranking, spearman +0.764, POSITIVE 10 OF 10 DAYS, strongest in the MOST LIQUID quintile) is precisely 'active coins stay active'. Two independent lines arriving at the same population. ⭐⭐⭐ AND ENGINE 2'S PROBLEM IS PRECISE AND ALREADY NAMED: at a fixed +3% target a 66% win rate with a -0.114% mean means THE 34% OF LOSERS AVERAGE ABOUT -6.2%. That is not an entry problem, it is the LOSS TAIL - the same loser-detection question as the rest of the day, but now inside a population that wins two thirds of the time, which is a far better place to be asking it than the 32%-loser population K produces. |
| engine2_buyers_won 2026-09-05 · Operator 2026-09-05: 'it has to come down to buyers and sellers who is winning, right?' then 'I would assume if the BUYERS are losing at this point it's a dead coin' — and 'see if you can refine that detection, check the other data factors.' | ⭐⭐⭐ ALL Engine-2 spikes -0.106% (66% win) -> + breakout 685 signals +0.486% (71%) -> + buyers-won band 341 +0.555% (74%) -> BOTH 274 signals +0.942% AT 79% WIN. Percentile 100.0 against a same-size random draw. HOLD-OUT POSITIVE IN BOTH HALVES at similar magnitude (+1.599pp and +1.286pp) - the check that killed the shape rules, the hour-of-low finding and the buyer-spike rule earlier the same day. Day-blocked 95% CI of the gap +0.207...+2.288, clears zero. ⭐⭐ THE cib QUINTILE TABLE IS U-SHAPED AND IT IS HIS DEAD-COIN CLAIM EXACTLY: 0.600-0.732 -> +0.017% (buyers barely won, the dead coin) · 0.832-0.918 -> +0.624% · 0.918-1.000 -> +1.200% at 81% win · EXACTLY 1.000 -> +0.397% (a single print at the high, not a won fight). ⭐ AND THE BTC GATE ADDS NOTHING (+0.937 with, +0.942 without) so it is excluded - a parameter that does nothing can only overfit later. ⚠ ⛔⛔ IN SAMPLE AND FITTED AT TWO POINTS: the 0.0349 breakout boundary is a quartile cut from this data and the 0.83 band edge was chosen AFTER seeing the quintile table. A max-statistic null over the whole day's search has not been run and would be the honest test. THAT IS WHY IT IS REGISTERED RATHER THAN ADOPTED. ⭐⭐⭐ THREE THINGS THE SAME SESSION MEASURED AND EXCLUDED, each with the number that excludes it: (1) BUYING THE PULLBACK - winners dip LESS (median -1.34% vs -2.29%) and a resting limit below the spike is monotonically worse (-0.365pp at 0.5% below, -1.873pp at 3%), because A LIMIT ONLY FILLS ON THE ONES THAT KEPT FALLING. Adverse selection on the entry itself. (2) wave_decay, since_last, ret24, abs_oi, funding each add <= +0.07pp inside the kept population. (3) deeper breakout horizons (72h/168h) are marginally better alone and add nothing on top. ⭐⭐⭐ AND THE SYMMETRY WITH ENGINE 1 IS THE DAY'S REAL SHAPE: for a CAPITULATION entry you want the SELLERS still winning at the close (the closing-2h filter, percentile 99.7, mirror at 1.3); for a BUYING-SPIKE entry you want the BUYERS to have won (percentile 100.0). SAME QUESTION - who is winning - OPPOSITE ANSWER DEPENDING ON WHICH SIDE YOU ENTER FROM. Both surviving filters of the session are the same reading. |
| quiet_tape_before_the_spike 2026-09-05 · Operator 2026-09-05: 'look for before the spike and not after it has happened. there is a tell that shows it. there has to be, it's mathematically impossible to not have some kind of change directly before the spike.' Then: 'and why are you looking at the candles, look at the indicators we built for this moment.' Then: 'it sounds like there are a few things that happen at the same time and then it moves.' | ⭐⭐⭐ THE TELL IS THAT IT GOES QUIET, NOT THAT BUYERS ARRIVE. large_trade_share shift -0.37sd (pct 95.2), thermometer shift -0.37sd (95.2), activity shift -0.31sd (84.8), volatility shift -0.26sd (63.7) — while EVERY buy-vs-sell reading is dead flat: flow_imbalance -0.09sd at percentile 0.2, count_imbalance 0.1, momentum 0.0. ⭐⭐ NAMED CONJUNCTIONS (five, declared from the mechanism): QUIET-3 (large trades gone + activity down + volatility down) DOUBLES the spike rate 9.7% -> 20.8%, lift +11.1pp at percentile 94.7 — ON the line, the null's 95th being +11.1pp. On MONEY a different arm wins: QUIET+BUYERS (+1.373pp day-matched, percentile 97.8, n=31 over 7 days) — the operator's original 'buyers winning' mechanism, which means NOTHING on its own and becomes the best money reading of the session inside a quiet tape. ⚠ ⛔⛔ THE SEARCH FOR A SECOND TELL FAILED AND THE FAILURE IS THE LESSON: 44 cuts -> 946 pairs, best pair lift +11.4pp, and the MEDIAN of the identical search on day-shuffled labels was +11.4pp (percentile 51). AT 66 SPIKES A PAIR SEARCH MANUFACTURES EXACTLY WHAT IT FINDS — which is why both arms are NAMED FROM THE MECHANISM and scored as five declared tests. ⛔ Two instrument corrections were needed to see any of this: (1) hourly bars CANNOT resolve a five-minute event — reporting 'the coin never cooled off' was reporting the resolution of the instrument; (2) the per-day-means statistic printed an EMPTY TABLE on a 12-day tape and had to be rebuilt as a within-day permutation test, because a test that silently returns nothing when the sample is small looks like 'no tell' and means 'no test'. ⭐ Registered, not adopted: clock 2026-09-05T21:00Z, thresholds frozen as NUMBERS in config/quiet_tape_arms.json (a tercile recomputed as data arrives is a rule that rewrites itself), cron at :40, rendered on the Experiments board. |
| the_tell_is_not_the_trade 2026-09-05 · Operator 2026-09-05: 'look at the point where we would want to buy and examine all the existing data points we have besides things like bitcoin that have nothing to do directly with that coin. there is something that says it is about to go for a second spike.' | ⭐⭐⭐ THE READINGS THAT PREDICT THE SPIKE LOSE MONEY, MONOTONICALLY AND IN OPPOSITE DIRECTIONS. vol_x quintiles: spike rate 5.5/5.5/9.1/14.3/21.2% while day-matched money runs +0.140/+0.170/+0.115/-0.052/-0.373pp. AUC 0.668 at within-day percentile 100 for the spike; percentile 100 REVERSED for the money. Same for range_x, ret3/ret6, dist_hi24, pullback, vol_trend. ⭐⭐ THE ONLY CLASS THAT PAYS IS POSITIONING, AND IT IS SILENT ABOUT THE SPIKE: oi_chg4 +0.328pp (99.8) at AUC 0.498; abs_oi_chg24 +0.277pp (99.5); oi_chg24 +0.215pp (95.0). ⭐⭐ AND THE MOMENT MATRIX PROVES THE OPERATOR'S NO-GLOBAL-RULES DOCTRINE ON THE COIN'S OWN CLOCK: ret6, dist_hi24, cib and abs_oi_chg24 all CHANGE SIGN across hours-since-last-spike (ret6 runs +0.096 at 1h to -1.400pp at 17h+), while vol_x and oi_chg4 never do. Best cell: 3-8h since the last spike AND oi_chg4 top quintile = 726 signals, 64 days, +0.478%, 67% win, +0.542pp day-matched, within-day percentile 100 — and its spike rate is BELOW base when volume has also cooled. ⚠ ⛔ ALSO CLOSED HERE: buying BEFORE the spike. The honest early rule (any hot bar closing >=0.75 above its weekly high) returns +0.952% and a DAY-MATCHED RANDOM HOT BAR on the same days returns +1.174% — excess -0.222pp, beats 0 of 40 controls. Its entire apparent edge was which days it stood in. ⛔ AND THE REGISTERED ENGINE-2 RULE DOES NOT CLEAR ITS OWN SEARCH: rebuilt from the prereg's prose (the original study was never committed and cannot be re-derived by anyone), the max-statistic null over 1,274 candidates puts breakout24+cib[0.83,1) at percentile 94.5 — inside. Its lead-in rival reaches 99.7 on the null but is NOT better on the day-matched control (+0.31pp vs +0.39pp): ~70% of its extra return is day selection. ⭐ THE NULL AND THE CONTROL ASK DIFFERENT QUESTIONS — a day-picking rule beats a day-block-shuffled null and dies to a day-matched control — and a rule must be told which business it is in. |
| thermometer_exit 2026-09-05 · Operator 2026-09-05: 'so we have the buy, what about the sell? the spike goes sometimes to 10% or 11% can we use the thermometers to read the exit?' and 'let's do a test to see if we can get a better exit; if we can't we will just start with the 3% plus round trip.' | ⛔⛔ DEAD, AND IT IS THE SEARCH-LEVEL NULL THAT KILLS IT. Three rules beat BOTH the +3% limit and their own random-exit control at 199-200/200 — large_trade_share<=-1.0 for 3m (+0.241% vs the limit's -0.188%, +0.428pp paired), count_imbalance<=-0.5 for 10m, flow_imbalance<=-0.5 for 10m. ⭐ BUT THE BEST-OF-48 UNDER THE SAME CONTROLS HAS MEDIAN +0.296pp AND A 95TH OF +0.397pp, WHILE THE REAL BEST IS +0.237pp — PERCENTILE 8.5, BELOW THE NOISE MEDIAN. At 97.5% per rule a 48-rule grid produces about one passing rule by chance; it produced three, and all three sit under what the search manufactures. ⭐ They were also 'hold 24h' in disguise: median hold 24.0h, i.e. the rules almost never fire, and most of their apparent gain over the limit is simply NOT TAKING the limit. ⚠ ⭐⭐ THE TWO PRE-SPECIFIED BASELINES ARE THE REAL OUTPUT, because neither was searched: on this population HOLDING THE FULL 24h (+0.057%) BEATS THE +3% LIMIT (-0.188%), reproducing the 2026-08-27 finding that the crudest rule is the best one. ⛔ But that population is EVERY hot-coin bar — an unselective entry with negative expectancy — not the registered entry, so it does not license changing K's exit. ⭐⭐⭐ AND THE PRIZE IS CONFIRMED LARGE AND UNREACHABLE: selling each trade at its own peak returns +4.603% on the same 681 trades. The operator's '10 or 11%' is really there; nothing we hold can find it in flight. ⭐ Cost was held at 0.681% for EVERY rule including the limit, though a resting limit would fill as maker at ~0.32% — the comparison was therefore biased AGAINST the baseline and the baseline still won. DECISION: start with +3% plus the round trip, as agreed before the test was run. |
| limit_level_sweep_and_ladder 2026-09-05 · Operator 2026-09-05: 'we should just buy and once it buys sell at 3% plus the round trip right when we buy' then, after seeing the sweep, 'can we do a range. 10%, 9%, 8%, 7% just to see what we get? it's just more lines to it right?' | ⭐⭐⭐ THE CURVE IS MONOTONE ALL THE WAY UP: 3% +0.229% · 5% +0.577% · 7% +0.808% · 8% +0.885% · 9% +0.965% · 10% +1.009% · 12% +1.101% · 15% +1.130%, against NO LIMIT AT ALL at +0.987%. ⛔ SO EVERY LIMIT BELOW ~10% IS A CAP THAT COSTS MONEY and the crossover where it stops costing is 10% on the full population and ~11% on the OI cell (10% +0.952% vs no-limit +0.963%). Fill rate falls 53% -> 21% from 3% to 10% and the median time to fill rises 7h -> 13h. ⭐⭐ THE LADDER LOSES AT EVERY SETTING: 7/8/9/10 +0.917% · 8/10/12 +0.998% · 5/10/15 +0.905% · 3/6/9/12 +0.752%, against a SINGLE LINE AT 10% at +1.009%. The cost is monotone in how low the lowest rung sits — the slices sold early are exactly the trades that kept going. ⚠ ⭐⭐ THIS INDEPENDENTLY REPLICATES REGISTRY 107 (scaling out is dead) AT A COMPLETELY DIFFERENT PLACE ON THE CURVE. That result was measured banking half at +1.5% under a 3% target and explained by 'the target already makes the gain certain, there is no right tail to scale into'. Up at 7-12% the right tail is precisely what is being harvested, and selling part early still loses — so the mechanism is more general than the original explanation. ⛔ A FEE BUG WAS CAUGHT BY A DEGENERATE CASE: the first ladder charged the round trip on slices whose trigger already contained it, taxing filled slices twice and making every ladder look ~0.5pp worse. It was caught because the '10 only' ladder disagreed with the sweep's own 10% row. ⭐ A degenerate case that must reproduce an existing number is the cheapest test available and it is now permanently in the table. ⛔ ALL OF THIS IS IN SAMPLE ON ONE REGIME (BTC never closed >= +3% in the window), and the registered exit therefore settles THREE ways forward — +10% attached, +3%, and the pure clock. |
| ride_or_sell_at_ten 2026-09-05 · Operator 2026-09-05: 'yeah but there has to be a way to read the exit if it goes up to as high as 15%.' | ⛔⛔ THE FIRST CUT SAID YES AND WAS WRONG. Against a fixed +10.0% baseline, ret_since_entry_close's top quintile made riding worth +3.151pp (percentile 100.0, ABOVE the max of 4,000 null draws), touch_cib +2.596pp (99.7), accel +2.294pp (97.7). ⭐⭐⭐ BUT THOSE READINGS ARE ONLY KNOWABLE AT THE BAR'S CLOSE, BY WHICH TIME A RESTING +10% LIMIT HAS ALREADY FILLED INTRA-BAR — so the ride arm was credited with starting from a price the sell arm was not. Priced at the same point (decide at the first hourly CLOSE above the trigger): SELL +12.193% vs RIDE +11.592%, riding costs -0.601pp, and the best reading INVERTS — peak_so_far's top quintile now -2.829pp at percentile 100. Raising the limit to +15% after the touch returns +12.105%, still below selling. ⚠ ⭐⭐⭐ THE DEFECT IS THE SAME ONE THIS PROJECT KEEPS FINDING IN NEW COSTUMES: two arms measured at different points. A fixed +10.0% for one side and an actual price for the other is not a comparison, and it manufactured a percentile-100 result. ⭐ ALSO SETTLED, on all 8,684 entries: a RESTING order at +10% returns +1.009% and CHECKING THE CLOSE returns +1.091% — the close check is +0.08pp better and gives up an exit that lives on the exchange and survives this machine dying. The registered attached limit stands. ⛔ Third independent confirmation in one session that nothing we hold reads the exit: stateless price rules (August), flow rules across the hold (exit_lab, percentile 8.5), and now a single reading at the one moment that matters. |
| exit_surge_direction_and_portfolio_shape 2026-09-06 · Operator 2026-09-06: 'figure out how to take it at a peak. I bet the one that peaked at 30% we could have sold and there was another spike incident we could have bought on... we might get 70% on one that moves that much in a day from multiple trades.' Then: 'the ones that are the losers, just one coin pays for all of them... we could be in profit everyday if this works.' | ⭐⭐ HIS MULTI-TRADE INTUITION IS ALREADY WORKING AND THE CASE PROVES IT: COTI moved +19.42% first price to peak, and the registered rule BANKED +20.0% net across two separate filled positions on the same coin in 12 hours - more than the move itself, because the double-buy rule ALLOWS re-entry once the limit has filled. That is the second-and-third-leg design paying out exactly as described. ⛔⛔ BUT THE EXIT STILL HAS NO READING. The case study exposed a real gap in my own search - the 48-rule grid only ever tested 'exit when the flow FADES' (<=), never 'exit when it SURGES' (>=), and at COTI's peak minute large_trade_share read +0.74 and flow_imbalance +0.56 against a day mean of 0.00. A SEARCH THAT CANNOT EXPRESS THE HYPOTHESIS RETURNS A CONFIDENT NO. Rebuilt to 96 rules across both directions: the surge rules look excellent per-rule (large_trade_share >= +1.5 for 3m in profit: +0.705% vs the limit's +0.057%, beating 200 of 200 of its own controls) and the SEARCH-LEVEL null still kills them - best-of-96 in noise has median +0.341pp and a 95th of +0.464pp against a real best of +0.333pp, PERCENTILE 40.0, below the noise median for the second time. ⚠ ⭐⭐⭐ THE PORTFOLIO SHAPE IS THE REAL ANSWER AND IT SPLITS HIS CLAIM IN TWO. THE CONCENTRATION IS CONFIRMED, EXTREMELY: on 53 deduped entries the top 3 trades carry 66% of the total and the top 5 carry 110% - everything else is net negative, exactly 'one coin pays for all of them'. ⛔⛔ BUT 'IN PROFIT EVERY DAY' IS REFUTED BY THE SAME NUMBERS: only 4 of 9 days are green, the median day is -0.6%, and THE BEST DAY ALONE IS 95% OF THE ENTIRE RESULT. Convexity does not produce steady days; it produces mostly-red days and a rare day that carries the year. ⭐ And on the SAME trades the pure 24h clock returns +1.149% against the registered +10% exit's +0.856% and the +3% limit's +0.179% - while the +3% limit buys MORE GREEN DAYS (5 of 9) for less money, which is the honest statement of that trade-off. ⛔ All of it in sample on arms chosen in this window. |
| daily_gate_vs_selectivity 2026-09-06 · Operator 2026-09-06: 'I bet we can figure out the 10, 5 or even 6 so we have a buffer and then drop the other coins for the rest of the day unless we get a massive buying indication.' Then, correcting himself: 'we shouldn't stop for the day once we have a winning group of trades. if anything we want to keep winning if we can keep finding winning trades.' | ⭐⭐⭐ HIS OWN CORRECTION IS CONFIRMED BY THE DATA AND IT IS NOT CLOSE. Baseline (no gate) 46 trades, +49.8%, 6 of 10 days green. EVERY stop-win loses money — +3% -> +22.0%, +5% -> +28.6%, +6% and +8% -> +17.6%, +10% -> +48.2%, +15%/+20% -> +39.8% — and NOT ONE of them buys a green day: the tight budgets make it WORSE, 6/10 -> 5/10, because stopping early leaves a marginal day short of the trades that would have turned it. The both-arms override recovers some and never beats the baseline. ⛔ THE DAILY STOP-LOSS IS INERT: -10/-15/-20% are byte-identical to the baseline because no day's REALISED P&L ever fell that far intraday - the damage arrives through positions opened before the day turned, which a daily gate cannot see. ⭐⭐⭐ THE LEVER IS SELECTIVITY, NOT A GATE: QUIET-3 ALONE returns +59.1% on 44 trades against +49.8% for taking everything, while BOTH ARMS TOGETHER — the strongest thing the system can say — is the WORST population at -4.8% on 14 trades, and QUIET+BUYERS alone manages +6.4%. ⚠ ⛔⛔ THAT INVERTS LAST NIGHT'S READING OF QUIET+BUYERS, which looked like the money arm (+1.373pp day-matched, percentile 97.8, on 31 signals). The two measurements are not the same measurement: that one was a per-signal day-matched excess over ALL firings including double buys; this one is a realised book, deduped one-position-per-coin, on the registered exit, accounted by the day a position CLOSES. Both are in sample on ten days and they disagree — which is exactly why both arms are on a forward clock and neither is adopted. ⛔ TWO BUGS OF MY OWN WERE FOUND AND FIXED IN THIS FILE BEFORE ANY NUMBER ABOVE WAS BELIEVED: the budget loop's `break` fired on BUDGETS[0]=None so not one budget was ever tested while the table printed a single baseline row as if that were the answer; and reading `realised[d]` on a defaultdict CREATED empty days, turning 10 days into 12 and moving mean/day from +4.98% to +4.15% with no trade changing — a lookup with a side effect, reported as a result. |
| sizing_slots_and_picking_the_winner 2026-09-06 · Operator 2026-09-06: 'the real issue is size of the trade. we make smaller trades so we can capture more opportunities. it's fine to go down to $20 a trade at first if it makes sure we get more of the big winners.' And: 'is there a way to find the big winner out of the crowd of good coins that day? I bet there is a way to find the top 3 early.' And, stated BEFORE seeing any of the ranking output: 'I bet we can detect at least one of the top 3 after the 1st spike everyday. it will have a signature that is different.' | ⭐⭐⭐ HIS SIZING POINT IS RIGHT ABOUT COVERAGE AND WRONG ABOUT RETURN, AND THE TRADE IS THE INTERESTING PART. At ~$750 exposure: 3 slots x $250 takes 24 trades, MISSES 41, returns +4.6% at 6.0% maxDD - and goes to -0.9% without its best day. 12 slots x $60 takes 45, misses 2, returns +1.8% at 1.8% maxDD - and STAYS POSITIVE at +1.1% without its best day. 20 x $40 misses NOTHING, +1.4%, maxDD 1.2%, +1.0% without its best day. ⛔ SO THE +4.6% AT THREE SLOTS IS NOT SKILL, IT IS WHICH 24 OF 65 SIGNALS HAPPENED TO FIND A FREE SLOT. Smaller tickets convert a one-day fluke into a smaller distributed edge and cut the drawdown fivefold. ⛔⛔ AND NOTHING PICKS THE WINNER: the best of 27 readings (prior_spikes, top-3 money +46.6% against a random-3 median of +8.9%, per-reading percentile 98.9) lands at PERCENTILE 67.5 against a null that replays the whole 27-reading search - best-of-27 in noise has a median of +43.4%. ⚠ ⭐⭐ THE ONE THING WORTH CARRYING FORWARD IS HIS BLIND CALL. He proposed 'the top-3 shows itself after the 1st spike, with a different signature' BEFORE any ranking output was shown to him, and the data's best reading of 27 is exactly that variable - prior_spikes, the wave number. The wave structure: wave 3 (the first) n=24 mean -0.09% at 46% win carrying -5% of the total; WAVE 4 n=13 mean +2.29% at 62% win carrying 60% OF EVERYTHING; wave 5 n=7 mean +0.34% at 29%. ⛔ BUT THE RULE HE NAMED FAILS ITS OWN SINGLE TEST: taking the day's DEEPEST-wave firing returns +3.9% against a random pick's median +2.3%, percentile 54.8, inside. And wave 4 is one bucket of 13 trades that is not monotone with wave 5 - the exact shape of 'one event wearing a factor's clothes' that this project has killed five times. It is a hypothesis to REGISTER and let the clock judge, not a rule to adopt. |
| no_progress_cut_and_the_arms_refactor 2026-09-06 · Operator 2026-09-06: 'when to get out of the other short runners earlier so we don't lose as much' and 'avoid that small draw down trade group that bleeds us with the fees.' | ⛔⛔ IT DOES FREE SLOTS AND IT LOSES MONEY ANYWAY, AT EVERY SETTING AND EVERY SLOT COUNT. At 3 slots: no cut 24 trades / 42 missed / +4.6%; cut at hour 2 takes 36 trades and misses only 28 - the mechanism works exactly as predicted - and returns +2.4%. Hour 4 under +1% cuts the misses from 42 to 12 and returns -1.5%. Every deeper cut is negative. At 8 slots the best cut (+1.6%) is below no cut (+2.7%). At 20 slots, where nothing is missed and there is no capacity to buy, every cut is worse than none (+1.7% vs -0.2% to +0.6%). ⭐ THE EXTRA TRADES A CUT BUYS ARE NOT WORTH WHAT THE CUT COSTS, and the reason is already in the record: among trades that dip badly early, as many recover as die. ⚠ ⭐⭐⭐ THE REFACTOR THAT CAME OUT OF IT MATTERS MORE THAN THE RESULT. Registering a third arm broke daily_budget.entries with KeyError: 'prior_spikes' because it held its own copy of the condition loop and the new arm's threshold is a literal count rather than a frozen percentile. THE DUPLICATE DID NOT SURVIVE ITS FIRST REAL CHANGE. A check then found the same loop in case_study and portfolio_day - four copies in total, and had the new threshold been a percentile like the others they would all have silently EXCLUDED the new arm while still printing full-looking tables. ops/arms.py is now the only place a registered condition becomes a decision, and its self-check PARSES the syntax tree rather than grepping - the text version fired on a docstring that merely described the bug, the second false positive of the day from a source-scanning test after `still_open` matched a local variable. |
| peak_reaction_trail_and_the_label_argument 2026-09-06 · Operator 2026-09-06: 'there is definitely a way to tell a peak we just didn't figure it out yet... if we can get closer to the peak, or SELL RIGHT AFTER A PEAK then we are going to be able to capture way more profit.' And: 'I bet once we start doing peak reaction we will eventually get peak prediction.' | ⛔⛔ DEAD ON BOTH NULLS. Per rule it is the best exit result this project has produced: 'trail 3% from the high once +5%' returns +0.555% against the +3% limit's +0.062% (+0.493pp) and beats 200 of 200 of its own random-exit controls. THE FAMILY-LEVEL NULL - the fair one - has a best-of-30 median of +0.235pp against a real best of +0.230pp: PERCENTILE 47.0, which is not merely inside the noise, it IS the noise median. The whole-grid null agrees at percentile 24.5. ⭐ FIFTH EXIT FAMILY KILLED: price stops (August), flow fading, flow surging, a single reading at +10%, and now the trail. The ceiling is unchanged and enormous - selling each trade at its own peak returns +5.285% against the registered exit. ⚠ ⭐⭐⭐ WHAT SURVIVES IS HIS SECOND SENTENCE, AND IT IS A DATA ARGUMENT RATHER THAN A TRADING ONE. Peak PREDICTION needs labelled peaks to learn from and this project has exactly one - the COTI case study, n=1, seven readings, which is precisely why that look could never be more than a hypothesis. A trail does not predict anything but it MARKS where the peak was, and every marked peak is a training label. So the label is now recorded on every settled position regardless of whether the trail is ever traded: peak %, the bar it occurred in, hours-to-peak and how much was given back, written to peak_labels.jsonl. Collecting it costs nothing today and is the only thing that makes the question answerable in a month; not collecting it means asking the same unanswerable question again with the same n=1. ⛔ AND ONE CORRECTION TO THE OPERATOR'S NEXT CLAIM: a trail does NOT produce more trades here. The best cell's median hold is 24.0h, unchanged, and at 20 slots every signal is already taken - the data rate is set by the SIGNAL rate (~17.6 firings/day), not by how fast positions close. |
| sizing_frontier_robust_not_optimal 2026-09-06 · Operator 2026-09-06: 'the sizing question is a big one so yes, we are going to work on that next.' | ⭐⭐⭐ CONCENTRATION LOOKS BEST AND DIES FIRST. 3 slots x $250 is the top row at the observed edge (median +58.5%, 5th pct +24.2%) and at NO EDGE it has a 5th percentile of -34.0% with a 46.0% chance of a drawdown worse than 20%; at the wrong sign, median -32.2% and 89.3%. 8 slots x $100 is worse still in the no-edge case (median -11.5%, P(dd>20%) 33.1%). ⭐⭐ 20 SLOTS AT $40 IS THE ONLY FAMILY THAT HOLDS ACROSS ALL FOUR: observed +19.0% median with a +10.7% 5th percentile; HALF the edge +9.5% with a 5th percentile STILL POSITIVE at +1.5%; NO edge 0.0% median, -7.4% at the 5th, ZERO probability of a 20% drawdown; wrong sign -9.6% median and -16.6% at the 5th. 40 x $20 is the same shape with less of everything. Risk of ruin is 0.0% everywhere - the account cannot be destroyed by this rule, only ground down. ⚠ ⛔ THE 5TH PERCENTILE IS THE COLUMN THAT DECIDES, NOT THE MEDIAN. The median is what a policy does on a good run; the left tail decides whether the account is still there to keep running it. ⛔⛔ AND THE 'NO EDGE' BLOCK IS THE BASE CASE RATHER THAN THE PESSIMISTIC ONE - ten days whose result was one day cannot distinguish a real edge from none, which is exactly why the forward clock exists. ⭐ The practical reading: at $1,403.85 the difference between 3x$250 and 20x$40 is not return, it is whether being wrong costs 7% or 34%. Until the forward test says otherwise, the small ticket is not a conservative choice, it is the only one whose downside is bounded by something we have measured. |
| when_is_the_days_winner_knowable 2026-09-06 · Operator 2026-09-06: 'how long does it take to determine who is the big winners for that day? that is an important metric that we didn't work on... when can we first pick which ones they are?' | ⭐⭐⭐ THE EARLY LEADERS ARE NOT THE WINNERS, AND AT HOUR ONE THEY ARE WORSE THAN RANDOM. Hour 1 catches 1.01 of 5 against a random-pick expectation of 1.25 on a 20-coin day; hour 2 catches 1.29, which is chance. The ranking only clearly beats chance from hour 3 or 4. ⭐⭐ AND THE WINNER'S MOVE STARTS LATER THAN EXPECTED: the eventual top coin first clears a quarter of its own move at hour 6 (median), 7.7 mean. ⭐⭐⭐ THE TRADE-OFF IS MONOTONE IN BOTH DIRECTIONS AND THE PRODUCT IS FLAT: accuracy climbs 20% -> 43% -> 73% -> 99% across hours 1, 8, 16 and 23 while the move still ahead falls 42% -> 18% -> 9% -> 1%, so EXPECTED CAPTURE sits between 8.3% and 9.1% for the whole of hours 1-7, decays gently to ~7% by hour 13, and collapses after hour 16 (1.7% by hour 22). ⚠ ⛔⛔ THIS ENTRY WAS WRONG WHEN FILED AND IS CORRECTED HERE RATHER THAN DELETED. It claimed the early leaders are 'worse than chance' at hour 1 because it compared 1.01 of 5 against a random expectation of 1.25 computed for a 20-COIN day. THE DAYS CARRY ~94 COINS. Random is 5x5/94 = 0.27, so hour 1 is FOUR TIMES chance and every row in the table is several times chance. The 20 came from a `len(rank) < 20` guard - a MINIMUM read as if it were the actual count. ⭐ The operator rejected the conclusion on instinct ('there has to be a mathematically traceable signal') and was right. What survives from the entry is the TIMING shape: accuracy rises and remaining upside falls, and their product is flat at 8-9% through hour 7 before decaying - so waiting costs nothing until about hour 8 and a great deal after 16. |
| early_detection_of_the_days_top_five 2026-09-06 · Operator 2026-09-06: 'did you examine it with all of the metrics we are using. There has to be a mathematically traceable signal that shows which ones are in the top 5... I guarantee it is something we can find with plenty of time to make the first trades right before it happens.' | ⭐⭐⭐ HE WAS RIGHT AND MY PREVIOUS CONCLUSION WAS WRONG. Random catches 0.26 of 5 on a 96-coin day. `hi_h` (the day's high so far against its open) catches 1.29 at hour 2, 1.65 at hour 4, 1.87 at hour 6 and 2.15 at hour 8 — FIVE TO EIGHT TIMES CHANCE — with AUC climbing 0.694 -> 0.799. SEARCH-LEVEL NULL over all 15 metrics x 4 hours: best-in-noise median 0.40, 95th 0.45, MAX 0.54, against a real best of 2.15. PERCENTILE 100.0, and four times the noise MAXIMUM. This is the first thing in the project to clear a search-level null by a margin rather than a hair. ⭐⭐ AND IT IS NOT ONE METRIC: ret_h (AUC 0.739 at hour 8), range_h (0.767), max_bar_range (0.772) all separate strongly; usd_x and range_x — today against the coin's OWN yesterday — reach 0.626 and 0.623 and are a different fact from the raw move; abs_oi_chg24 holds 0.58-0.60 at every hour and is the only NON-price reading that works; and ret_prev (yesterday's return, AUC 0.563, 0.85 of 5) is knowable BEFORE THE DAY STARTS. ⚠ ⛔ WHAT THIS DOES NOT SAY. It identifies the day's biggest MOVERS, not profitable entries — the move it names is partly already spent (at hour 2 about 34% of it remains, at hour 8 about 18%), and five exit families tested today say we cannot sell at the top. Detection and profit are different claims and the second is not established by the first. ⭐ WHAT IT CHANGES: the arms currently fire on a coin-local condition and ignore the cross-sectional question entirely. A day-level top-5 screen is a genuinely new input to them, it is available from hour 2, and unlike everything else tested today it has 82 days behind it rather than ten. |
| volume_forms_and_the_sided_tape_for_top5 2026-09-06 · Operator 2026-09-06: 'what about buyers vs sellers? what about volume? there is a factor that will show this.' Then: 'the volume should be in comparison to the other coins that are moving. perhaps what percentage of the money that is moving that day that each of the coins get.' | ⭐⭐⭐ THE FORM OF THE QUESTION DECIDES THE ANSWER, AND ABSOLUTE READINGS ARE WORTHLESS. usd_h AUC 0.533 catching 0.16 of 5 — BELOW random's 0.26. usd_share, the operator's share-of-the-day's-money, is the same: AUC 0.534, 0.16 of 5, also below random, because the big coins always take the biggest share and always have. ⭐⭐ BOTH SELF-RELATIVE FORMS WORK AND THEY ARE THE SAME FACT: usd_x AUC 0.626 catching 0.65 of 5, usd_share_x AUC 0.624 catching 0.65 — numerically identical, because 'busier than itself' and 'taking more of the day than it usually takes' are one reading with two names. range_x agrees at 0.623 and catches 0.86. ⛔ AND THE SIDED TAPE IS WEAK: on 11 days buy_share reaches AUC 0.596 and catches 0.36 of 5 — better than chance, worse than every self-relative volume reading and far worse than price. buy_share_shift, trade rate and large-trade share are all at or under 0.59. ⭐ PRICE DOMINATES EVERYTHING: hi_h AUC 0.799 catching 2.15 of 5, ret_h 0.739 catching 2.05. ⚠ ⛔ THE HONEST CAVEAT ON THE WINNER: `hi_h` is the day's high so far and the target is the day's eventual high, so part of its power is the target's own autocorrelation — the best early detector of 'will be a big mover today' is largely 'is already a big mover today'. That is legitimate (it is knowable at the hour it is read) but it is a weaker claim than a mechanism, and the non-price readings are what would have made it one. ⭐ THE USEFUL RESIDUE: abs_oi_chg24 holds AUC 0.58-0.60 at EVERY hour and is the only non-price reading that does; ret_prev (yesterday's return, AUC 0.563) is knowable before the day opens; and the self-relative volume family is genuinely independent of the price family in form even if it is weaker. A screen built from price alone would be a restatement of the target; one built from price PLUS positioning PLUS self-relative volume would not. |
| top5_detection_at_the_second_spike 2026-09-06 · Operator 2026-09-06: 'I'm sure we can detect the big 5 movers from the pack within 30 minutes of it making at least its second spike. let's focus finding that indication at that moment. I think that is the next best time and place to look.' | ⭐⭐⭐ 16.2% OF SECOND-SPIKERS FINISH IN THE DAY'S TOP 5 (236 of 1,458). The top decile by RETURN SO FAR AT THE SPIKE finishes top-5 58.2% of the time — a lift of +42.0pp and 3.6x the base rate — with AUC 0.827; hi_so_far is 57.5% at AUC 0.851. SEARCH-LEVEL NULL over all 20 readings with top-5 labels shuffled inside each day: best-in-noise median 22.6%, 95th 26.2%, MAX 30.3%. Real best 58.2% ⇒ PERCENTILE 100.0, nearly DOUBLE the noise maximum. ⭐⭐ AND THE SPIKE'S OWN SHAPE MATTERS INDEPENDENTLY: first_spike_size top decile 46.6% (+30.4pp, AUC 0.747) and spike_size 45.2% (+29.0) — how big the first leg was is nearly as informative as the whole day's move. ⭐ hours_since_first INVERTS (AUC 0.396): the SOONER the second spike follows the first, the more likely a top-5 finish. ⚠ ⛔ THE AUTOCORRELATION CAVEAT IS REDUCED, NOT ELIMINATED. ret_so_far and hi_so_far still measure how far the coin has moved and the target is how far it eventually moves; conditioning on a second spike equalises the fact of moving but not the amount. The readings that are genuinely independent of price are the positioning family — oi_chg4's top decile is 30.1% (+14.0pp) and abs_oi_chg24's is 28.8% (+12.6) — and both sit at or just above the noise 95th of 26.2% while below its max of 30.3%, so they are suggestive and not established. ⭐ THE PRACTICAL SHAPE: this is a SCREEN for the existing arms rather than a new entry. The arms fire on a coin-local condition and ask nothing about the field; taking a firing only when the coin is in the top decile of return-so-far would be a cross-sectional filter on an existing registered rule, and it has 79 days behind it rather than ten. |
| top5_detection_at_the_first_spike 2026-09-06 · Operator 2026-09-06: 'so did we narrow down a way to truly tell which ones turned into the top 5 after the FIRST spike?' | ⭐⭐⭐ YES, AND EARLIER IS ONLY SLIGHTLY WEAKER. At the FIRST spike 9.8% of coins finish in the day's top 5 (320 of 3,278); the top decile by hi_so_far finishes top-5 42.4% of the time — a lift of +32.6pp and 4.3x the base rate — at AUC 0.839, with ret_so_far at 42.1% and AUC 0.824. SEARCH-LEVEL NULL over all 17 readings: best-in-noise median 12.8%, 95th 14.8%, MAX 17.8%, against a real 42.4% ⇒ PERCENTILE 100.0 at 2.4x the noise maximum. ⭐⭐ THE COMPARISON THAT ANSWERS THE QUESTION: first spike 9.8% -> 42.4% (4.3x lift), second spike 16.2% -> 58.2% (3.6x lift). Waiting for the second spike buys higher absolute precision from a higher base; the MULTIPLE is actually better at the first. ⭐⭐⭐ AND THE SPIKE'S OWN SIZE IS THE READING THAT NEEDS NO WAITING AT ALL: spike_size top decile 34.5% (+24.7pp) at AUC 0.768, measured on the spike bar itself. ⚠ ⭐⭐ A CLEAN SPLIT INSIDE THE SPIKE ITSELF: the spike's PRICE MOVE predicts (spike_size AUC 0.768) and the spike's VOLUME does not (spike_vol_x AUC 0.504, top decile 10.4% against a 9.8% base — nothing). The range multiple is weak too (0.589). So it is not how much traded, it is how far it went. ⛔ The autocorrelation caveat still applies to hi_so_far and ret_so_far and applies LESS to spike_size, which is one bar's own move rather than the day's accumulated one. ⭐ Positioning again holds a small independent signal: abs_oi_chg24 17.1% (+7.3pp) against a noise 95th of 14.8% — above it, below the max of 17.8%, so suggestive and not established, exactly as at the second spike. |
| preday_screen_before_the_first_spike 2026-09-06 · Operator 2026-09-06: 'is it possible to detect them before the first spike and get rid of at least some of the other coins that turn out to be false runners?' | ⭐⭐⭐ YES, AND THE MECHANISM IS VOLATILITY CLUSTERING. Base rate 5.2%. `atr_prev` — yesterday's average bar range — has AUC 0.697 with a TOP decile of 15.0% (+9.8pp, 2.9x the base) and a BOTTOM decile of 1.3%. `days_since_big` (how recently the coin last had a 10% day) inverts at AUC 0.283, top decile 13.6%, bottom 1.8%. `dist_lo168` (how far above its own weekly low it opens) AUC 0.672, top 12.1%, bottom 2.0%. SEARCH-LEVEL NULL over all 17 readings with labels shuffled inside each day: best-in-noise median 6.3%, 95th 7.2%, MAX 8.4%, against a real 15.0% ⇒ PERCENTILE 100.0, nearly double the noise maximum. ⭐⭐ AND THE OPERATOR'S ACTUAL ASK — GETTING RID OF COINS — IS THE STRONGER HALF: dropping the bottom decile by yesterday's volatility takes the top-5 rate from 5.2% to 1.3%, a FOUR-FOLD reduction, before a single trade is placed. ⚠ ⭐⭐ THE THREE STRONGEST PRE-DAY READINGS ARE ALL THE SAME FACT IN DIFFERENT CLOTHES: yesterday's range, days since the last big day, and position within the weekly range are all measuring whether this coin has been MOVING lately. Volatility clusters; quiet coins stay quiet. That is a known market property and it arrives here cleanly rather than as a surprise, which is a point in its favour — it is the first result in this project with a textbook mechanism behind it rather than a pattern. ⛔ It is a WATCHLIST, not an entry: 15% of the top decile finishing top-5 still means 85% do not, and the screen says nothing about when or whether to buy. Its value is the bottom decile, where four coins in five hundred become one. |
| full_sweep_candles_vs_tape_same_days 2026-09-06 · Operator 2026-09-06: 'let's do the sweep of indicators we have and include the ones we only have 12 days of data on. this might take more time to get and it's probably going to be the AI getting smarter path we talked about that will find this over time.' | ⛔⛔ THE TAPE ADDS NOTHING AT EITHER ANCHOR, ON ITS OWN DAYS. At the FIRST spike the best tape reading (ret_last5, 16.3%) lands EXACTLY ON THE NULL MEDIAN of 16.3%, while ret_so_far reaches 36.0% against a base of 10.5% and clears the 29-reading null at percentile 100. At the SECOND spike the best tape reading (momentum_5, 39.5%) sits above the null median of 31.6% but below its 95th of 36.8%... marginally, while ret_so_far reaches 68.4% against a 19.6% base. In both tables every one of the top five readings is a candle reading. ⭐ The candle results REPLICATE on the small sample at higher AUC than on the full one (hi_so_far 0.863 and 0.911 here against 0.839 and 0.851 on 79 days), which is what a real effect looks like when the sample shrinks and nothing else changes. ⚠ ⭐⭐⭐ AND THE OPERATOR'S OWN FRAMING IS THE CORRECT READING OF THIS, NOT A CONSOLATION. Thirteen days carry roughly 60 top-5 slots. A reading needs to be very strong to clear a search-level null over thirty candidates on that much data, so THIS RESULT MEANS 'NOT VISIBLE YET' AND NOT 'NOT THERE' — and the study says so before it prints, rather than after. The tape found the ENTRY tell (the quiet before a spike, percentile 95.2) on this same window, so it is not a dead instrument; it is a thin one, and it thickens at ~17 events a day on its own. ⛔ A DEFECT FIXED MID-RUN: the coverage floor was a flat 200 rows, which was silently fine at anchor 1 (430 rows) and emptied the candidate set at anchor 2 (189), where the study CRASHED with `max() iterable argument is empty` instead of refusing. A study too thin to run must say so; crashing is the same failure with worse manners. |
| the_first_fill_is_a_tell_for_the_rest_of_the_day 2026-09-06 · Operator 2026-09-06, watching RAY fill at +10% and then lose 9.6% on its re-entry: 'that might be a really good tell — if it moves quick we work with that coin during the day, if the others are slow then we drop them.' | ⭐⭐⭐ BOTH HALVES CLEAR, AND THE DROP HALF IS THE BIG ONE. Every re-entry averages +1.153%. If the first position NEVER FILLED the later entries average -0.166% at 44% win — and that is 5,983 of 7,873, SEVENTY-SIX PERCENT of all re-entries. Percentile 0.0 against the within-day null. Dropping them removes three quarters of the re-entry population and it is the losing three quarters. The FAST half also clears: a first fill within 2h gives later entries +2.756% at 66% win, +1.603pp above all re-entries, percentile 100.0. ⛔⛔ BUT THE MIDDLE IS NOT MONOTONE AND THAT MATTERS: 2-6h returns +0.400%, 6-12h +6.430%, 12-24h +7.281%. Speed is therefore NOT the mechanism. A first position that took 12-24h to fill means the coin ran LATE, so its re-entries were taken while the move was still building; a 2h fill means the move happened early and the re-entries came after it. The buckets are confounded by WHEN the re-entry sits relative to the fill. ⚠ ⭐⭐ THE HONEST STATEMENT OF WHAT WAS FOUND: it is not 'fast is good, slow is bad'. It is 'DID THE COIN REACH +10% AT ALL TODAY'. Filled at any speed, the later entries average +5.34%; never filled, -0.166%. ⛔ And that split is itself partly the target measuring itself — 'the coin got to +10%' and 'later entries on that coin do well' share a cause. The DROP rule survives that objection better than the KEEP rule, because it acts on an ABSENCE: nothing has happened yet, and we decline to keep paying. ⛔ ALSO STRUCTURAL AND WORTH KEEPING: under the live one-position-per-coin rule with a 24h hold there were only 61 re-entry pairs in 77 days, because a coin can only be re-entered the same day if its first position FILLED. The rule already implements most of the operator's 'drop the slow ones' by accident — a coin whose position is still open cannot be re-entered anyway. What the finding adds is the case where the first position closed on the CLOCK rather than the limit. |
| two_tier_exit_by_rank 2026-09-06 · Operator 2026-09-06: 'if we moved on all that had a clear spike coming, we catch the spike and sell everything that is below the 7 top movers in the first spike at 3% — we should still win those trades and we are still in the top players.' | ⛔⛔ WORSE AT EVERY SETTING, AND THE CEILING IS WORSE THAN THE LIVE VERSION. Everything on +10% returns +0.662% over 27 of 75 green days. Two-tier on the LIVE rank: top 3 +0.209%, top 5 +0.306%, top 7 +0.328%, top 10 +0.420%, top 15 +0.418% — every one below the +10%-everywhere baseline, and none of them buys a green day (26-27 of 75 throughout). ⭐⭐⭐ AND THE HINDSIGHT ROWS ARE WORSE STILL: top 3 -0.033%, top 7 +0.038%, top 15 +0.163%. PERFECT KNOWLEDGE OF WHICH COINS FINISH THE DAY HIGHEST MAKES THE RULE WORSE, which is the strongest possible statement that the ranking is pointed the wrong way. ⚠ ⭐⭐⭐ THE REASON IS THE MOST USEFUL THING HERE AND IT GENERALISES. `ret_so_far` identifies the day's top MOVERS precisely because they have ALREADY MOVED — that is why it reaches 42-58% top-5 precision at the spike anchors. But the exit question is about movement STILL TO COME, and a coin that is already up 15% has less of it left than one that is up 2%. So the reading that best DETECTS the day's winners is close to the worst reading for deciding which position deserves patience: giving the patient exit to the highest-ranked coins caps the ones with room and frees the ones without. ⛔ DETECTION AND FUTURE UPSIDE ARE DIFFERENT QUANTITIES, and this is the first time that distinction has cost measurable money rather than appearing as an autocorrelation caveat in a footnote. |
| trail_on_the_detected_runners 2026-09-06 · Operator 2026-09-06: 'the trail will pay well especially since we are getting into the top runners early.' The trail died on the full population where most moves are small; on the detected runners the moves are +20-30% and a 3% giveback should be a rounding error against them. | ⭐⭐ THE ABSOLUTE NUMBERS ARE MUCH BIGGER AND THE VERDICT IS UNCHANGED. On this population the +3% limit returns +0.253% at 67% win in 9.6h and holding returns +0.238%, while `trail 1% from the high once +5%` returns +1.221% — nearly five times the limit, +0.968pp against it, beating 200 of 200 of its own random-exit controls, with 2%/3%/5% right behind it at +1.044/+0.919/+0.855%. TWENTY-ONE of 126 rules beat both the limit and their own control. ⛔⛔ AND BOTH NULLS STILL REFUSE IT. Family-level, the fair bar for an idea the operator named: best-of-30 in noise has a median of +1.339pp and a 95th of +1.643pp against a real best of +1.272pp — PERCENTILE 36.0, below the noise median for the second time. Search-level: median +1.736pp against +1.694pp, percentile 42.0. ⚠ ⭐⭐⭐ THE REASON IS WORTH MORE THAN THE RESULT AND IT SETTLES THE FAMILY: THE EDGE AND THE NOISE BOTH SCALE WITH THE POPULATION'S VOLATILITY, SO THEY CANCEL. On the full population the trail's best was +0.230pp against a noise median of +0.235; on the runners it is +1.272pp against a noise median of +1.339. Five times the numbers, the same ratio, the same verdict. A random exit taken at the same holding times captures the same fat moves the trail does, because on a population selected for moving there is a lot to capture no matter when you leave. ⭐ The prize is confirmed larger here — selling each trade at its own peak returns +8.148% against +5.285% on the full population — so there IS more on the table; we still cannot reach it. ⛔ POWER: 184 trades over 10 days and 25 coins is thin and the null is correspondingly wide, so this is 'not visible at this sample' rather than proof of absence. But it is the SIXTH exit family or variant to land inside its own null, and the consistency is now itself evidence. |
| the_peak_is_visible 2026-09-06 · Operator 2026-09-06: 'if you didn't detect it then you are not looking at the right indicator... you keep giving up on something that has to be mathematically provable.' | ⭐⭐⭐ YES, AT ABOUT TEN STANDARD ERRORS. In the five minutes BEFORE the peak: thermometer score z +0.483, volatility +0.481, activity +0.434, large_trade_share +0.244, flow_imbalance +0.205 — fourteen readings at percentile 100 against a null where a random minute of the same trade pretends to be the peak. At n=482 the standard error on a mean z is ~0.046. THE ANATOMY: the surge builds from -30m, spikes at -5m, and collapses THROUGH the top (score -0.098 at +5m, -0.268 at +30m). ⛔ THE HARD BAR IS MUCH HARDER: picking the true peak out of the ~25.9 local highs of the same trade, volatility gets 9.6% against a 3.9% random baseline — real (percentile 100) and weak. A shape-based version (surge/collapse) did WORSE at 6.8%. ⚠ ⛔⛔ THE METHODOLOGICAL FINDING IS BIGGER THAN THE RESULT AND IS RECORDED AS A LAW (LAW-discover-not-disprove.md). Six exit studies concluded the exit was unreadable; all six asked 'does this RULE beat a random exit at the same holding times', which bundles signal with threshold, window, direction and population. I reported the consistency of six failures as evidence and it was one design flaw counted six times. |
| buyers_exhausted_exit 2026-09-06 · Operator 2026-09-06: 'the last person is willing to buy. no more buyers, only sellers... we can even have a certain percentage of buyer to seller we are looking for to choose as the peak. if we get out too early, we still made money.' And: 'the target is make money. that is all that matters.' | ⭐⭐⭐ PER TRADE IT LOSES AND IN THE ACCOUNT IT WINS. Per trade: +1.943% against the +10% limit's +3.614%, but 38.3% capture of the peak against 34.6%, a 96% WIN RATE and a 5.1 HOUR hold against 24. IN THE ACCOUNT AT THREE SLOTS: 73 legs taken against 34, +26.0% RETURN against the registered rule's +18.1%, and a 0.0% MAX DRAWDOWN against 2.6%. At 8 slots +23.9% against +21.9%. At 20 slots the registered rule wins (+22.5% against +17.9%) because there is no scarcity left to exploit. ⭐ THE FASTER EXIT IS WORTH EXACTLY AS MUCH AS THE SLOT IS SCARCE, and no per-trade average can express that — at three slots the book refused 448 of 482 signals. ⚠ ⛔ TWO EXTENSIONS TESTED AND BOTH FAIL. MULTI-LEG (flip out, re-enter, repeat) generates 1,701 legs from 482 runs and returns +2.7% at three slots against one leg's +26.0%: every leg pays its own 0.681% round trip, which costs 2.40% per run in friction, and the re-entries crowd out fresh coins. The buyers DO come back (465 of 469 cases, 99%) and there IS a pullback (median trough -1.16% within 2h, 57% dip more than 1%) — the second leg simply does not pay for itself. SKIPPING THE FIRST SPIKE also fails: every spike +26.0%, skip the first +24.2%, third-on +20.8%. ⭐ The reason is the best part: skipping was meant to avoid the losers, and under the flip exit THERE ARE NO LOSERS TO AVOID — the win rate is 96-100% everywhere. ⛔ ALL IN SAMPLE on 12 days the arms were fitted to, and the parameters came from a 15-cell grid. |