prop-challenge-lab research console

axis

conditioning

search space

4

alpha

0.02

outcome

range/ATR r = -0.131, partly algebraic, 1.7% of variance

effect

—

interval

—

Statement

Adverse excursion is FORESEEABLE before entry: pre-trade volatility measures predict MAE/stop-distance, so days on which the stop is likely to be run can be identified in advance.

Mechanism written before the number existed

H-STOPWIDTH established MAE is LARGE -- 77% of trades exceed the stop, median 3x -- but not whether it is foreseeable. The mechanism here is the strongest available to us and is not a market-edge claim at all: MAE is essentially a realised-volatility quantity, and volatility CLUSTERS. That is one of the most robust facts in financial time series. If it holds intraday on this setup, a day's likely excursion is partly knowable from the volatility that preceded it.
Note what this would and would not buy. It is NOT a directional edge and cannot become one. At most it is a SIZING or STAND-ASIDE filter: skip, or size down, on days where the stop sits well inside the expected noise. That is worth having only if a strategy with a real entry ever exists -- so this is groundwork, honestly labelled.

Power plan

{
 "cluster_size": 2,
 "effect": 0.15,
 "intra_cluster_r": 0.9,
 "test": "correlation"
}

{
 "estimator": "occams.stats.spearman",
 "holds_constant": "data + setup logic (frozen script imported)",
 "tool": "scripts/backfill_m5.py"
}

M5 REPRODUCTION — supersedes the `mins to breakout` values in 2026-08-01-four-predictors. NO ALPHA SPENT: same estimand, same data, same setup; only the rank function changed.

The original used argsort(argsort(x)), which gives tied values arbitrary DISTINCT ranks in row order. `mins to breakout` is the only integer-valued predictor in the set -- 94 distinct values across 3,260 observations, 97.1% tied -- and it is the only one that moved. MNQ -0.0304 -> -0.0388, a 28% change.

THE CONCLUSION IS UNCHANGED. range/ATR, the one predictor C2 reported as surviving, reproduces EXACTLY at -0.1312. Minutes to breakout was noise before and is noise now. But a recorded number was wrong, and this appends rather than edits it.

Detail: docs/M5-BACKFILL.md

{
 "instruments": [
  "MES",
  "MNQ"
 ],
 "outcome": "log(MAE / stop_distance)",
 "predictors": [
  "range/ATR",
  "overnight gap/ATR",
  "5d mean range/ATR",
  "mins to breakout"
 ],
 "search_size": 4,
 "span": "2019-2026",
 "test": "Spearman"
}

OUTCOME: ONE of four predictors survives, and it is weaker than it looks.
range/ATR: r = -0.131 pooled, -0.129 MES, -0.120 MNQ. Consistent across instruments and above the detectable floor (r=0.067 at effective n=1,736). The other three are noise: overnight gap -0.039, 5-day mean range +0.008, minutes-to-breakout -0.046.
THE CAVEAT, and it matters more than the coefficient. The stop is defined as extreme +/- 0.2 x range HEIGHT, so stop distance scales with height by construction, while MAE scales roughly with daily volatility. MAE/stop is therefore ALGEBRAICALLY close to ATR/height -- the inverse of the predictor. Some unknown share of r=-0.131 is definitional rather than informative.
What argues against it being PURELY definitional is the magnitude: a clean inverse relationship would show rank correlation near -1, not -0.13. So MAE is dominated by day-specific noise rather than by the scaling. But that cuts both ways -- it also means the predictable component is small.
PRACTICAL VERDICT: r=-0.131 explains ~1.7% of variance. As a stand-aside filter that is negligible, and it would be filtering a strategy that has no obtainable entry anyway. NOT worth acting on.
Disentangling the definitional component would be a NEW hypothesis with its own alpha -- deliberately NOT done here, because the registration said one look and no fifth predictor, and the practical verdict does not change either way.

SCRIPT: experiments/C2-MAE-PREDICTABLE/2026-08-01-four-predictors/exp_mae_predictable.py
OUTPUT: experiments/C2-MAE-PREDICTABLE/2026-08-01-four-predictors/output.txt

axis

objective

search space

5

alpha

0.01

outcome

P(pass) 0.042 against a 0.55 gate

effect

—

interval

—

Statement

Scored against the REAL objective -- P(pass) on the verified Zero tier -- the obtainable fade clears the G1 gate of 0.55, and P(pass) is non-monotonic in risk size, peaking BELOW the expectancy-maximising size.

Mechanism written before the number existed

The challenge is a path-dependent survival problem: trailing drawdown floor, profit target, minimum days. Expectancy ignores path. A lower-variance configuration should therefore reach the target more often before touching the floor, which would put the P(pass) maximum at a smaller size than the one maximising expectancy.

{
 "supersedes_note_of": "2026-08-01-zero-tier"
}

OUTCOME: hypothesis REJECTED on BOTH claims, and the second rejection is the useful one.
(1) The obtainable fade does not clear G1. P(pass) 0.042 at 90 days against a 0.55 gate, with P(breach) 0.440. The SEALED variant -- the one whose entry cannot be obtained -- reaches only 0.258, so even the artifact would have failed. Verdict v3's NO-GO stands twice over.
(2) MY REGISTERED PREDICTION WAS WRONG. P(pass) is MONOTONICALLY INCREASING in risk size, not non-monotonic: 0.000 at $75, 0.009 at $125, 0.042 at $175, 0.088 at $250, 0.174 at $350. There is no interior peak.
WHY, and this is worth keeping: with NEGATIVE expectancy the only route to a +$3,000 target is variance, so P(pass) rises with size -- and P(breach) rises faster (0.811 at $350). At $75 the strategy neither passes nor breaches; it simply drifts. The interior peak I predicted REQUIRES a positive edge, because only then does reducing variance preserve the target while moving the floor further away.
NEW LAB CAPABILITY implied: the SHAPE of P(pass) versus size is a diagnostic for whether an edge exists at all, independent of any expectancy estimate. Monotonic-increasing means variance harvesting and no edge; an interior peak means a real edge. Worth adding to the harness as a standard readout.

{
 "account": 50000,
 "calendar_filter": true,
 "horizons": [
  60,
  90
 ],
 "instrument": "MES",
 "min_days": 1,
 "risk_ladder": [
  75,
  125,
  175,
  250,
  350
 ],
 "target": 3000,
 "tier": "verified Zero 50k",
 "trailing_dd": 2000
}

PLACEHOLDER - replaced below once the outcome is written up.

SCRIPT: experiments/E2-OBJECTIVE/2026-08-01-zero-tier/exp_objective.py
OUTPUT: experiments/E2-OBJECTIVE/2026-08-01-zero-tier/output.txt

axis

order flow

search space

1

alpha

0.0

outcome

MES spread 1.0 tick, MNQ 2.0 -- assumption optimistic

effect

—

interval

—

Statement

The sealed cost model -- $1.25 per side commission and 1 tick of slippage -- is accurate for our order sizes on MES and MNQ. Specifically: the effective spread is 1 tick, and a typical order of 1-6 contracts does not move price.

Mechanism written before the number existed

Commission was measured at 0.05-0.09R per trade, which is decisive at an edge of ~0.03R gross. Slippage enters the same arithmetic and is currently ASSUMED. If real slippage is 2 ticks rather than 1, every net figure in the register shifts by roughly the same magnitude as the edge itself. The trades tape carries aggressor side and size, so both the effective spread and our size relative to typical prints can be measured rather than assumed.

{
 "assumed_model": "1 tick slippage, $1.25/side",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "sessions_per_inst": 60,
 "source": "R2.1 trades tape, 09:45-10:15 ET"
}

OUTCOME: SPLIT. The assumption holds for MES and is OPTIMISTIC for MNQ.
MES: effective spread 1.00 ticks median, and even a 6-lot fills with 0 ticks of travel (p90 1). 6 contracts sits at the p90 of print size. The sealed 1-tick assumption is CORRECT for MES.
MNQ: effective spread 2.00 ticks median (mean 1.60) -- DOUBLE the assumption. A 3-lot travels 1 tick median / 3 at p90; a 6-lot 1 / 4. Only 3.6% of MNQ prints are 6 contracts or larger, so a 6-lot is a large order there, unlike MES where it is ordinary.
MAGNITUDE, stated so it is not over-read: one extra tick per side on MNQ is $1.00 per contract, roughly 0.006R at 1 contract and 0.017R at 3. That is real and it makes every MNQ figure in the register slightly optimistic, but it is an order of magnitude below the gap that matters (0.03R gross against 0.25R required). It changes no conclusion; it tightens the error bar.
ACTION: raise slippage_ticks to 2 for MNQ in the instrument costs, leave MES at 1, and re-state that MNQ results predating this were measured on an optimistic model.

SCRIPT: experiments/E3-COSTS/2026-08-01-tape-measured/exp_costs.py
OUTPUT: experiments/E3-COSTS/2026-08-01-tape-measured/output.txt

axis

price

search space

1

alpha

0.01

outcome

detectable

effect

-0.2393

interval

[-0.2910, -0.1843]

Statement

Requiring price to trade through the stop, back through it, and then reclaim the original entry before entering will improve the fade's per-trade expectancy.

Mechanism written before the number existed

User observation from chart inspection: on several 2026-07 setups the eventual direction was correct but the stop was hit first. Stated BEFORE the test was run. The proposed mechanism is that a reclaim confirms the fade thesis and filters days that simply keep going.

{
 "cluster_size": 2,
 "engine": "M1-M7 audited",
 "intra_cluster_r": 0.9,
 "reproduces": "2026-08-01-full-history (10/10)",
 "seed": 20260830
}

F5 recompute. The archived run recorded bare means with no interval, no floor and no effective n, so it could not be resolved as a measured result. This reproduces it 10/10 and then states the estimand: the FILTERED strategy expectancy against the sealed arm. Strategy level, not trade level -- the filter changes which days are traded.

SCRIPT: experiments/H-RECLAIM/2026-08-30-audited/f5_reclaim.py
OUTPUT: experiments/H-RECLAIM/2026-08-30-audited/output.txt
CALCS: experiments/H-RECLAIM/2026-08-30-audited/calcs.json

{
 "calendar_filter": false,
 "folds": "pooled (NOT the sealed OOS/lockbox split)",
 "grid_size": 1,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "k_stop": 0.2,
 "risk_usd": 175.0,
 "rule": "reclaim-before-entry",
 "years": "2019-2026"
}

OUTCOME: hypothesis REJECTED. The reclaim is anti-confirmation, not confirmation: you enter at the same price with the same stop but only after the market has demonstrated it can punch through that stop, so the filter selects for days where the stop is weakest. It removes ~40% of trades and removes the wrong 40%. CAVEAT, recorded rather than buried: controls_passed=False because this run used a re-implementation, pooled years and no calendar filter -- it is NOT the sealed engine. The V0-vs-V5 COMPARISON is valid because both share those simplifications; the ABSOLUTE levels are not, and reconciling them is task X1.

axis

price

search space

1

alpha

0.02

outcome

gross +0.011 -> net -0.082 R; costs ate the signal

effect

—

interval

—

Statement

Use the failed-breakout as a SIGNAL only, take no fade, and instead enter WITH the original breakout direction when price returns through extreme +/- 0.2h (the level where the fade would have been stopped). Stop at the range boundary; target one range height beyond entry.

Mechanism written before the number existed

H-STOPWIDTH established that 77% of fade setups see price travel through the stop, median 3x the stop distance. If the fade is systematically on the wrong side, the second push through the extreme is the tradeable event: the pullback has shaken out the faders and the original breakout resumes. Distinct from the engine's 'follow' null, which enters at the BOUNDARY rather than at the stop level.

{
 "calendar_filter": false,
 "entry": "stop order at extreme +/- 0.2h, WITH the breakout",
 "folds": "pooled",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "risk_usd": 175.0,
 "span": "2019-05-06..2026-07-03",
 "stop": "range boundary",
 "target": "one range height beyond entry"
}

OUTCOME: REJECTED as a standalone strategy -- net -0.082R (MES) / -0.015R (MNQ), negative in 6 of 8 years on MES. BUT the decomposition is the real finding, and it is bigger than this hypothesis. BOTH directions are GROSS POSITIVE and both are made negative by COMMISSION: fade +0.042/-0.088, second push +0.011/-0.093 on MES. Cost drag is 0.05-0.09R PER TRADE because true-risk sizing buys many contracts when the stop is tight -- MES median 6 contracts x $2.50 round turn = $15 against a $175 risk budget = 0.086R of pure commission. This reconciles H-STOPWIDTH's ladder, where net improved monotonically toward zero (x1 -0.046 -> x8 -0.013 on MES) precisely as widening stops shrank position size and therefore commission. THE STRATEGIC IMPLICATION: gross signal on this setup family is roughly +0.02 to +0.04R. The challenge needs +0.25R. Even at ZERO cost the family is an order of magnitude short, so no further price-based variant of it can close the gap. This is the strongest available argument for R2 (a new information axis) over more search on price.

axis

price

search space

5

alpha

0.02

outcome

ladder flat at every rung, 77.5% exceed stop

effect

—

interval

—

Statement

The fade stops out too early: widening the stop, or removing it, will improve per-trade expectancy because the direction is usually right over a longer horizon.

Mechanism written before the number existed

User observation from chart inspection, corroborated on 7 live setups where adverse excursion ran 5-12x the stop distance. If the stop sits inside ordinary intraday noise, it should be converting winners into losers.

{
 "calendar_filter": false,
 "folds": "pooled (NOT the sealed OOS/lockbox split)",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "risk_usd": 175.0,
 "sizing": "fixed risk, so a wider stop buys fewer contracts",
 "span": "2019-05-06..2026-07-03"
}

OUTCOME: the OBSERVATION is confirmed, the HYPOTHESIS is rejected. 77% of trades do have an adverse excursion exceeding the stop, median 3x -- the stop genuinely sits inside ordinary noise. But widening it does not help: win rate climbs steeply (31%->64% on MES) while expectancy stays flat and negative. That is the signature of NO EDGE -- moving along an iso-expectancy curve, trading win rate against loss size. Removing the stop is worst on both instruments and produces single-trade losses of -$4,444 and -$6,349 against a $175 risk budget (25x and 36x), which disqualifies it structurally regardless of expectancy. CAVEAT: controls_passed=False. This V0 baseline reads -0.046R / -0.026R where the sealed verdict read +0.1R. The ladder COMPARISON is valid (all rungs share the simplifications); the ABSOLUTE levels are not. Two implementations disagreeing by 0.15R across 3,295 setups makes X1 MORE urgent, not less.

axis

order flow

search space

1

alpha

0.03

outcome

superseded

effect

—

interval

—

Statement

Signed aggressor imbalance during the breakout leg predicts the outcome of the failed-breakout setup. Specifically: a breakout pushed by weak aggression (low or opposing signed imbalance) fades more reliably than one pushed by strong aggression.

Mechanism written before the number existed

The fade's own story is an order-flow story -- 'the breakout had no participation behind it' -- yet every test to date has used price alone, and price alone is public information that H-SECONDPUSH showed carries only ~+0.03R gross against a +0.25R requirement. Aggressor side is information not visible on a chart. If the fade thesis is mechanically true, the trades tape at the breakout should separate the setups that revert from the ones that resume.

axis

order flow

search space

1

alpha

0.03

outcome

superseded

effect

—

interval

—

supersedes

H4-ORDERFLOW

Statement

Signed aggressor imbalance during the breakout leg predicts the outcome of the failed-breakout setup: a breakout pushed by weak aggression fades more reliably than one pushed by strong aggression.

Mechanism written before the number existed

Unchanged from H4-ORDERFLOW. The fade's own story is an order-flow story that has only ever been tested through price, and H-SECONDPUSH showed price carries ~+0.03R gross against a +0.25R requirement.

axis

order flow

search space

1

alpha

0.03

outcome

r = -0.072, UNDERPOWERED (n_eff 474, floor 0.131)

effect

—

interval

—

supersedes

H4-ORDERFLOW-v2

Statement

Signed aggressor imbalance during the opening-range breakout predicts the DIRECTION of the subsequent move, measured as the signed return from the failure close to the 16:00 ET close, normalised by the range height.

Mechanism written before the number existed

SUPERSEDES H4, which asked whether imbalance improves the FADE. The fade is dead -- its entry cannot be obtained (#3a) -- so that question is moot and its answer would be unusable either way.
The fresh question is prior to any strategy: does the tape carry directional information at the breakout AT ALL? A breakout pushed by heavy one-sided aggression is a different event from one that drifts through on thin flow, and that distinction is invisible in OHLC. If no such information exists, no strategy built on this setup can work and the programme should say so. If it does, the strategy is designed AFTERWARDS, around an entry that is obtainable by construction -- which is the inversion of how the fade was built.

Power plan

{
 "alpha": 0.05,
 "effect": 0.1,
 "power": 0.8,
 "test": "correlation"
}

{
 "outcome": "signed forward return to 16:00 ET, in range-heights",
 "predictor": "signed aggressor imbalance, breakout bar through failure close",
 "sessions": "525 x2 purchased, 453/448 overlapping",
 "test": "Spearman, pooled and per instrument",
 "window": "09:45-10:15 ET"
}

OUTCOME: NOT ESTABLISHED, but the first result this session that was not killed outright. Handle carefully.
Pooled Spearman r = -0.0723 (MES -0.060, MNQ -0.089) -- consistent in SIGN and rough magnitude across both instruments. The pre-specified tercile secondary is a clean monotonic gradient: low imbalance +0.216 range-heights forward, mid -0.030, high -0.198. A spread of 0.41 range-heights between terciles is economically meaningful IF it is real.
DIRECTION, and it is counterintuitive: NEGATIVE r means heavy aggression WITH the breakout predicts REVERSAL, while breakouts that drift through on weak or opposing flow tend to CONTINUE. That is the classic exhaustion shape -- aggressive buyers lifting into a breakout are late, and their flow is the fuel for the move back.
WHY IT IS NOT ESTABLISHED, stated plainly:
(a) UNDERPOWERED FOR WHAT IT FOUND. The sample was powered to detect r=0.10 (n>=783) and passed that gate at n=901. The observed r=0.072 is SMALLER than the effect we powered for; detecting it would need n~1,500.
(b) THE POOLED n IS NOT 901. MES and MNQ are ~90% correlated intraday, so 453+448 setups are not 901 independent observations. Effective n is nearer 453, where the detectable effect is r=0.13 -- almost double what we saw. My power plan treated pooling as free and it is not. That is a defect in the plan, not in the result.
WHAT WOULD SETTLE IT: ~780 INDEPENDENT setups, meaning ~3 years per instrument rather than the ~2 we bought, and the correlation between instruments handled explicitly rather than by pooling. NOT bought on the strength of this -- a suggestive gradient is exactly what a second look is most likely to flatter.

SCRIPT: experiments/H5-FLOW-PREDICTS/2026-08-01-first-look/exp_flow_predicts.py
OUTPUT: experiments/H5-FLOW-PREDICTS/2026-08-01-first-look/output.txt

axis

price (1-minute bars, owned)

search space

3

alpha

0.05

outcome

-0.00036 ATR, d -0.0015; no forward information

effect

—

interval

—

Statement

A fair value gap (3-bar imbalance) carries directional forward information. Specifically: after price returns to the gap's consequent encroachment (50%), the signed forward return in the direction of the displacement that created the gap has mean > 0 over a 30-minute horizon, in ATR units.

Mechanism written before the number existed

ICT doctrine holds that a displacement leg leaves an imbalance -- one side of the book was not traded -- and that price returns to 'rebalance' it before continuing. Stated as a market claim rather than a metaphor: a fast directional move leaves resting interest unfilled; that interest is still there when price returns, so the return is met by the same direction of pressure that caused the move. If true, a touch of the gap midpoint marks a point where the original direction resumes.

This is TESTED AS A PRIMITIVE, not as a strategy, and that is the whole point of registering it this way. All four strategies in the source playbook -- 2022 Model, Silver Bullet, Judas Swing, Turtle Soup -- enter with a limit at the FVG or its CE. The entry is their only shared component. If the gap carries no forward information, all four die on this one register entry rather than four.

The outcome is a strategy-free signed forward return: no entry assumption, no stop, no target. This is deliberate. The fade's defect entered through a P&L outcome that smuggled in an unobtainable fill, and a signed return from a touched price to a later price cannot carry one.

PRIOR, stated before the number: I expect this to be at or near zero. FVG detectors ship with every charting platform, which makes this the most crowded thing on our map. The one argument the other way is that crowding at a LIMIT level can be self-fulfilling -- resting orders are the liquidity -- where crowding at a breakout is self-defeating.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.07,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "test": "mean_shift"
}

{
 "days": 1849,
 "entry_ref": "CE (gap midpoint)",
 "fvg": "3-bar imbalance, first of RTH session",
 "horizon_bars": 30,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "seed": 20260803,
 "unit": "daily ATR"
}

ICT-P1. Strategy-free: signed forward return from a DERIVED CE touch. Primary conditional on touch; unconditional (no touch = 0) reported as pre-specified secondary. Mirror-level control for the fill rate.

SCRIPT: experiments/ICT-P1-FVG/r1/exp_ict_fvg.py
OUTPUT: experiments/ICT-P1-FVG/r1/output.txt

axis

price (1-minute bars, owned) -- control on ICT-P2-SWEEP

search space

1

alpha

0.0

outcome

drift -0.008, CI crosses zero; the corrected reading

effect

—

interval

—

supersedes

ICT-P2-SWEEP

Statement

P2's -0.159 ATR reversal reading is dominated by the reference price, not by forward drift. Decomposing total = penetration + drift, the penetration term accounts for the majority of it, and the drift term is small.

Mechanism written before the number existed

Stated from construction, before the decomposition is run. On an up-sweep bar price is ALREADY above the swept level when that bar closes. Measuring forward return from the level therefore includes the penetration distance, which is positive by construction. A market that froze the instant it swept would still print a positive continuation reading. The identity total = (C[i]-level) + (C[end]-C[i]) is exact, so this control cannot manufacture a better number -- the parts must sum to P2's r1 result. That is what makes it a control and not a second look at the hypothesis.

Registered because P2 r1 returned Cohen d = -0.43 where R1 established the gross signal on these instruments at +0.02 to +0.04R. An effect six times the declared detectable floor, appearing suddenly on the family we have already killed twice, is far more likely to be an artifact of measurement than a discovery. That is the #3a lesson: the fade also passed a sealed verdict before anyone asked what price it was using.

{
 "horizon_bars": 30,
 "identity": "total = penetration + drift",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "seed": 20260803
}

Control on ICT-P2-SWEEP r1. Must reproduce its total exactly.

SCRIPT: experiments/ICT-P2-CONTROL/r1/exp_ict_sweep_decomp.py
OUTPUT: experiments/ICT-P2-CONTROL/r1/output.txt

axis

price (1-minute bars, owned)

search space

3

alpha

0.05

outcome

-0.159 ATR, d -0.43 -- ARTIFACT, 94.8% reference price

effect

—

interval

—

Statement

A sweep of the prior session's RTH extreme carries directional forward information in the REVERSAL direction. After the first sweep of the prior day's high (low), the signed forward return measured short (long) has mean > 0 over 30 minutes, in ATR units.

Mechanism written before the number existed

The claim is that obvious highs and lows hold resting stop orders, that price is drawn to them because filling large orders requires the liquidity those stops provide, and that once taken the move reverses because the fuel is spent. Unlike the FVG claim this is a story about participants rather than a statistical property, which is why it scores 'fair' and not 'strong' on mechanism.

Tested as a primitive: 2022 Model, Judas Swing and Turtle Soup differ ONLY in which level is swept. Prior-day extreme is the cleanest and most objective of those levels, so it is the one measured. One measurement speaks to three named strategies.

This family is our failed-breakout fade under another name. Three pieces of counter-evidence are already on the record and are stated here BEFORE the run so they cannot be recalled selectively afterwards: R1 measured the gross signal at +0.02 to +0.04R against a +0.25R requirement; H-STOPWIDTH found expectancy flat across a four-rung stop ladder, so a tighter stop has already been probed on this family and did nothing; and Z0 priced four placeable orders on it, all negative. What is genuinely new is only the LEVEL (prior-day rather than opening-range) and the strategy-free outcome.

PRIOR: I expect zero or a small NEGATIVE, i.e. continuation rather than reversal, because a sweep of an obvious level is also a breakout, and the breakout reading is the more crowded one.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.07,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "test": "mean_shift"
}

{
 "horizon_bars": 30,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "level": "prior included RTH session high/low",
 "seed": 20260803,
 "sign": "reversal",
 "unit": "daily ATR"
}

ICT-P2. Strategy-free reversal-signed forward return from the swept LEVEL. Ambiguous bars breaking both extremes in one minute are dropped rather than guessed -- intrabar order is unknowable.

SCRIPT: experiments/ICT-P2-SWEEP/r1/exp_ict_sweep.py
OUTPUT: experiments/ICT-P2-SWEEP/r1/output.txt

axis

objective (capability, not a market question)

search space

1

alpha

0.0

outcome

P(pass) 0.042 vs P(first payout) 0.012 -- 73% of qualifiers never paid

effect

—

interval

—

supersedes

E2-OBJECTIVE

Statement

The engine can express P(first payout), not only P(pass). Demonstrated on the fade, whose entry is already known unobtainable, so that no decision rides on the number.

Mechanism written before the number existed

NOT A HYPOTHESIS TEST, and registered at ZERO ALPHA for that reason. The fade is dead on obtainability (#3a); no payout figure can revive it, so nothing rides on the result and no alpha is spent. What is being established is that the question can be asked at all.

The gap it closes: consistency rules differ BY STAGE and the engine held only the first. The tier we sealed has no evaluation consistency requirement; the qualified tier applies a largest-day rule to PAYOUT eligibility. A strategy can therefore qualify comfortably and never be paid, and ChallengeState alone could not express that. Every P(pass) this programme has quoted -- E2's 0.042, the 0.55 gate, the whole feasibility frontier -- measured getting through the door.

THE PAYOUT PARAMETERS ARE ILLUSTRATIVE AND UNVERIFIED. We verified the evaluation geometry and never verified the funded-stage terms. Every number this produces is conditional on parameters that a user must confirm before any of it is cited. Stated here rather than discovered in a footnote later.

{
 "eval_horizon": 90,
 "funded_horizon": 180,
 "instrument": "MES",
 "payout": {
  "VERIFIED": false,
  "consistency_frac": 0.4,
  "min_days": 5,
  "payout_frac": 1.0,
  "threshold": 2000,
  "trailing_dd": 2000
 },
 "risk_usd": 175.0
}

Q1 capability demonstration, zero alpha. Payout parameters are ILLUSTRATIVE and unverified — the figures show a shape, not a venue.

SCRIPT: experiments/Q1-PAYOUT-PATH/r1/exp_q1_payout.py
OUTPUT: experiments/Q1-PAYOUT-PATH/r1/output.txt
CALCS: none — the script called no audited estimator

axis

objective (capability, not a market question)

search space

1

alpha

0.0

outcome

4.2% pass, 44.0% breach, 35.4% timeout ahead, 16.4% timeout underwater

effect

—

interval

—

supersedes

E2-OBJECTIVE

Statement

Every failed attempt can be assigned exactly one principal reason, and those reasons plus P(pass) sum to one. Demonstrated by decomposing E2's already-recorded runs.

Mechanism written before the number existed

NOT A HYPOTHESIS TEST — zero alpha. The decomposition asks nothing new of the market; it re-reads runs E2 already produced. No number here can revive the fade.

The gap: E2 reported P(pass) 0.042 and P(breach) 0.440, which do not sum to one, so roughly half of all attempts ended in some unnamed way. 'Unnamed' contains at least three genuinely different outcomes with genuinely different remedies -- an attempt that ran out of horizon while still alive wants MORE DAYS, one that breached wants LESS SIZE and more days would only have breached sooner, and one blocked by a consistency rule wants neither because it had the money and could not claim it. A single probability is a scoreboard; this is a diagnosis.

The rule that makes it honest: EXACTLY ONE REASON PER PATH, chosen by a DECLARED precedence rather than by whichever check happens to run first -- the latter reads as an ordering detail and behaves as a definition. Shares are computed over ALL attempts including passes, so P(pass) plus the shares sums to one; reporting shares of failures only would invite double-counting.

One reason exists purely to protect a past lesson: NEVER_TRADED. Verdict #1 was void for instrument failure, and a run whose equity never moved must be named as that rather than counted as a market result.

{
 "change": "TIMEOUT split into in_profit vs underwater",
 "horizon": 90,
 "instrument": "MES",
 "risk_usd": 175.0,
 "supersedes_run": "r1"
}

Q2 r2. r1 put every surviving-but-short attempt in one TIMEOUT bucket. Its own example detail showed an attempt '4,032 short of the 3,000 target' — i.e. DOWN 1,032, not nearly there. 'Ran out of days' implies the remedy is a longer horizon; for a losing strategy more days produce more BREACHES. Not estimand-shopping: this is a zero-alpha descriptive bucketing, no tested quantity changed, and r1 stays on the record.

SCRIPT: experiments/Q2-ATTRIBUTION/r2-timeout-split/exp_q2_attribution.py
OUTPUT: experiments/Q2-ATTRIBUTION/r2-timeout-split/output.txt
CALCS: none — the script called no audited estimator

{
 "horizon": 90,
 "instrument": "MES",
 "precedence": [
  "breach_drawdown",
  "never_traded",
  "consistency",
  "min_days",
  "timeout"
 ],
 "risk_usd": 175.0
}

Q2 capability demonstration, zero alpha. Decomposes E2's own runs.

SCRIPT: experiments/Q2-ATTRIBUTION/r1/exp_q2_attribution.py
OUTPUT: experiments/Q2-ATTRIBUTION/r1/output.txt
CALCS: none — the script called no audited estimator

axis

price (1-minute bars, owned; no new data)

search space

6

alpha

0.05

outcome

null

effect

-0.0213

interval

[-0.0893, +0.0523]

Statement

At a FIXED stop distance, raising the target multiple (target_r) improves expected value on a funded-account contract. PRIMARY: EV in dollars on the best-ranked contract is monotone non-decreasing in target_r across the declared ladder {1.0, 1.5, 2.0, 3.0, 4.0, 6.0}. SECONDARY: net R per trade across the same ladder, with a cluster-bootstrap interval per rung.

Mechanism written before the number existed

A funded-account evaluation is a BARRIER-CROSSING problem -- reach +target before a trailing drawdown ratchets up behind you -- not an expectancy problem. For a barrier, asymmetry dominates the mean: FEASIBILITY.md's frontier shows the required win rate falling from 60-65% at a 1.0R payoff to 45% at 2.0R.

The evidence for running it forwards comes from running it backwards. The archived H-STOPWIDTH ladder widens the STOP at a fixed target, which mechanically LOWERS the payoff ratio: MES goes 2.09 -> 1.21 -> 0.79 -> 0.54 across x1..x8. Net R per trade IMPROVES over that range (-0.046 -> -0.013) because cost in R falls 15x as the fixed per-contract cost is amortised over a larger per-contract risk -- and yet EV through the contract menu COLLAPSES (-$38 -> -$269). Expectancy and EV disagree, and the barrier is why.

So the mechanism predicts the mirror image: moving the TARGET outward at a fixed stop raises the payoff ratio, and if the market's hit-rate decay with target distance is slower than the payoff gain, EV rises. This axis has never been tested. The sealed grid varied k_stop and width_max and recorded 'Fixed, not searched: target = far side'.

BOTH DIRECTIONS, INTERPRETED IN ADVANCE. A rising EV curve means asymmetry is the untested lever and the axis reopens -- NOT that the ORB family reopens, which is closed on its own evidence. A flat or falling curve means hit-rate decay exactly offsets the payoff gain, and the 'exhausted' conclusion then extends to the one dimension that had never been searched, which is a stronger negative than the programme currently holds.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.05,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "search_space_size": 6,
 "test": "mean_shift"
}

{
 "cluster_size": 2,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "intra_cluster_r": 0.9,
 "ladder": [
  1.0,
  1.5,
  2.0,
  3.0,
  4.0,
  6.0
 ],
 "range_minutes": 15,
 "risk_usd": 175.0,
 "seed": 20260830,
 "stop_range": 0.5,
 "supersedes_run": "2026-08-30-ladder (refused by the power gate: no top-level n emitted)",
 "vehicle": "ORB (closed family)"
}

T1. Target ladder at a fixed stop -- the axis the sealed grid recorded as "Fixed, not searched".

SCRIPT: experiments/T1-TARGET-LADDER/2026-08-30-ladder-v2/exp_t1_target_ladder.py
OUTPUT: experiments/T1-TARGET-LADDER/2026-08-30-ladder-v2/output.txt
CALCS: experiments/T1-TARGET-LADDER/2026-08-30-ladder-v2/calcs.json

axis

price (1-minute bars, owned; no new data)

search space

4

alpha

0.05

outcome

null

effect

+0.0230

interval

[-0.0361, +0.0840]

Statement

Selecting trades by their COST IN R improves net expectancy on the retained subset. PRIMARY: mean net R per trade on trades whose cost-in-R falls below a threshold exceeds the unconditioned mean by more than the detectable floor, across the declared threshold ladder {no filter, 0.10, 0.07, 0.05}.

Mechanism written before the number existed

Computed today, not assumed. Cost in R = (fixed per-contract cost) / (per-contract risk), and per-contract risk = (stop_dist + slippage) * multiplier + 2 * commission. The contract count cancels, so cost in R is INVARIANT to risk per trade -- but it varies 15x with stop distance on MES (0.273 at a 2-point stop to 0.018 at 40 points), because a fixed cost is amortised over a larger per-contract risk.

Stop distance is set by the opening-range height, which is known BEFORE the trade is taken. So cost in R is an ex-ante observable, and filtering on it removes the most cost-burdened trades without touching the target, the stop multiple, or the signal.

This is the one way to raise net expectancy that does NOT buy it with variance. T1 established that pushing the target outward raises the mean (+0.0816R across the ladder) while raising the standard deviation from 0.79 to 1.71, so the detectable floor grew 0.053 -> 0.114 and the test lost power exactly where the answer became interesting. Filtering on cost changes the mean through the cost term while leaving the per-trade return distribution's shape alone.

BOTH DIRECTIONS, INTERPRETED IN ADVANCE. A monotone improvement across the ladder means cost burden is a selectable property and the geometry axis has a second lever. A flat or falling curve means the trades with low cost-in-R are also the trades with worse gross outcomes -- large ranges being harder to trade -- and the two effects cancel, which closes cost-conditioning and is a stronger negative than the programme holds today.

Power plan

{
 "alpha": 0.05,
 "cannot_detect": "the 0.024R gap to EV break-even",
 "cluster_size": 2,
 "effect": 0.1,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "search_space_size": 4,
 "test": "mean_shift"
}

{
 "cannot_detect": "the 0.024R gap to EV break-even; smallest available floor is 0.0526R",
 "cluster_size": 2,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "intra_cluster_r": 0.9,
 "ladder": [
  "nofilter",
  0.1,
  0.07,
  0.05
 ],
 "range_minutes": 15,
 "risk_usd": 175.0,
 "seed": 20260830,
 "stop_range": 0.5,
 "target_r": 2.0,
 "vehicle": "ORB (closed family)"
}

T2. Cost-in-R is an ex-ante observable (it follows from the opening-range height), so filtering on it raises the mean through the cost term without touching target, stop multiple or signal -- the one lever found so far that does not buy the mean with variance. Registered with its power limit stated: this cannot resolve +EV and does not claim to.

SCRIPT: experiments/T2-COST-CONDITIONED/2026-08-30-cost-ladder/exp_t2_cost_conditioned.py
OUTPUT: experiments/T2-COST-CONDITIONED/2026-08-30-cost-ladder/output.txt
CALCS: experiments/T2-COST-CONDITIONED/2026-08-30-cost-ladder/calcs.json

axis

price (1-minute bars) on metals and energy -- NEW instruments, not owned

search space

1

alpha

0.05

outcome

NOT RUN - programme closed before the data was bought

effect

—

interval

—

Statement

The cost-in-R filter that improved net expectancy on index futures also improves it on metals and energy. PRIMARY, single pre-specified test with NO freedom: on MGC and MCL alone, mean net R per trade at cost_R <= 0.05 exceeds the unfiltered mean on the same instruments by more than the detectable floor. The threshold 0.05, the vehicle (ORB, stop_range=0.5, target_r=2.0), the risk ($175) and the range window (15 min) are all FIXED at T2's values and may not be re-searched. SECONDARY: the pooled four-instrument estimate, reported only if the validity gates below pass.

Mechanism written before the number existed

T2 established the direction on MES and MNQ: net R rose monotonically as the cost filter tightened (-0.0851 unfiltered -> +0.0230 at cost_R <= 0.05, keeping 36% of trades), an improvement of +0.1081R against a floor of 0.1154R -- 94% of the way and SHORT, with an interval crossing zero.

The mechanism is arithmetic rather than a story about participants: cost in R = fixed per-contract cost / per-contract risk, so it falls as the stop widens, and the stop follows the opening-range height which is known before the trade. Nothing in that reasoning is specific to equity index futures, so it should hold on metals and energy -- and if it does not, the mechanism was never what produced T2's number.

BOTH DIRECTIONS, INTERPRETED IN ADVANCE. Replication means the effect is a property of trade geometry rather than of two correlated index contracts, and the pooled sample then resolves what T2 could not. Failure means the effect is either instrument-specific or was noise dressed as a monotone trend -- and a monotone ladder of four rungs is exactly what noise looks like often enough to matter. EITHER ANSWER IS WORTH $19.16; the failure is the more informative one, because it would retire the only positive number this programme has produced.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.1081,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "precondition": "measure cross-asset rho before scoring",
 "search_space_size": 1,
 "test": "mean_shift"
}

axis

price (1-minute bars, owned; no new data)

search space

4

alpha

0.05

outcome

null

effect

-0.0218

interval

[-0.0635, +0.0195]

Statement

The cost-in-R filter improves net expectancy at target_r=1.0 as it did at 2.0. PRIMARY: across the same declared threshold ladder {no filter, 0.10, 0.07, 0.05}, mean net R per trade is monotone non-decreasing as the threshold tightens, and the improvement from unfiltered to the tightest rung exceeds the detectable floor.

Mechanism written before the number existed

A ROBUSTNESS CHECK, not a search for a better number, and it is registered that way so it cannot be reported as the latter.

cost_R = fixed per-contract cost / per-contract risk, and both terms come from the opening-range height. It is therefore a property of the SETUP and independent of the target: the filter selects the SAME TRADES at any target_r. Only the outcome distribution differs. So if cost filtering is a real mechanism it must improve expectancy at 1.0 as it did at 2.0, and if it works only at the target it was found on, it is an interaction or it is noise.

There is a second, statistical reason to prefer 1.0. T1 measured the standard deviation of net R per trade at each target: 0.79 at 1.0 rising to 1.71 at 6.0. The detectable floor scales with sd, so the same filter at 1.0 is measured against a floor roughly 25% lower than T2's -- attacking the floor through VARIANCE rather than through sample size, which is the only lever left that costs nothing.

BOTH DIRECTIONS, INTERPRETED IN ADVANCE. Monotone and clearing its floor CORROBORATES T2 on an independent target and makes the mechanism claim substantially harder to dismiss. Flat, non-monotone, or failing its floor UNDERMINES T2: a four-rung monotone ladder that does not reproduce one target step away is what noise looks like, and this programme has already published three results that looked real and were not.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.1,
 "intra_cluster_r": 0.3818,
 "power": 0.8,
 "search_space_size": 4,
 "test": "mean_shift"
}

{
 "cluster_size": 2,
 "corroborates_or_undermines": "T2-COST-CONDITIONED at target_r=2.0",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "intra_cluster_r": 0.3818,
 "ladder": [
  "nofilter",
  0.1,
  0.07,
  0.05
 ],
 "range_minutes": 15,
 "rho_note": "MEASURED on 1,637 paired days, not the asserted 0.9",
 "risk_usd": 175.0,
 "seed": 20260830,
 "stop_range": 0.5,
 "target_r": 1.0,
 "vehicle": "ORB (closed family)"
}

T4. Robustness check on T2, registered as one. cost_R is a property of the SETUP (it follows from the opening-range height), so the filter selects the same trades at any target -- if the mechanism is real it must work here too. Also attacks the floor through variance rather than sample size: T1 measured sd 0.79 at target_r=1.0 against 1.05 at 2.0.

SCRIPT: experiments/T4-COST-FILTER-AT-LOW-TARGET/2026-08-30-low-target/exp_t4_cost_low_target.py
OUTPUT: experiments/T4-COST-FILTER-AT-LOW-TARGET/2026-08-30-low-target/output.txt
CALCS: experiments/T4-COST-FILTER-AT-LOW-TARGET/2026-08-30-low-target/calcs.json

axis

price (1-minute bars, owned)

search space

2

alpha

0.05

outcome

detectable

effect

-0.0890

interval

[-0.1345, -0.0431]

Statement

Volatility compression predicts subsequent expansion. PRIMARY: the count of consecutive contracting RTH sessions ending at day N-1 is POSITIVELY rank-correlated with day N's realised RTH range divided by its trailing ATR. SECONDARY: day N's opening-range height over its trailing 20-day median is NEGATIVELY rank-correlated with the post-range expansion of the same day.

Mechanism written before the number existed

Volatility clusters and mean-reverts. Low-volatility states are followed by higher-volatility ones more often than chance. This is the strongest prior available to this programme because it is a DOCUMENTED STATISTICAL PROPERTY of financial time series rather than a story about what participants are thinking -- every other family in the catalogue rests on the latter.

WHY THE PRIMARY USES PRIOR DAYS ONLY, decided on design grounds before any number exists. The obvious predictor -- today's opening range -- is contaminated: the opening range and the rest of the same session are both measurements of that day's volatility, and intraday volatility PERSISTS. A quiet open genuinely predicts a quiet day. That confound would swamp any across-day mean reversion and the result would be uninterpretable. The consecutive-contraction count is computed entirely from COMPLETED PRIOR SESSIONS, so it cannot contain any part of the outcome. This is the C2 lesson (a predictor partly ALGEBRAIC with its outcome) and the ICT-P2 lesson (a measurement contaminating its own answer) applied at the design stage rather than discovered afterwards.

TWO DIRECTIONS, BOTH INTERPRETED IN ADVANCE, so neither can be reinterpreted after the fact:
  * PRIMARY POSITIVE -> the mechanism holds: a longer run of contracting sessions is followed by a larger range.
  * PRIMARY NEGATIVE -> the mechanism is refuted in the sharpest way: compression predicts MORE compression, i.e. volatility persists across days rather than mean-reverting at this horizon.
  * PRIMARY ~ZERO -> compression carries no information about what follows, and the family is dead.
  * SECONDARY POSITIVE is EXPECTED EVEN IF THE MECHANISM IS FALSE, because of the same-day persistence named above. It is reported as a confound check, not as evidence.

The outcome is strategy-free -- a realised range, with no entry, no stop and no target -- because a P&L outcome is how the fade's unobtainable fill entered the record. The eventual entry, if this ever advances, is a STOP order beyond the compressed range, which arms while price is inside it and is therefore obtainable by construction. That obtainability is why this family was worth queuing first.

PRIOR: genuinely uncertain, which is unusual here and is the point. Volatility clustering is real and well documented; whether it is exploitable at a ONE-DAY horizon on two of the most heavily traded futures on earth is a different question, and R1 established that public price structure on these instruments carries +0.02 to +0.04R gross against a +0.25R requirement. I expect a small positive that fails to clear the frontier.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.07,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "test": "correlation"
}

{
 "floor": 0.07,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "median_window": 20,
 "outcome_primary": "day N RTH range / trailing ATR",
 "outcome_secondary": "same-day post-range expansion / ATR",
 "predictor_primary": "consecutive contracting RTH sessions ending N-1",
 "predictor_secondary": "opening-range height / trailing 20d median",
 "unit": "Spearman r"
}

X1. First experiment on the completed maths engine (M1-M7): all statistics from occams.stats, calibration-gated, calcs.json written. PRIMARY uses PRIOR SESSIONS ONLY -- the same-day opening range is contaminated by intraday volatility persistence, and that was decided on design grounds before any number existed. Both directions were interpreted in the registration.

SCRIPT: experiments/X1-COMPRESSION/r1/exp_x1_compression.py
OUTPUT: experiments/X1-COMPRESSION/r1/output.txt
CALCS: experiments/X1-COMPRESSION/r1/calcs.json

axis

instrument-trust

search space

1

alpha

0.0

outcome

reproduces verdict #3 to 3 d.p.; instrument trusted

effect

—

interval

—

Statement

The comparator's -0.046R and the sealed verdict's +0.1R can be reconciled, and the discrepancy is explained rather than assumed.

Mechanism written before the number existed

Registered as a gate, not a bet: a lab whose two implementations disagree by 0.15R across 3,295 setups cannot evaluate anything, and cross-implementation disagreement is the defect class that voided verdict #1.

{
 "engine": "occams/fade.py via the sealed loader",
 "k_stop": 0.2,
 "range_minutes": 15,
 "risk_usd": 175.0,
 "width_max": null
}

X1 CLOSED. The sealed engine reproduces verdict v3 to three decimal places, so the engine is correct and the comparator was wrong. THE CAUSE IS THE ENTRY MECHANISM, and it is not a bookkeeping detail: fade.py books a fill AT the boundary at the failure close, which in reality requires an order already RESTING there -- a stop armed during the breakout, while price is outside the range and the boundary is a valid stop level. Protocol #4a instead places a LIMIT after the failure close. That fills 95% of the time, so it is not a missed-trade problem; it is a TIMING problem. Entering on the retest forfeits 0.129R (MES) and 0.193R (MNQ) -- the entire measured edge. CONSEQUENCE: the live paper campaign is currently running a variant with NO edge, and protocol #4a needs superseding by #4b. I identified this reading on 2026-07-31, labelled it 'most faithful to the engine', and recommended against it for operational convenience. That was the wrong call and this is the measurement of what it cost.

axis

price (1-minute bars, owned)

search space

2

alpha

0.05

outcome

null

effect

+0.0025

interval

[-0.0433, +0.0483]

Statement

The overnight gap carries DIRECTIONAL information about the session that follows. PRIMARY: the signed gap, (RTH open minus prior RTH close) / trailing ATR, is POSITIVELY rank-correlated with the signed forward return over the first session hour, measured from the close of the 09:30-09:31 bar.

Mechanism written before the number existed

The gap prices everything that happened while the market was shut. If that repricing were complete at the bell there would be nothing left to trade; the claim is that absorption is NOT instantaneous, so a large gap leaves the market still moving toward the new level during the first hour. That is the underreaction reading, and it predicts CONTINUATION.

The competing folklore reading is overreaction -- 'gaps fill' -- which predicts the opposite sign. Both are widely believed, which is exactly why the direction is committed to here rather than after the number exists.

DIRECTIONS INTERPRETED IN ADVANCE:
  * PRIMARY POSITIVE -> incomplete absorption; the mechanism holds.
  * PRIMARY NEGATIVE -> overreaction; gaps fade rather than extend. A real finding, and the opposite strategy.
  * PRIMARY ~ZERO -> the gap is fully priced at the bell and carries no directional information. Family dead.

REFERENCE PRICE, decided by design and not by default (M2). The gap is known at 09:30 from the OPEN; the measurement starts from the CLOSE of the first minute, by which time the gap is observable and an order could have been placed. Predictor and outcome therefore share no bar. Measuring from the open would fold the first minute's reaction into both, which is the ICT-P2 defect in a new place.

CONFOUND, NAMED BEFORE THE RUN, exactly as in X1. Gap SIZE almost certainly predicts subsequent VOLATILITY -- a big gap means a busy day -- and that is volatility persistence, not directional information. It is registered as a secondary CONTROL and a large positive there is expected even if the primary is zero. In X1 the analogous control came back at five times the detectable floor while the hypothesis itself was refuted; it will not be read as evidence here either.

PRIOR: I expect approximately zero. Index futures trade nearly 23 hours, so the 'overnight' gap on MES and MNQ is a gap across a brief maintenance break rather than a genuine information vacuum -- the structural argument that makes gap strategies work in equities is much weaker here, and that is worth saying before rather than after. What keeps it in the queue is that it is structurally DIFFERENT from every family that has already died, all of which traded intraday price structure.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.07,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "test": "correlation"
}

{
 "control": "|gap| -> |forward move|",
 "floor": 0.07,
 "horizon_bars": 60,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "outcome": "signed forward return, 60 bars from the close of bar 0",
 "predictor": "(RTH open - prior RTH close) / trailing ATR, signed",
 "reference": "AT_PRICE \u2014 predictor and outcome share no bar"
}

X2. Serial after X1 (closed DEAD). Strategy-free outcome. Both directions and the volatility-persistence confound were interpreted in the registration before any number existed.

SCRIPT: experiments/X2-OVERNIGHT-GAP/r1/exp_x2_gap.py
OUTPUT: experiments/X2-OVERNIGHT-GAP/r1/output.txt
CALCS: experiments/X2-OVERNIGHT-GAP/r1/calcs.json

axis

order flow (trades tape, 525 sessions x2, owned)

search space

2

alpha

0.05

outcome

null

effect

+0.0260

interval

[-0.0610, +0.1125]

Statement

The SIZE DISTRIBUTION of aggressive prints carries directional information. PRIMARY: the share of 09:45-10:00 volume transacted in prints at or above that window's 90th-percentile print size is POSITIVELY rank-correlated with the window-direction-signed forward return over the following hour, measured from the 10:00 bar close.

Mechanism written before the number existed

Aggressor IDENTITY is invisible in the tape. Aggressor SIZE is not. A move carried by a few large prints is a different event from one carried by many small ones: the first is consistent with a participant who has a reason and more to do, the second with retail flow and market-making churn. If size carries any signal about who is behind a move, a move dominated by large prints should extend further.

This is the ONLY NON-PUBLIC AXIS this programme holds. Every family that has died -- ORB, the fade, second push, ICT, X1, X2 -- traded price structure visible to everyone with a chart. That is why X3 stayed in the queue after two price families died.

MULTIPLICITY, STATED PLAINLY: H5 already spent alpha on this same tape, asking whether aggressor IMBALANCE (buy vs sell volume) predicts direction. That found r = -0.072, underpowered, unconfirmed. X3 is a DIFFERENT QUESTION OF THE SAME DATA -- distribution rather than balance -- and the register must show both so the multiplicity is visible rather than inferred.

THE POWER LIMITATION, AND IT IS SEVERE. We own 525 sessions per instrument, so at most 1,050 raw observations and 552 independent-equivalent after the MES/MNQ cluster discount. The 0.07 floor used in X1 and X2 is IMPOSSIBLE here and the power gate would rightly refuse it. The declared effect is therefore 0.13, which this sample can see with margin. The consequence must be stated BEFORE the run and not discovered after: **X3 can only refute a LARGE effect. A null here means 'no effect big enough for 552 effective observations to see', which is a materially weaker statement than X1's or X2's nulls.** Declaring a bigger effect because the sample is small is only honest if that consequence is carried with it.

DIRECTIONS INTERPRETED IN ADVANCE:
  * POSITIVE -> large prints mark informed flow that extends. The mechanism holds and it is the first thing in this programme that has.
  * NEGATIVE -> large prints mark exhaustion or absorption; a move carried by size is more likely to reverse. A real finding, opposite strategy.
  * ~ZERO -> print size carries no directional information at a scale this sample can detect.

CONFOUND, NAMED BEFORE THE RUN -- the third time, and by now the expected pattern. Large-print share almost certainly predicts subsequent VOLATILITY: a chunky tape is a busy tape. It is registered as a control and a large positive there is expected even if the primary is zero. X1's analogous control returned 5.00x the floor and X2's 2.33x, while both hypotheses were refuted.

SEPARATION OF PREDICTOR AND OUTCOME. Predictor is computed from 09:45-10:00 only; the outcome starts at the 10:00 close. They share no bar and no print. Momentum, if it exists in this window, shifts the MEAN signed return but cannot create a correlation with print-size share -- so the primary is clean of it by construction.

PRIOR: the most uncertain of the queue, which is why it was kept. H5's first look at this tape was the only suggestive result this programme has produced. I expect a small positive that fails to clear 0.13.

Power plan

{
 "alpha": 0.05,
 "cluster_size": 2,
 "effect": 0.13,
 "intra_cluster_r": 0.9,
 "power": 0.8,
 "test": "correlation"
}

{
 "control": "share -> |forward move|",
 "direction": "sign of the 09:45->10:00 price move",
 "floor": 0.13,
 "instruments": [
  "MES",
  "MNQ"
 ],
 "outcome": "direction-signed forward return, 60 bars from the 10:00 close",
 "predictor": "share of 09:45-10:00 volume in prints >= window p90 size",
 "reference": "AT_PRICE \u2014 predictor and outcome share no bar or print",
 "sessions_owned": 525
}

X3. Serial after X1 and X2, both DEAD. The only non-public axis we hold. POWER IS THE BINDING CONSTRAINT and the floor is 0.13, not 0.07 — a null here means 'no effect big enough for 552 effective observations to see', which is materially weaker than X1/X2. Stated in the registration before the run. H5 spent alpha on this same tape asking a different question; the multiplicity is on the register.

SCRIPT: experiments/X3-PRINT-SIZE/r1/exp_x3_print_size.py
OUTPUT: experiments/X3-PRINT-SIZE/r1/output.txt
CALCS: experiments/X3-PRINT-SIZE/r1/calcs.json

axis

instrument-trust

search space

3

alpha

0.0

outcome

assumed +0.080/+0.186 R; all four placeable orders negative

effect

—

interval

—

Statement

The sealed engine's entry -- a fill AT the range boundary at the failure close -- is achievable by some real order, and therefore the measured +0.1R is tradeable.

Mechanism written before the number existed

Registered as a verification, not a bet. #4a's failure showed the entry mechanism is worth ~0.13R, so before amending the protocol again the question had to be asked directly: which actual order produces the engine's fill?

{
 "engine": "occams/fade.py via the sealed loader",
 "k_stop": 0.2,
 "risk_usd": 175.0,
 "span": "2019-2026 pooled"
}

Canonical SOP-run of the Z0 finding. The earlier record predates the harness and its option (b) had no frozen script; this run is the reproducible one.

SCRIPT: experiments/Z-ENTRY-IMPLEMENTABLE/2026-08-01-sop-verified/exp_entry_obtainable.py
OUTPUT: experiments/Z-ENTRY-IMPLEMENTABLE/2026-08-01-sop-verified/output.txt

{
 "engine": "occams/fade.py via the sealed loader",
 "k_stop": 0.2,
 "risk_usd": 175.0,
 "span": "2019-2026 pooled"
}

HYPOTHESIS REJECTED, and the consequence is large. EVERY implementable order is negative on both instruments; only the engine's ASSUMED fill is positive.
WHY: at the failure close price is INSIDE the range, so the boundary is a price you cannot obtain. A resting stop fills earlier -- 60.3% of armed fills occur before any failure close, and those lose -0.277R. A market order takes the close, which is worse than the boundary. A limit enters on a retest, which is late. The engine books a fill at the boundary while conditioning on a close that had not yet happened when price was there.
That is LOOK-AHEAD, and it is the entire measured edge: the armed-stop subset where the close DID come first reads +0.122R (MES) / +0.226R (MNQ), but that subset is only selectable after the fact.
IMPLICATION: verdict v3's 'edge real (+0.1R, persists OOS)' is most likely an artifact of an unobtainable entry price. Protocol #4 is validating a non-edge, and H4 would be trying to improve something that is not there. This needs a DATED ADDENDUM to the verdict -- the verdicts are append-only and are never edited.

axis

instrument-trust

search space

1

alpha

0.0

outcome

CLEAN - ORB does not share the fade's defect

effect

—

interval

—

Statement

The ORB verdicts (#1 and #2) do NOT share the fade's defect: every entry the ORB engine books is producible by an order that could actually have been placed at that moment.

Mechanism written before the number existed

The fade booked a fill at the range boundary at the failure close, when the boundary was already behind the market. ORB arms stops OUTSIDE the range while price is inside it, which is the correct side for a stop, so the same failure should not apply. But #1 was already VOIDED for a different instrument defect, so the family's entries have never been verified and reading the code is not checking it.

{
 "checks": [
  "placeable when armed",
  "fill matches the engine"
 ],
 "family": "ORB",
 "instruments": [
  "MES",
  "MNQ"
 ],
 "risk_usd": 175.0,
 "span": "2019-2026",
 "stop_range": 0.5,
 "target_r": 2.0
}

OUTCOME: CLEAN. 6,774 armed orders across 3,482 instrument-days, ZERO unplaceable and ZERO fill disagreements. ORB does NOT share the fade's defect.
WHY it is clean, and the contrast is the useful part: ORB arms its stops OUTSIDE the opening range while price is still INSIDE it. A buy stop above the high and a sell stop below the low are on the correct side at the moment they are placed, so the fill is produced by the order rather than assumed alongside it. The fade booked a fill at the boundary AFTER price had returned inside, when the boundary was already behind the market and nothing was resting there.
The engine's gap handling also checks out: it books max(level, open) + slippage, which is exactly what a real stop produces when a bar opens through the level -- conservative, and matching the gate on all 5,289 fills tested.
CONSEQUENCE: verdicts #1 and #2 are CLEARED of this defect. Their outcomes stand as recorded -- #1 VOID on the instrument defect it was already voided for, #2 NO-GO because the family never worked. The entry-obtainability problem is specific to the fade, not systemic across the sealed record.

SCRIPT: experiments/Z04-ORB-OBTAINABLE/2026-08-01-audit/exp_orb_obtainable.py
OUTPUT: experiments/Z04-ORB-OBTAINABLE/2026-08-01-audit/output.txt