Built 2026-09-16 17:00 UTC data 20s ago · nothing placed, nothing traded
8 strategies scored on 7,525 recorded signals, across ~94 independent 4h blocks (wa…
8 strategies scored on 7,525 recorded signals, across ~94 independent 4h blocks (want 40+). Scoring many rule sets against one small sample guarantees one looks excellent by chance. Ranking below is for watching, not for acting: until the block count is high enough, the leader is whichever strategy was luckiest. Adding strategies makes this worse, never better.
The quiet tape — the thermometers, read at the entry instant
Registered 2026-09-05. Clock started 2026-09-08T15:00Z.
Registered 2026-09-05. Clock started 2026-09-08T15:00Z. The first thing on this board built from our own minute tape rather than from candles — and the first use ever made of the thermometers, which had been built, tested and wired into nothing.
What it says: in the last five minutes before a second buying spike the tape goes quiet — the large trades withdraw, activity falls, volatility falls. Buy/sell pressure on its own says nothing at that moment (percentile 0.2). It is the conjunction that carries it: buying pressure matters only inside a quiet tape.
⛔ The forward column is the only evidence here. The history column chose these arms and can only agree with itself. QUIET-3 sat at percentile 94.7 in sample against a null whose 95th percentile was +11.1pp — on the line, not over it. A pair search over 946 combinations on the same data returned the null’s own median: at 66 spikes a search finds exactly what it manufactures, which is why these two arms were named from the mechanism instead.
The recorder’s tape starts 2026-08-25, so 819 of the 900 symbol-days this study wanted do not exist. It grows about 12 spikes a day.
The exit, decided 2026-09-05:
The exit, decided 2026-09-05: the 24h clock is the primary exit and a +10% net limit is attached at entry (a Kraken conditional close), so it lives on the exchange from the moment we are filled. ⭐ At +3% the limit is a cap and costs half the edge — swept over 8,662 entries and 77 days, net rises monotonically with the level (+0.224% at 3%, +0.997% at 10%, +0.972% with no limit), and a fixed-pot run agrees (+0.9% vs +7.0%). At 10% it is free insurance on the spike that really runs. All three exits are settled forward on every firing so the live record can overturn the choice.
Arm
Signals
Settled
Open
Bars
Days
Mean net
Median
Win%
Exit
THE-FLIP ⭐ FLAGSHIP 10/250 settled instants · 2/40 days · forward 10.0% · +0.079pp passes at: beats QUIET-3 IN THE BOOK at 3 and 8 slots, not per trade · in sample 117 on 15d · 14.5% · -0.010pp
29 +57 sup
24
5
64
5
-1.33%
+1.24%
54%
the buy-share flip
TOP-DECILE 2/150 settled instants · 2/40 days · forward 50.0% · +1.585pp passes at: +1.0pp day-matched money, above QUIET-3 · in sample 42 on 11d · 21.4% · +0.093pp
8 +13 sup
6
2
19
4
-2.49%
-0.60%
33%
+10% limit, else the 24h clock
QUIET-3-AT-15 10/250 settled instants · 2/40 days · forward 10.0% · +0.079pp passes at: beats QUIET-3 on the SAME entries · in sample 117 on 15d · 14.5% · -0.010pp
29 +57 sup
23
6
64
5
-1.67%
-2.46%
30%
+15% limit, else the 24h clock
QUIET-3 10/250 settled instants · 2/40 days · forward 10.0% · +0.079pp passes at: +7pp spike lift · in sample 117 on 15d · 14.5% · -0.010pp
31 +57 sup
25
6
65
6
-0.01%
-0.54%
40%
+10% limit, else the 24h clock
QUIET+BUYERS 7/150 settled instants · 2/40 days · forward 28.6% · +0.128pp passes at: +0.5pp day-matched money · in sample 68 on 14d · 22.1% · -0.054pp
16 +22 sup
14
2
36
6
-2.98%
-3.35%
14%
+10% limit, else the 24h clock
LATER-WAVE 6/120 settled instants · 2/40 days · forward 0.0% · +1.194pp passes at: +1.0pp day-matched money, above QUIET-3 · in sample 74 on 15d · 16.2% · +0.259pp
17 +34 sup
13
4
43
5
+0.19%
-0.46%
46%
+10% limit, else the 24h clock
THE-FLIP-SELECTED 4/150 settled instants · 2/40 days · forward 0.0% · +1.745pp passes at: beats THE-FLIP in the book at 8 and 20 slots · in sample 31 on 11d · 9.7% · +1.121pp
2 +7 sup
2
0
8
2
-4.02%
-4.02%
50%
the buy-share flip
THE-SECOND-ENTRY 0/60 settled instants · 0/40 days · no settled instant yet passes at: positive net, above a random re-entry minute · nothing claimed in sample
5
5
0
16
4
+5.79%
+5.60%
80%
+20% limit, else the first position's remaining clock
THE-FLIP-STRONG 4/100 settled instants · 2/40 days · forward 0.0% · +1.745pp passes at: beats THE-FLIP-SELECTED, and a lower never-arm rate · in sample 23 on 9d · 8.7% · +1.663pp
2 +7 sup
2
0
8
2
-4.02%
-4.02%
50%
the buy-share flip
THE-BOUNCE 0/60 settled instants · 0/40 days · no settled instant yet passes at: positive net, above holding the first position · nothing claimed in sample
0
0
0
0
0
—
—
—
+10% limit, else the first position's remaining clock
THE-PACE 4/150 settled instants · 2/40 days · forward 0.0% · +1.745pp passes at: beats THE-FLIP-SELECTED on the same entries · in sample 31 on 11d · 9.7% · +1.121pp
2 +7 sup
2
0
8
2
-4.05%
-4.05%
50%
+3% then a 1% trail ยท slow ones cut and re-bought
THE-SECOND-ENTRY registered 2026-09-07 · fires on a coin THE-FLIP has already exited, when the price is 8.0% BELOW that exit — a median ?h later, not straight away in sample: ? fires on ? exits · ?% · ? · null percentile ? ⛔ its window is the REMAINDER of the first position's 24h, not a fresh one — the figures are a floor
5 +11 watching
5
0
16
4
+5.79%
+5.60%
80%
+20% limit, else the first position's remaining clock
These are the lane's own firings, and every figure is settled on THAT ARM'S OWN EXIT
⭐ These are the lane's own firings, and every figure is settled on THAT ARM'S OWN EXIT — the flagship on the buy-share flip, the others on their own limit or the 24h clock. Net of the 0.681% round trip. Signals counts what could be HELD; a firing the one-position-per-coin rule suppressed is shown beside it and is deliberately not in the money, because it is evidence about the arm and not money we could have made.
⛔ Nothing here is a result. The registered bar is under each arm's name and it is counted in SETTLED INSTANTS against 250 across 40 days, which is a different and much slower count than these trades.
⛔ The counts above are 26 hours behind, and that is by design. An instant can only be scored once 24 bars have closed after it, so the newest 26h of trading is never in the pass-line total. 7 firings are already on the lane waiting to become countable at roughly 18 firings a day. A stale-looking count here is a settlement delay, not an absence of trades — the per-arm table above is current to the last closed bar.
What the lane has actually done
104 bar(s) judged · 41 taken, 64 suppressed as double buys · 9 filled at +10%, 7 still open — every decision taken at the close, median latency 62s against a 120s gate.
⭐ ONE POSITION PER COIN (operator, 2026-09-06). A coin that has already been bought and SOLD may be bought again — that is a second leg of the same run and is the thing we want. A coin whose position is still open may not. The arm still fires either way and the suppression is written down, so the arm's hit rate stays measurable while the book stays honest about what it could have held.
⭐ THE-FLIP is the flagship arm and the only one whose EXIT is derived from the event rather than from a price level chosen in advance — it sells when the buy share collapses, not at +10%. It holds about 5 hours instead of 24 and hands the slot back. ⛔ Its exit is a MONITORED CONDITION, not a resting order — it needs this machine alive every minute of the hold, which no other arm does.
⛔⛔ CORRECTION, 2026-09-07 — THE IN-SAMPLE CASE THAT MADE THIS THE FLAGSHIP DOES NOT HOLD, AND IS WITHDRAWN. It was registered on “+26.0% against the +10% limit’s +18.1% in the book at three slots”. That comparison was computed on positions filtered to those that later rose at least 3% above entry — a condition nobody can know at the moment of entry. Same days, same simulator, same code, the filter removed:
positions kept
THE-FLIP
+10% limit
flip win%
peaked ≥3% (as registered)
+30.8%
+23.5%
99%
peaked ≥2%
+24.3%
+12.1%
91%
peaked ≥1%
+8.2%
+10.1%
79%
no filter — every position
-4.2%
+2.2%
68%
3 slots, 880 positions unfiltered against 501 filtered, 13 days. At 8 slots the limit takes +6.0% against the flip’s -3.1%; at 20 slots +3.4% against -3.4%.
The advantage tracks the strength of the filter and reverses without it. The 96–99% win rate and the 0.0% maximum drawdown were the filter describing itself: every position in that sample had, by construction, been up 3% at some point, and the flip arms at +2%. On the honest population the registered +10% limit is the only rule that stays positive, at every slot count — though at +2.2% over 13 in-sample days it is thin, and what died here is the COMPARISON, not just the flip.
⭐ WHAT THIS DOES NOT SAY: that the flip is a bad exit. On the unfiltered population the flip still wins 68% of its trades against the +10% limit’s 36%, and takes 65 trades where the limit takes 36. What it cannot do is protect a trade that never reaches +2% and so never arms — those ride to the 24h clock with no exit at all, and they are exactly the positions the ≥3% filter had removed. The failure is in the population, not obviously in the exit, and “is the flip a good exit CONDITIONAL on firing” is a different question that has not been asked yet.
⭐ Nothing has been changed in the arm. THE-FLIP is registered on a forward clock to 2026-10-15 and that clock is what judges it; an in-sample defect in how it was CHOSEN is not grounds to edit a live arm mid-clock. What is withdrawn is the in-sample case, which is what this board was asserting.
⛔ These are DECISIONS, not the pass line. The table above counts instants whose 24-hour outcome has settled, which is why it can read zero while this table lists trades — a firing from an hour ago has no outcome yet. Both numbers are real and they are not the same number.
Coin
Bar
Arm
State
How it was taken
SUPER/USD
09-10T07:00Z
QUIET-3 + LATER-WAVE + QUIET-3-AT-15 + THE-FLIP
open · -0.18% · the flip has not fired falls back to the 24h clock
tape-at-close · 63s
ATOM/USD
09-10T04:00Z
QUIET-3 + QUIET+BUYERS + QUIET-3-AT-15 + THE-FLIP
open · -1.76% · the flip has not fired falls back to the 24h clock
tape-at-close · 62s
ESPORTS/USD
09-10T01:00Z
QUIET-3 + QUIET-3-AT-15 + THE-FLIP
open · -1.12% · the flip has not fired falls back to the 24h clock
open · -6.03% · the flip has not fired falls back to the 24h clock
tape-at-close · 63s
Net of 0.681% round-trip fees. Money is a day-matched excess, never a raw mean. Nothing here places an order, and no order path is opened by it.
Waiting for the second flip — the operator's question, measured
The question, 2026-09-07:
The question, 2026-09-07:“instead of it exiting after the first signal to exit it was told now it is alright to exit after the next signal like this — what would happen?”
It aims at the exact mechanism that breaks the flagship on big runs: the rule cannot tell a PAUSE from an ENDING. Inside a long run the 10-minute buy share dips under the exit line at every consolidation, so the biggest runs are the ones it quits soonest and captures least of. A pause makes one dip and the buyers return; an ending makes a dip that does not. Counting the dips is the cheapest way to tell them apart and needs no new reading.
How a signal is counted: arm when the 10-minute buy share is ≥ 0.60 and the trade is at least +2% in profit; flip when it falls below 0.50; then it must climb back above 0.60, still in profit, before the next dip counts as a new signal. ⛔ Without that re-arm condition one long decline would tick off dozens of “signals” in consecutive minutes and N would mean nothing. A trade that never reaches its Nth flip rides the 24h clock — that is most of what changes.
The whole book — 882 positions, 13 days, no hindsight filter
exit rule
mean
median
win
hold
how it ended
exit on flip #1 (REGISTERED — the live arm)
-0.092%
+0.984%
63%
12.8h
572 fired · 310 to the clock
exit on flip #2
+0.053%
+1.010%
58%
15.3h
484 fired · 398 to the clock
exit on flip #3
+0.211%
+1.020%
57%
16.2h
447 fired · 435 to the clock
exit on flip #4
+0.205%
+0.631%
55%
17.2h
412 fired · 470 to the clock
+10% limit
+0.686%
-0.429%
45%
—
—
hold to the clock
+0.619%
-0.551%
44%
—
—
And by the run that was actually available
available peak
n
flip #1
flip #2
flip #3
flip #4
the clock
2-5%
274
+0.727%
-0.146%
-0.296%
-0.665%
-1.200%
5-10%
168
+2.026%
+2.544%
+3.131%
+3.367%
+1.601%
10-20%
131
+2.233%
+3.315%
+3.711%
+3.811%
+6.401%
20%+
47
+2.382%
+5.329%
+5.969%
+6.880%
+16.866%
⛔ This split is on the peak available over the whole hold, which is HINDSIGHT. It shows what the rules did; it is not a rule, because nobody knows at the exit which bucket a trade is in.
⛔⛔ THE VERDICT: IT RAISES THE NUMBER AND LOSES THE INFORMATION.
⛔⛔ THE VERDICT: IT RAISES THE NUMBER AND LOSES THE INFORMATION. Against 400 draws of a random exit taken at the same holding times, the best N (flip #3) scores +0.211% against a null median of +0.116% — percentile 83.5. Flip #1 sits at percentile 100 on the same test. Once you wait, what you are really doing is holding longer, and a random exit at those same longer holding times does just as well. The skill was in the FIRST flip.
⭐ And it is a MIX problem, which is what the operator said it might be. On the biggest runs waiting nearly triples the result, and on the smallest it destroys the only bucket where the flip beats everything else. The two cancel, which is why the whole-book number barely moves. The right behaviour genuinely differs by bucket — what is missing is a live-knowable tell, at the moment the first flip fires, that says this one is going to run. ⛔ It is probably not “how far up is it on the day”: the two-tier exit by rank was tested on 2026-09-06 and made things WORSE at every setting, because a coin already up 15% has less room left — the reading that best DETECTS the day's winners is close to the worst one for deciding which position deserves patience.
⛔ In sample. N was chosen by reading this table, so the winning N is a SEARCH RESULT and its margin is not a forward claim. Nothing here is registered and the live arm is unchanged.
The forward test — a separate experiment
Not ranked in the table above, on purpose.
⛔ Not ranked in the table above, on purpose. Different signal population (a 4h capitulation spike, not a minute-level rank climb) and a different cost basis (0.681% measured, not 0.40%). Ranking them together would reward the cheaper fee. Clock started 2026-09-04T14:00Z · judged at ≥250 signals across ≥40 days.
LIVE: 538 signal(s) fired across 13 day(s) · 12.1 of the ≥40 days elapsed…
LIVE: 538 signal(s) fired across 13 day(s) · 12.1 of the ≥40 days elapsed · 197 forward AI decision(s), 163 of them taken AT the bar close ⛔ None of that is a result — it is the count of a test that has not finished. The pass line is ≥250 signals across ≥40 days, and the day-blocked interval does not clear zero until roughly 180 days.
Arm
Signals
Settled
Bars
Days
Mean net
Median
Win%
AI decisions
K (registered, gated)
87
70
33
11
-0.17%
+3.00%
63%
64
K-U (ungated)
177
157
44
12
-0.46%
+3.00%
59%
132
K + |ΔOI| ≥15% (positioning)
18
15
13
7
-1.79%
-0.96%
40%
—
K + n_x ≥3.14 (the crowd)
34
25
22
10
+0.89%
+3.00%
76%
—
K + n_x + close ≤0.25 (both)
21
16
14
8
+0.47%
+3.00%
75%
—
⛔ Nothing here is a result. The registered bar is 250 signals across 40 days, and the day-blocked interval does not clear zero until ~180 days.
The positioning arm — the first reading not made of price
Every reading this project has ever tested was a transformation of PRICE and VOLUME…
Every reading this project has ever tested was a transformation of PRICE and VOLUME on one venue's spot market — all 110 of them. This one is not. It is the magnitude of the change in a coin's perpetual-futures OPEN INTEREST over the 24 hours ending at the spike, from Kraken's futures API. ⛔ NOT the venue we trade — we trade SPOT. Open interest is read as INFORMATION about positioning; no price, entry or exit in this system comes from the futures series any more (2026-09-08). Open interest is how much leveraged money is actually in a position: a published measurement of what people are doing, not an inference from what price did.
Why it is worth more than another price reading: price can trend forever, leverage cannot. When positioning is extreme somebody is paying a toll to hold, and a toll has a forced end. That is a mechanism rather than a pattern — the first one here.
In sample (282 signals, 97 coins, 88 days) the quartiles run +0.37 / +0.06 / +0.71 / +1.43% with win rates 59 / 56 / 63 / 80%, against Candidate K's +0.61% and 68%.
⛔ IT PASSED FOUR CHECKS AND FAILED THE FIFTH.
⛔ IT PASSED FOUR CHECKS AND FAILED THE FIFTH. Passed: monotone in the money; a day-blocked interval on the gap that clears zero (+0.34…+1.91); the hold-out in both directions — the exact check Candidate K's own parameter choice failed; and near-zero correlation with what we already read (0.13 against volume). Failed: it does not clear its rank null (percentile 85.6 of a luckiest-of-six shuffle). Four of five is not five of five, which is why it is an arm on a clock and not a change to the rule. ⛔ And an earlier story is retracted: on 45 coins this looked like “falling open interest means a liquidation flush”. On the full 97-coin set the signed reading flips sign on the hold-out. Only the magnitude survived — which is the third time in one day that magnitude has beaten direction here. Coverage: only 97 of our 159 coins have a Kraken perpetual, so this arm can judge about three quarters of the signals. The rest are not judgeable — that is not the same as rejected, and the two are never pooled.
Why the two new arms — the trade count
Every OHLC bar carries n, the number of trades — and for 120 days nothing rea…
Every OHLC bar carries n, the number of trades — and for 120 days nothing read it. We computed vol × vwap for dollars and never once asked how many people produced them. Dollar volume decomposes: vol_x = n_x × size_x. The rule already gates on vol_x ≥ 2; it just could not tell which half of the spike it was.
Measured on 374 settled signals over 99 days, the answer is unambiguous: it is the CROWD, not the SIZE.
the volume spike was…
mean net
win
many people, big tickets
+1.17%
77%
many people, small tickets — a crowd
+0.87%
74%
few people, BIG tickets
+0.17%
56%
few people, small tickets
+0.23%
65%
A 4–5× split on who acted and essentially nothing on how big each actor was. Average trade size is dead on its own (AUC 0.469, and its hold-out flips sign), so there is no size arm.
It survived the three checks that killed the last three candidates. It is not a thinness tilt but the reverse — stronger in the liquid half (AUC 0.664 vs 0.557), where an order can actually be filled, only the second reading this project has ever produced that does that. It passes its hold-out in both directions (pick the cut on the bear half, score the bull: +0.30pp; and the reverse: +0.67pp) — the exact check Candidate K’s own parameter choice failed at percentile 20. And within a single day it picks the right coin among those firing together (+1.38pp, day-bootstrap 95% CI +0.41…+2.44), so no market regime can be doing the work.
The two readings are independent — spearman(n_x, close-in-bar) = −0.015 — which is the whole reason a “both” arm exists instead of being a near-duplicate. In sample, each alone has an interval that touches zero; together it clears by a full point: K +0.61% (CI +0.04…+1.12) → K+n_x +1.02% → both +1.43% (CI +1.00…+1.83).
⛔ READ THE CAVEATS WITH THE NUMBERS.
⛔ READ THE CAVEATS WITH THE NUMBERS. 1. In sample, and fitting by construction. Every figure above was found by looking at outcomes on the same 121 days Candidate K itself was chosen from. That is exactly why these are arms on a forward clock and not a change to K. Candidate K remains the registered rule; nothing on this page may nominate a replacement. 2. n_x is 0.853 correlated with the volume the rule already reads. It is largely the same reading under a different name. What saves it is that it still separates inside a fixed volume band — but it inverts in the lowest band, and on its rank statistic alone it does not clear a luckiest-of-4 null. 3. The threshold 3.14 is the in-sample median, frozen as a constant — the centre of a plateau, not its peak. It is frozen because a threshold recomputed from the live population each run is a definition that moves, and an arm whose definition moves cannot be scored against a clock. 4. A sharper read is a slower one. The “both” arm takes only 32% of K’s signals, so its own 250-signal bar sits roughly 3× further out. 5. The board-wide flush idea was refused. How many coins capitulate in the same bar reads AUC 0.499 — nothing. Its one spectacular cell (10+ coins together, +2.91%, 97% win) is 36 signals across exactly two days. That is one event wearing a factor’s clothes, and it would have been the headline if the day-count check had not been built into the tool.
Every experiment running
Strategy
Trades
Mean net
Median
Win%
narrow_board
2,808
+0.05%
-0.69%
38%
combined
1,429
-0.04%
-0.48%
43%
baseline
6,908
-0.18%
-0.71%
37%
lower_entry
3,864
-0.25%
-0.61%
41%
not_already_run
6,334
-0.27%
-0.69%
37%
operator_btc_rule
3,667
-0.49%
-0.92%
37%
baseline_4h
7,416
-0.70%
-0.75%
26%
operator_btc_strict
174
-1.50%
-2.43%
22%
What each one is
Strategy
Rule
baseline
What the system does today: every money-floor signal, bought at signal, held 24h. The control every other strategy must beat.
baseline_4h
Same, exited at 4h. Isolates holding period alone.
operator_btc_rule
Operator: do not trade while BTC is losing. Trailing 2h measure, held 24h.
operator_btc_strict
The operator's fuller rule: below +1% on the day is 'tough'. Stricter gate.
lower_entry
Rest a limit at the deeper of the pre-jump low or -2%. Tests entry price alone.
not_already_run
Skip coins already up >2% on the day. The pct_24h factor, which held on both the runner and the money test.
narrow_board
Only trade when under 40% of the board is green -- a mover in a thin tape is doing something of its own.
combined
Lower entry + not-already-run + narrow board. Every factor that survived its own test, together.
⛔ THE FEE WAS WRONG UNTIL 2026-09-05 AND IT CHANGED EVERY SIGN HERE.
⛔ THE FEE WAS WRONG UNTIL 2026-09-05 AND IT CHANGED EVERY SIGN HERE. This page was net of 0.40%, labelled “conservative”. The round trip measured on this venue is 0.681% taker, and these are modelled as taker entries — so every figure was overstated by 0.281pp. The leader read +0.24% and is really −0.04%. Now corrected. The whole edge of the best rule this project has is 0.537pp, so a 0.281pp error in the cost basis was over half of it. On a thin edge the fee is not an accounting detail, it is the result.
Net of 0.681% round-trip fees. Every figure is paper: no order has ever been placed on any venue.
Public Kraken market data only. No account balance, no position, no key material on this page. Published every 5 minutes; this page reloads itself every 150s. The build stamp above is the truth — if it stops moving, the publish has stopped.