Where the next edge is

Date: 29 August 2026 Status: DECISION Evidence cut: 52 numbered historical trials; 61-row inventory digest; 45 newly ranked candidates; 2022+ lockbox unopened.

STATUS: DECISION · DATE: 29 August 2026. The next return-bearing work is three Tier-B hypotheses, not a claim that edge has been found. Data, power, execution, and decomposition work stays DIAG and does not consume a return trial.

Decision summary

• STATUS: RANKED · DATE: 29 August 2026. A 0 · B 3 · C 0 · D 24 · DIAG 18 across 45 candidates; 34 candidate claims hit a mechanical or judge kill.
• STATUS: RUN ORDER · DATE: 29 August 2026. Rebuild the manifest-backed price read model; restore durable method/data contracts; re-ingest the already-owned fundamentals bundle; then run profitability trend, overnight-component momentum, and iHML in judge order, subject to their pre-run power gates.
• STATUS: LIVE BOUNDARY · DATE: 29 August 2026. Momentum is the only possible 15 September lead, and only after its one-shot lockbox pass, Round-4 receipt cutover, prospective policy-v2/construction decisions, rehearsals, graduation, and owner arm. GP/A requires its own pass and production path. iHML and ETF trend are research slots, not holdings.
• STATUS: FALSIFIER · DATE: 29 August 2026. A positive point estimate below its registered MDE, a failed 2017–2021 gate, a 2×-cost failure, post-2021 contact, an uncosted benchmark, or a non-PIT universe is a refusal, not “promising.”

1. What 52 trials taught us

STATUS: RECONCILED · DATE: 29 August 2026. The inventory contains 61 digest rows: 52 numbered ledger positions, of which one position is the prospective trial-null recency amendment rather than a numbered return trial, plus nine pre-ledger/non-trial programme rows. The literal row-status count is 51 numbered trials because the sequence deliberately skips trial 43. “52 trials” is the programme accounting label; “61 rows” is the digest size.

1. STATUS: ADOPTED · DATE: 29 August 2026. Execution timing is the first attack. Gap fade moved from +96.60%/yr, t=43.17, to −56.17%/yr, t=−25.90 after honest timing; overnight premium moved from +25.52% to −4.20%. Decide at close (t), fill at adjusted open (t+1), exit at the registered next open.
2. STATUS: ADOPTED · DATE: 29 August 2026. Magnitude is a bug detector. Daily-bar return above roughly 30%/yr or Sharpe above 3 is presumed timing or data error until a causal audit survives.
3. STATUS: ADOPTED · DATE: 29 August 2026. A date is not a timestamp. Insider clusters fell from +8.77%/yr, t=3.21, to −3.30%, t=−1.09 under filing-date+2. The pessimistic availability convention is the fundable result.
4. STATUS: ADOPTED · DATE: 29 August 2026. The benchmark is separately costed. Candidate-net minus a separately costed equal-weight eligible universe prevents benchmark-cost rebates.
5. STATUS: ADOPTED · DATE: 29 August 2026. Survive 2× costs and report 3×. Ex-dividend fell from +8.25% at 1× to −3.86% at 2× and −15.97% at 3×.
6. STATUS: ADOPTED · DATE: 29 August 2026. PIT eligibility and all date columns are prerequisites. Trial 49 touched malformed post-2021 rows in a precheck, computed no statistic, and was still charged and marked invalid.
7. STATUS: ADOPTED · DATE: 29 August 2026. Small samples, a failed negative control, or alpha living only on the unavailable short side are decisive. Thirty-five 13F months were no evidence; net share issuance had +6.18% L/S but only +0.45%, t=0.34 long-only.
8. STATUS: ADOPTED · DATE: 29 August 2026. Capacity and tails precede promotion. Both survivors depend on a small right tail; removing the best 14 months leaves momentum +0.41%/yr and GP/A +0.82%/yr.
9. STATUS: ADOPTED · DATE: 29 August 2026. Positive case first, falsifiers second, then a binary verdict. A friendly null does not become a lead.

The two screening survivors

1. Momentum 12−1 long-only. STATUS: PREREGISTERED / LOCKBOX UNOPENED · DATE: 29 August 2026. Fundable top book: +15.02%/yr net, Sharpe 0.49, max drawdown −68.5%; active versus separately costed EW +4.87%, t=1.25. The registered 60-session/10%/1.0-cap overlay produced +6.75%, Sharpe 0.55, drawdown −27.6%, mean leverage 0.428. Positive case: 2×/3× costs remain +13.86%/+12.70%. Falsifiers: right-tail dependence, 143-name whole-share failure at live size, non-PIT cost buckets, absent delisting proceeds, and the unopened lockbox.
2. Gross profitability, GP/A. STATUS: CONDITIONAL PRE-REGISTER / LOCKBOX UNOPENED · DATE: 29 August 2026. Active net +2.827%/yr, Sharpe 0.580, max drawdown −15.4%, NW t=2.461; 2×/3× +2.795%/+2.764%. Positive case: 2017–2021 +5.05%/yr, t=2.40, permutation p=0.010, correlation to momentum 0.20. Falsifiers: best-14-month removal leaves +0.82%, t=0.91; SF1 snapshot revisions, terminal proceeds, and a second production StrategySpec path remain unresolved.

What the two survivors look like together

STATUS: REPORT-ONLY · DATE: 29 August 2026 · trial:null · variants_allowed:0. These diagnostics reproduce the frozen survivor series; they do not select an allocation, amend either strategy, or consume a return trial. No post-2021 market-return rows were read.

• Positive case: 5,767 daily rows map to 275 GP/A exit intervals from 1 February 1999 through 1 December 2021. A vol-scaled 50/50 blend of GP/A total net and momentum total net has 9.97%/yr component-net arithmetic mean ×12, 9.34% CAGR, 14.23% volatility, Sharpe 0.70, and max drawdown −34.41%. Equal-risk weights are 37.14% GP/A / 62.86% momentum; a naive 50/50 allocation contributes 64.99%/35.01% of risk.
• Falsifier / scope: the valid total-net series correlation is 0.69658 Pearson / 0.81512 Spearman over the full sample and 0.82439 / 0.81228 in 2017–2021. The approximately-zero result, −0.04656 / −0.06713 full sample and 0.08965 / 0.13031 in 2017–2021, compares GP/A active with momentum total and therefore describes co-movement only; it is not portfolio diversification evidence. The report has no holdings or trades, so it cannot measure opposite-trade overlap or portfolio-level netting costs.
• Policy-v1 calibration: STATUS: REPRODUCED · DATE: 29 August 2026 · trial:null. All 9 consistency checks pass. H1 fires 12 times. The continuous-HWM H2 episode peaks 7 March 2000, breaches 18 July 2002, troughs 12 March 2003 at −27.629%, and recovers 7 September 2005 after 792 inclusive crossing-to-recovery sessions. Seven soft breaches and one synthetic historical retirement on 9 February 2010 are descriptive, never live authority.

#### Worst 10 narrated drawdown episodes — vol-scaled blend

STATUS: LIMITATION · DATE: 29 August 2026. Narrations identify contemporaneous market windows from general chronology, not causal proof. Distinct episodes do not overlap; exposure, cohort, sector, corporate-action, and per-component contribution fields are unavailable in this diagnostic.

2. Ranked shortlist and pre-registration gaps

Tier A

STATUS: EMPTY · DATE: 29 August 2026. No candidate earned Tier A. This is a result, not a missing section.

Conjunctive lockbox verdict map for every Tier-B look

STATUS: MUST-FIX BEFORE REGISTRATION · DATE: 29 August 2026. None of the three Tier-B outlines below is a sealed pre-registration. A future self-hashed document must bind the rule hash, frozen parameters and exclusions, live contract, trial accounting, and all six post-lockbox gates below. The gates are conjunctive; any miss is REFUSED / NO RETRY.

1. STATUS: REQUIRED · DATE: before registration. Lockbox active net excess is positive and at least 50% of the screening estimate.
2. STATUS: REQUIRED · DATE: before registration. Lockbox Newey–West t-statistic is at least 2.0 on the registered primary difference series.
3. STATUS: REQUIRED · DATE: before registration. At least 3 of 5 lockbox calendar years are positive.
4. STATUS: REQUIRED · DATE: before registration. The registered primary remains positive at 2× costs; 3× is reported.
5. STATUS: REQUIRED · DATE: before registration. DSR is at least 0.95 with raw and clustered N both disclosed.
6. STATUS: REQUIRED · DATE: before registration. The named negative control does not beat the registered book.

Tier B1 — profitability trend, dGP/A primary

STATUS: NOT PREREGISTRABLE — POWER VALUE UNSET · DATE: 29 August 2026. Score 0.06804; judge expected incremental net 1.2%/yr. Publish and seal a numeric MDE and required sample before this can become a pre-registration; the advancement threshold below is not an MDE.

• Hypothesis / positive case: attention-constrained investors under-react to same-fiscal-quarter year-over-year improvement in gross profit on assets. This is the cheapest plausible fundamental complement to GP/A and may partly explain price momentum.
• Universe: byte-identical GP/A survivor eligibility: US common stocks; no category filter; formation closeunadj ≥$10; 20 consecutive trailing sessions of close×volume ADV ≥$10m; latest available ARQ revenue and cost of revenue present, assets present and nonzero. Net income and a missing lag affect rankability, never universe membership.
• Exact signal fields and availability: SF1 ARQ revenue, cor, assets, netinc, datekey, calendardate, reportperiod, permaticker. Use the earliest bounded as-first-reported row available only after datekey+2 exchange sessions. Join the exact four-quarter-back permaticker sequence with a 330–400 day reportperiod guard. Primary (dGPA=((revenue-cor)/assets)t-((revenue-cor)/assets){t-4}), residualized cross-sectionally on GP/A level and registered 12−1 momentum inside the attention-constrained lower-ADV half. dROA and the original 50/50 composite are disclosure-only; no weight search.
• Rebalance / holding: form at each exchange month-end close; fill next adjusted open; 12 staggered annual cohorts; aggregate and net monthly trades; hold each cohort 12 months. The original six-month variant is removed.
• Construction: equal-weight top quintile of the residualized dGP/A rank; primary is candidate net minus a separately costed EW PIT-safe eligible benchmark. Apply the survivor overlay: prior 60 completed net daily returns, one-session lag, 10% annual volatility target, leverage cap 1.0.
• Cost: actual initial, rebalance, forced-exit, terminal, and exposure turnover at 10/25 bps per side; report 1×/2×/3× and an OPG 10/25/40-bps row.
• Primary endpoint: annualized active net return and NW t on the monthly difference series. Advancement requires active net ≥2.3%/yr and NW t≥2, despite the judge prior being lower.
• Screening and recency: every current and lagged date field ≤2021-12-31; screening 1998–2021 where the lag exists; 2017–2021 active net >0 with t≥1.
• Lockbox: UNOPENED. No read, precheck, census, row count, or statistic on a field dated after 2021-12-31.
• Data needed: manifest-backed SEP/ACTIONS/TICKERS read model plus re-ingested SF1 ARQ. Reproduce GP/A from the rebuilt ETL before registering the new look; a positive snapshot result must pass an SEC as-first-reported audit before advancement.
• Expected shape: original active gross +2.0–3.5%/yr; refuter-amended honest net +0.8–1.5%, point 1.2%. Expected 200–450 names/cohort, median position ADV about $18m, binding 1%-ADV sleeve about $25m; publish p10.
• Overlap: expected rho to momentum +0.20–0.45 and GP/A +0.10–0.35. Residualization and within-GP/A/momentum sorts must show incrementality.
• Falsifiers / refuter amendments: prior below MDE means no-run; no increment within GP/A/momentum, momentum rho>0.5, 2017–2021 failure, deteriorating quintile beating EW/top, 2× net ≤0, revised-snapshot dependence, or >6% active as a timing smell refuses. Inspect extreme dROA rows; no variant is available.

Tier B2 — overnight-component momentum

STATUS: NOT PREREGISTRABLE — ENDPOINT CONTRACT INCOMPLETE · DATE: 29 August 2026. Score 0.061875; judge expected incremental net 0.5%/yr. Before registration, freeze the coefficient's forward-return interval and observation timestamp, a numeric advancement threshold/MDE, and the separately costed benchmark. The long-only book remains diagnostic and cannot satisfy the primary endpoint.

• Hypothesis / positive case: the overnight component of the registered 12−1 signal contains information incremental to total momentum; the intraday component should reverse. This extracts new information from owned daily fields without creating an overnight trading schedule.
• Universe: the registered fundable US-common universe, closeunadj ≥$10 and ADV20 ≥$10m with 20 observations. Freeze exclusions for anomalous volume, null/zero volume, and open==close; report the overall and fundable censuses.
• Exact signal fields: SEP open, close, closeadj, closeunadj, volume, date, ticker; ACTIONS splits. Define split factor (sd=closed/closeunadjd), price-only overnight return (r^{on,px}d=(opend sd)/(close{d-1}s{d-1})-1), and a total-return overnight diagnostic only. The 12−1 session span is the registered exact endpoint interval, skipping the formation month. Rank-standardize ON within formation month.
• Primary endpoint: a powered monthly Fama–MacBeth coefficient for ON conditional on registered TOT121, with NW lags 0/5/11. The portfolio head-to-head is diagnostic, not a second primary.
• Rebalance / holding / construction: form month-end close, fill next adjusted open, hold one month, equal-weight top decile for the diagnostic book, roughly 140 names; use the same 60/10%/1.0 overlay.
• Cost: 10/25 bps per side plus 5 bps auction slippage; actual turnover at 1×/2×/3×; midpoint-fill sensitivity. Turnover above 45%/month triggers recosting and a new MDE before interpretation.
• Screening / lockbox: 1998–2021 only; 2017–2021 must be alive, with 2017–2019 and 2020–2021 shown separately. LOCKBOX UNOPENED; no ≥2022 field may be touched.
• Data needed: durable manifest-listed price read model carrying open; ACTIONS split audit; TICKERS only for the disclosed non-PIT cost bucket.
• Expected shape: standalone active gross +5–8% and net +4–7% were the strategist prior; the decision delta versus registered momentum is only −2 to +3 pp and the judge point prior is +0.5 pp. Treat the judge number as controlling.
• Capacity / overlap: original inherited estimate $14.9m median and $9.5m p10 binding sleeve is not accepted without measurement. Expected rho to momentum 0.85–0.97 and GP/A near zero; this is a replacement test, never a second sleeve.
• Falsifiers / refuter amendments: precomputed MDE >3 pp means no-run; failure of split reconciliation, dividend orthogonalization, the intraday negative control, liquidity-above-median effect, or midpoint-fill robustness refuses. A portfolio claim needs a 90% CI excluding −2 pp for noninferiority and zero with t≥2 for superiority. Pre-commit the action: a null keeps the frozen survivor.

Tier B3 — intangible-adjusted value, iHML

STATUS: NOT PREREGISTRABLE — TIMING/CONTROL CONTRACT INCOMPLETE · DATE: 29 August 2026. Score 0.0175; judge expected incremental net 1.0%/yr. Before registration, freeze the acceptance-timestamp-to-exchange-session map, the byte-exact plain-B/M comparator construction, and one named negative control. As-first-reported PIT ingest remains a no-run prerequisite.

• Hypothesis / positive case: expensing R&D and organisation-building SG&A understates book value; capitalizing those investments should restore a value rank specifically in intangible-intensive firms and diversify momentum/GP-A.
• Universe: fundable US common stocks, closeunadj ≥$10, ADV20 ≥$10m; latest available ARQ revenue, cor, assets, and equity present; assets nonzero; 20 contiguous permaticker quarters. Count every burn-in loss. The honest start is empirical, expected around 2003 with the first complete staggered cohort around 2004—not 1998 by assertion.
• Exact signal fields: as-first-reported, vintage-preserving ARQ/ART rnd, sgna, equity, assets, intangibles, revenue, cor, sharesbas, acceptance timestamp, reportperiod, and permanent security identity; SEP open/close/closeadj/closeunadj/volume; ACTIONS. Each filing becomes usable only after its actual acceptance timestamp maps to an exchange session; datekey+2 is a pessimistic diagnostic, not an allowed substitute for PIT vintage. Annual (R=sum rnd), (S=sum sgna); (K^{know}y=.85K^{know}{y-1}+Ry); (K^{org}y=.80K^{org}{y-1}+.30Sy), with the SG&A/R&D overlap convention frozen. Five-year burn-in. (B^*=equity+K^{know}+K^{org}-intangibles). Primary market cap is closeunadj×the latest then-known sharesbas; DAILY market cap is diagnostic. Winsorize adjusted B/M 1/99 and rank high-to-low.
• Rebalance / holding: monthly formation; 12 staggered annual cohorts; next-adjusted-open entry and anniversary-open exit; forced delisting exit under the registered proceeds rule; monthly aggregate netting.
• Construction / endpoint: equal-weight top quintile. Primary is the monthly active-net delta iHML minus plain B/M on the identical panel; industry-neutralization and sales-to-price are mandatory mechanism comparators, not alternate winners. Apply the same 60/10%/1.0 overlay.
• Cost: actual turnover at 10/25 bps per side; separately costed benchmarks; 1×/2×/3×. Expected one-way turnover 4–7%/month.
• Screening / lockbox: post-burn-in through 2021-12-31; every date column hard-bounded. 2017–2021 positive, then 2017–2019 and 2020–2021 splits. LOCKBOX UNOPENED.
• Data needed: price read model plus an as-first-reported/PIT fundamentals store, most directly SEC filing vintages and acceptance headers joined through a survivorship-complete permanent-ID map. The already-owned SF1 snapshot may reproduce field coverage and serve as a diagnostic comparator, but it cannot run the return endpoint. DAILY is not a blocker for the primary.
• Expected shape: original active gross +2.5–4.5%; refuter-amended net +0.8–1.8%, judge-controlled point 1.0% active. More than 4% is presumed a timing/knowability bug. Expected 200–450 names/cohort and $25–30m median binding 1%-ADV sleeve; publish p10 and auction qualifier.
• Overlap: expected rho to GP/A −0.2 to +0.2 and momentum −0.3 to 0.0, but measure the iHML-minus-BM delta against GP/A; |rho|>0.5 kills the diversification claim.
• Falsifiers / refuter amendments: no as-first-reported vintage spine means no-run; delta ≤0 or t<2; 2017–2021 failure; effect absent from high-intangible firms or from both R&D and organisation components; non-null-R&D and industry-neutral failure; bottom quintile beating top/EW; 2× failure; availability/restatement sensitivity; or a missing negative control refuses.

Tier C and Tier D

• STATUS: EMPTY · DATE: 29 August 2026. Tier C has zero candidates.
• STATUS: CLOSED · DATE: 29 August 2026. Tier D has 24 candidates. Twenty-three have decisive killed lines below. The remaining candidate, distress-filtered momentum, was not mechanically killed but scores only 0.00132: its positive case is crash-loss reduction; its falsifier is near-identity to momentum, deletion of the paying right tail, and negligible incremental portfolio value. It receives no return run.

3. DIAG — zero-trial diagnostics and infrastructure

STATUS: RUN WITHOUT RETURN CLAIM · DATE: 29 August 2026. These rows use trial:null / variants_allowed=0. A diagnostic may repair data or explain a survivor; it cannot silently become a sleeve.

4. Portfolio view — 15 September through Q4

1. STATUS: CONDITIONAL · DATE: 15 September 2026. Live now means only a lockbox-passing, production-ready survivor. Momentum may lead. GP/A joins only after its own pass and StrategySpec/risk-policy path.
2. STATUS: RESEARCH SHAPE · DATE: Q4 2026. Eventual 3–4 components: momentum + GP/A + validated iHML/value + one exact post-SFP-audit ETF-trend sleeve. The last two are research slots.
3. STATUS: PLANNING RANGE · DATE: Q4 2026. Two-survivor screening arithmetic is about 0.80%/month net at 50/50 and 0.95% at the reported 79% GP/A inverse-vol mix. Use 0.3–1.0%/month forward, never the desired 1–3%; negative Q4 months remain plausible.
4. STATUS: FALSIFIER · DATE: Q4 2026. Both survivors are right-tail dependent. More sleeves do not manufacture diversification and worsen whole-share rounding at this account size.

LL-4b Round 4, the first-order ladder, and policy v2

STATUS: #348 MERGED PENDING MAIN CI · DATE: 29 August 2026. Round 4 makes live-lane-execution-plan/v3 canonical: producers, CLI loading, submit/reconcile validation, and rehearsal replay reject v2/unknown receipts, so the 4–11 September chains start fresh. Flat plans record applicable:false with zero quantities and null deployment/headroom metrics; emergency kill/flatten bypasses planner gating. The implementability diagnostic calls the same planner kernel for quantities, metrics, and verdicts.

STATUS: CORRECTED DIAGNOSTIC · DATE: 29 August 2026. At $12,700 NAV and 0.428 leverage, 0/277 target months pass all three gates; worst name fraction is 18.0488%, worst deployment 13.9534%, and the latest month has 45/217 names, 16.0794% deployed, and 0.007789× headroom. The former $881,948.56 figure is only the pre-headroom name/deployment minimum. All-name one-share headroom produces $2,521,437,128,668,714.50, driven by a June-2013 raw-price outlier selected by the frozen rule. The diagnostic must not filter it away. Whether live authority uses all-name or per-name/fractional headroom is an owner decision, not an engineering correction.

STATUS: OWNER DECISION · DEADLINE: 3 September 2026. With 143 names, stage 1 at 10 bp, 0.428 overlay, and GBPUSD 1.27, perNameUsd = targetGbp × volScale × GBPUSD / 143, then quantity is floor(perNameUsd/latestPrice). At £100k NAV: £100 → £42.80 → $54.356 total / $0.3801 per name; £5k: £5 → £2.14 → $2.7178 / $0.0190; £10k: £10 → £4.28 → $5.4356 / $0.0380; £25k: £25 → £10.70 → $13.589 / $0.0950. Assuming prices remain at least the registered $10 floor, all stated NAV bases yield 0/143 shares; at £100k, stages 1/2/3 provide only $0.380/$0.950/$1.901 per name. Exact 15-September counts remain unknowable until the 14-September close.

• Option A — larger genuine NAV. STATUS: OWNER DECISION · DEADLINE: 3 September 2026. Raising the £10k cap needs new preregistration and policy authority; even the $10 floor implies at least £2.631m stage-1 NAV for 143 names.
• Option B — sealed prefix N. STATUS: OWNER DECISION · DEADLINE: 3 September 2026. Use the deterministic prefix-N construction below. It may return zero; £10k permits no positive N under the earlier ladder arithmetic, while £25k permits only N=1 in the stated price interval and lies above the current cap.
• Option C — one-share canary. STATUS: OWNER DECISION · DEADLINE: 3 September 2026. The bounded safety canary below requires prospective authority and must be unwound before stage 2.
• Option D — delay. STATUS: SAFE DEFAULT · DEADLINE: 3 September 2026. Place no live order and reseal graduation later.

Recommendation — C canary plus only a viable, sealed B. STATUS: OWNER DECISION · DEADLINE: 3 September 2026. The canary is the only source-recommended way to preserve a first-order operational check. Prefix N is conditional: it is neither a measured positive N at £10k nor evidence of return. If the sealed construction returns zero, choose D.

Amendment v1.2-r2 — rank-bound canary and prefix-N construction

STATUS: PROPOSED — OWNER RATIFICATION AND IMPLEMENTATION HASHES REQUIRED · DATE: 29 August 2026. This amendment is prospective and void unless sealed before lockbox opening and before the formation rank is generated or disclosed. It cannot repair or replace the original full-decile verdict.

1. Rank proof. STATUS: MUST-FIX BEFORE AUTHORITY · DATE: 3 September 2026. Keep strategy and index mechanics byte-identical. The planner rederives the prior-month-end eligible rank from frozen anchors: recent.closeAdj / old.closeAdj - 1, formation closeunadj ≥$10, and 20 formation-bounded bars with mean closeunadj×volume ≥$10m; sort descending score then ticker. Refuse unless the first D=ceil(eligible/10) names match the frozen selector. A v3 formation receipt seals ranked_candidates as exactly ranks 1..D, no traversal outside the frozen top decile, with identity, rank, dates, anchors, prices, score, ADV, derivation, and code/data/calendar hashes. Missing, extra, or mismatched proof refuses; the owner remains blind to names.
2. Canary lifecycle. STATUS: MUST-FIX BEFORE AUTHORITY · DATE: 3 September 2026. Traverse the sealed 14-September rank, skip candidates above £100, and select the first at or below £100; otherwise seal canaryRefused:"all_ranked_candidates_over_cap". Submit one whole-share OPG limit with a floored USD limit and deterministic client ID. A zero-fill reject/cancel/expiry is terminal: no substitute, reprice, or retry. Permit one same-ID replay only after confirmed 404; second ambiguity or field mismatch is H3. Partial fill requires cancel, resnapshot, and reconcile; only final quantity 0 or 1 is valid, otherwise H3 and canonical flatten. No stage-2 progression is authorized until a pre-opening, hash-bound binary canary go/no-go rule defines every terminal outcome; absent that rule, the result is NO-GO.
3. Risk and unwind. STATUS: MUST-FIX BEFORE AUTHORITY · DATE: 3 September 2026. Before POST, rerun H1–H4 from fresh reconciled sleeve state. H1 uses daily reconciled sleeve NAV; H2 uses sleeve equity versus since-arm HWM. A £100 loss is 2% at £5k, 1% at £10k, and 2.618% at the £3,820.56 H2 floor, below H1 3.202155812%; all economics count. Stage 2 is forbidden until the canary is canceled/resolved, sold next open at verified quantity, and reconciled to zero.
4. Prefix-N construction. STATUS: MUST-FIX BEFORE AUTHORITY · DATE: 3 September 2026. Seal capital G, stage s∈{10,25,50} bp, lagged overlay L∈[0,1], 14-September FX F, formation prices P_i, exact 20-session ADV_i, available-capital limit C=max(0,£10k-other approved pre-overlay targets), identities, and broker/FX/SEP/calendar manifests. Bind s and C to exact approval-event hashes and cutoff timestamps. Before ratification, seal the snapshot retry cadence and terminal cutoff; missing, stale, nonpositive, short, duplicate, conflicting, or still-ambiguous inputs refuse. Use decimal strings, arbitrary precision, and floor arithmetic.
5. Exact algorithm. STATUS: MUST-FIX BEFORE OWNER RATIFICATION · DATE: 3 September 2026. Set B=floor(min(G*s/10000,C)*L*F*100) USD cents. For each n=1..D-1, allocate a=floor(B/n) cents and q_i=floor(min(a,floor(0.01*ADV_i*100))/(100*P_i)); deployed=sum(floor(100*P_i*q_i)). Pass only when B>0, every q_i≥1, and deployed/B≥0.90; select the largest passing N, else zero. Preserve residual cash; never skip or reallocate. Canonical pseudocode bytes and their domain-separated hash must bind the algorithm, inputs, event cutoffs, retry cadence, ordering, rounding, cap, refusal, unwind, and LL-4b configuration before the owner sees a decision event.
6. Refusal and consumer semantics. STATUS: MUST-FIX BEFORE AUTHORITY · DATE: 3 September 2026. Prefix selection precedes LL-4b; LL-4b recomputes the same selected set and the v3 receipt binds order. N=0 seals constructionRefused:"prefix_n_zero" and submitEligible:false; it is not a submit-eligible applicable:false flat plan. A nonzero rehearsal must fail; no strategy POST is allowed. Cancel plus canonical resumable next-open flatten/reconcile is the only post-refusal action. Current rehearsal code consumes only v2 eligible plans; v3 compatibility must be implemented and proven before any rehearsal uses this authority.
7. Trial 53 and one shot. STATUS: FIX-FIRST · DATE: before lockbox. Register trial 53 against the current 66-line ledger whose reviewed SHA-256 abbreviates to fe5a1ce0…939c; the full digest must be sealed from the current file before ratification. The prior 65-line / 5aed4adb…3127 binding is stale. Raw and effective DSR N are 53, with absent #43 spent and diagnostics remaining trial:null. The safety-only canary is not a trial. After the authorized opening, evaluate fixed N once. PASS requires 12×mean(monthly top-N net − monthly EW-universe net)>0, DSR≥0.95 at N=53, and PBO≤0.20; otherwise refuse forever. Before sealing, define the six-block remainder allocation, which blocks are IS/OOS in all 15 tests, which side owns the one-rebalance-month purge, and the exact annualized net-active Sharpe formula. Tie goes to full decile; zero variance is −∞ except zero mean/variance=0; no rerun.
8. Event and hash domains. STATUS: MUST-FIX BEFORE OWNER RATIFICATION · DATE: 3 September 2026. Split the pre-opening authority event from future runtime-result fields. Domain-separate and seal schema, rank-proof, canary-lifecycle, construction, policy, executor, and receipt hashes; the pre-opening event cannot claim values that exist only after execution.
9. Implementation boundary. STATUS: BOUNDED · DATE: after fixes and owner authority. The packet may touch canonical execution plan, broker adapter, reconciliation, rehearsal, risk, validation, live-lane CLI, and corresponding tests. Keep strategy/** frozen. Do not modify the daemon: it has no strategy-submit API and already invokes canonical hard action.

#### Q13–Q16 — owner authority

1. Q13 · STATUS: PENDING OWNER · DEADLINE: 3 September 2026. A approves the exact canary/risk lifecycle; B means no canary.
2. Q14 · STATUS: PENDING OWNER · DEADLINE: 3 September 2026. A approves prefix-N/refusal/unwind; B means no concentrated authority.
3. Q15 · STATUS: PENDING OWNER · DEADLINE: 3 September 2026. A approves trial 53, N=53, PBO, and one shot; B neither registers nor evaluates the trial.
4. Q16 · STATUS: PENDING OWNER · DEADLINE: 3 September 2026. A permits only the final hash-pinned policy-v2 successor plus fail-closed v3 executor; B leaves Q13/Q14 dormant and authorizes no order.

STATUS: AUTHORITY BOUNDARY · DATE: 29 August 2026. Prose or an A choice alone authorizes nothing. Before owner ratification, close the corrected ledger/PBO definitions, split the authority/runtime events, seal final domain-separated digests, land reviewed code, and prove a loader-valid policy v2. The pre-opening event seals owner identity and UTC, Q13–Q16, successor-policy/executor hashes, and schema/rank/canary/construction/LL-4b domains; runtime refusal fields, selected N, and submitEligible belong to later receipts.

Policy-v2 fix-first review

STATUS: MUST-FIX BEFORE OWNER RATIFICATION · DATE: 3 September 2026. The draft is unsealed: owner/date/document hashes remain placeholders; the declared thresholds hash 4ae7357c…65dcaf mismatches the actual cbbe1cca…c2ad5a; inherited arm age and HWM are absent from the event; and the loader correctly refuses because the required thresholds.variants structure is missing. Acceptance requires one canonical-v2 parser that validates identity, version, rule, predecessor, document and threshold hashes; calibration/expiry inheritance; nested H1/H2/soft schedule; operational semantics; sizing/gates; production-v1 downgrade refusal; and policy-bound execution recomputation.

STATUS: PASS · DATE: 29 August 2026. Policy-read thresholds are the sole runtime authority: no codeConstantsSha256; producers and replay use LoadedRiskPolicy.

STATUS: OWNER DECISION · DEADLINE: 3 September 2026. Choose A: 0.90/0.90; B: 0.80/0.80 or 0.95/0.95; or C: reject the gate. Separately choose all-name versus fractional/per-name headroom, prefix/canary authority, outlier treatment, and ratification. Per-name/no-headroom-as-zero is a defensible response to the $2.5e15 pathology but is not yet authorized. Preregistration owns construction; policy v2 owns submission thresholds and the construction hash. Policy v2 must not silently widen LL-4b, and replay must recompute metrics rather than trust stored values.

5. Wave 7 run plan

1. STATUS: PREREQUISITE · DATE: before any look. Build one durable read-only DuckDB model from manifest-listed SEP/ACTIONS/TICKERS. Exclude the two orphan partitions; carry open; pin manifest and census hashes.
2. STATUS: PREREQUISITE · DATE: before any look. Re-author durable COMMON.md, SPEC.md, and DATA_DICTIONARY.md from retained prompts and killcard code.
3. STATUS: PREREQUISITE · DATE: before fundamentals. Add SF1/SF2/SF3/EVENTS/SFP/DAILY schemas and content-addressed ingest; re-ETL the already-owned 2.74GB bundle at $0 marginal.
4. STATUS: ORDERED · DATE: Wave 7. Run dGP/A first after GP/A reproduction, overnight-component after the open audit, then iHML after SF1/PIT checks. This ordering is inferred from judge rank and dependencies.
5. STATUS: COMPUTE CAP · DATE: every job. At most two concurrent DuckDB jobs. Hard maximum memorylimit is 4GB; ordinary free looks retain the established 1.5GB and threads=2. Use a job-local temp directory, maxtempdirectorysize ≤3GB, delete it after, and pass a df free-space check before work. The 4GB setting is reserved for the LL-4b diagnostic; it is not permission to inflate every look.
6. STATUS: BUDGET · DATE: every look. Runtime ≤40–45 minutes, no network, no pandas full scan of 45.3m SEP rows, one streamed ETL, window functions partitioned by ticker.
7. STATUS: GOVERNANCE · DATE: before execution. Append the ledger row before running; one declared variant maximum, and zero after a refuter amendment removes it. Same config + data snapshot + code + seed must rerun bit-identically.
8. STATUS: OUTPUT · DATE: every look. Emit executable, machine report, human report, and series CSV. The JSON must bind the ledger/config/data/code hashes; primary, separately costed benchmark, NW lags, recency, 1×/2×/3× costs, controls, twins, drawdown, capacity, diagnostics, runtime, and verdict.

6. implementation packet briefs for Tier A

STATUS: NONE TO ISSUE · DATE: 29 August 2026. The judge ranked zero Tier-A candidates, so the required “one packet per Tier-A look” set is empty. Creating a Tier-B implementation packet here would misstate rank and prematurely authorize a run. When a candidate earns A, its packet must bind:

• bounded file set: exactly one runner under lib/research-screens/ plus one looks-report JSON;
• frozen estate: every other file, all live-lane strategy/policy/ledger bytes, and the post-2021 lockbox;
• acceptance: required report fields, a pre-existing ledger row, one variant maximum, manifest/code/config/seed hashes, and a byte-identical deterministic rerun.

7. Process changes adopted

1. STATUS: ADOPTED · DATE: 29 August 2026. Publish difference-series volatility, MDE, and required sample before every look; no-run when the prior is below power.
2. STATUS: ADOPTED · DATE: 29 August 2026. Compute turnover/event/auction break-even first; require 2× costs and report 3×.
3. STATUS: ADOPTED · DATE: 29 August 2026. Freeze one primary, one separately costed benchmark, and one verdict map.
4. STATUS: ADOPTED · DATE: 29 August 2026. Bind PIT timestamps, lags, revisions, identity joins, and effective-dated classifications; current TICKERS remains diagnostic-only.
5. STATUS: ADOPTED · DATE: 29 August 2026. Use actual delisting proceeds and corporate-action truth; live dividend/split reconciliation must pass before arm.
6. STATUS: ADOPTED · DATE: 29 August 2026. Power, data, execution, cost, and decomposition work is trial:null; charge correlated families together.
7. STATUS: ADOPTED · DATE: 29 August 2026. Bind every date and manifest; forbid post-2021 prechecks and JSONL globs; exclude the two orphans.
8. STATUS: ADOPTED · DATE: 29 August 2026. Measure binding 1%-ADV capacity and whole-share deployment/tracking/tilt before preregistration or live sizing.

8. Killed candidates

1. implementable-n-v1 (position-count and minimum-notional sizing for a £10k sleeve) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics 0.88: the concentration return claim is not supported; preserve only the whole-share fidelity diagnostic.
2. overlay-cadence-v1 (monthly leverage update instead of daily) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal process 0.93 and unanimous >=0.85: an in-sample cadence comparison cannot amend the sealed overlay.
3. rank-band-turnover-v1 (buy top decile, hold to top quintile) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal process 0.88: a small turnover benefit does not justify changing the frozen rule.
4. two-sleeve-netted-v1 (momentum + GP/A run as one book with cross-strategy trade netting and a joint vol target) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes 0.90-0.94: repeats the settled GP/A portfolio decision and lacks live attribution.
5. risk-budget-recalibration-v1 (the H2 hard-kill threshold is tighter than the strategy's own history) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes 0.93-0.95: recalibrating a gate to its observed breach is post-hoc.
6. opg-vs-moc-execution-audit — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes 0.85-0.90: the proposed audit cannot authorize a changed fill; diagnostic only.
7. date-staggered-momentum-timing-luck — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics 0.85: likely lowers return and cannot replace the registered rule.
8. pooled-calendar-exposure-overlay — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process 0.93/0.94 and unanimous >=0.85: mechanical underexposure plus unpowered event gates.
9. quarter-end-window-dressing-reversal — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal process 0.89: 20 recency events and 24 December events are no evidence; costs consume the prior.
10. opening-auction-participation-profile-INTRADAY — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal knowability/process 0.99/0.98 and unanimous >=0.85: no lawful pre-2022 panel or current executor path.
11. qmj-lite-composite (does a 4-pillar quality composite beat the best single component?) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process 0.87/0.88: underpowered composite likely dilutes GP/A.
12. junk-screen overlay on the registered momentum book (mom x negative-signal exclusion) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process 0.90/0.93 and unanimous >=0.85: dead-signal recombination with frozen-survivor overlap.
13. forward-looking crash overlay: VIX term structure vs the registered realised-vol overlay — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes 0.95-0.97: retrospective data, exposure mismatch and excessive flip costs.
14. pre-committed equal-weight quality composite (GP/A + COP + low accruals + low net issuance + low asset growth) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process 0.96/0.95 and unanimous >=0.85: post-hoc weak-pillar assembly.
15. cross-asset trend sleeve on the survivorship-free ETF universe (Sharadar SFP) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal process 0.88: missing SFP audit and non-unique K rule; only the data-integrity diagnostic survives.
16. insider cluster tilt on true EDGAR acceptance timestamps — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes 0.94-0.97: optimistic fundable evidence is weak and this is the fourth family attempt.
17. process reset: power gating, benchmark pre-commitment, lockbox re-power, and a zero-capital shadow forward book over all 52 looks — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal knowability/process 0.91/0.96 and unanimous >=0.85: retroactive gate edits and unreproducible all-52 scope; prospective pieces become DIAG.
18. 52wh-fresh-anchor (52-week high with anchor recency, split-only prices) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Two fatal votes >=0.8: a dead price-signal family with adverse turnover and low power.
19. industry-momentum (Moskowitz-Grinblatt, Fama-48 industries) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal knowability/process and unanimous >=0.85: no PIT industry classification.
20. price-delay (limited investor recognition, annual cohorts) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics 0.82: projected effect cannot change a decision.
21. momentum-bear-state-exposure (Daniel-Moskowitz crash hedging, frozen-constant version) — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/knowability and unanimous >=0.85: the gate removes the rebound payer.
22. xs-rank-ridge-price-only-v1 — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes and unanimous >=0.85: tuned duplication without a credible incremental endpoint.
23. mispricing-composite-syy-longonly-v1 — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal knowability 0.90: non-PIT inputs/date semantics invalidate the claim.
24. gbrt-xs-rank-gkx-v1 — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal knowability 0.87: no genuine vintage feature spine.
25. cost-aware-banded-rebalance-v1 — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process >=0.8: claimed uplift exceeds the measurable cost ceiling; diagnostic only.
26. dividend-initiation-drift-exdate — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics 0.87: event supply is too small for a fundable modern test.
27. corporate-action-universe-hygiene-overlay — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics 0.87: hygiene is necessary infrastructure, not a return edge.
28. acquirer-post-completion-exclusion — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Two fatal votes 0.88: only 3.15% flagged weight, so benefit is economically capped.
29. split-liquidity-expansion-drift — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes and unanimous >=0.85: 29 modern qualifiers, only 17 clean.
30. insider-cluster-edgar-acceptance-time — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes >=0.8: optimistic fundable t=1.137 and the family is spent.
31. bond-credit-tsmom-long-or-cash — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal process 0.82: return claim is pre-known weak and data/process blockers dominate.
32. etf-xsec-momentum-12-1-pit — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process >=0.8: likely survivor duplicate and not robust to 3x costs.
33. fx-carry-etf-wrapper-venue-resolution — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Three fatal votes and unanimous >=0.85: no qualifying live wrapper.
34. vix-term-structure-exposure-head-to-head — STATUS: KILLED · DATE: 29 August 2026. Positive case was tested or specified; decisive falsifier: Fatal economics/process and unanimous >=0.85: flip costs exceed expected benefit and endpoint is confounded.

9. Data purchases and what they unlock

10. Confirm-card addition — worst 10 drawdown episodes, narrated

STATUS: REQUIRED TEMPLATE · DATE: next confirm-card. Add a section after the primary underwater plot:

1. Extract the ten deepest distinct peak-to-trough episodes from the primary active-net compounded equity curve. Nested subepisodes belong to the parent and do not fill extra slots.
2. For each episode report peak date/equity, trough date/equity, depth, trading-session duration, recovery date and duration, or “unrecovered at screening end.”
3. Narrate the registered mechanism first: market regime, signal exposure, overlay/leverage, and why the loss is or is not consistent with the payer.
4. Then report falsifiers: largest name/sector/cohort contributors where the data supports them; turnover/cost spike; corporate-action, missing-entry, delisting, stale/open, or PIT flags; benchmark contribution; whether one event dominates.
5. Map the path to precommitted risk gates: H1/H2/soft-review result at the then-valid policy threshold. Do not retrofit thresholds.
6. State what is unavailable. Missing attribution is labelled, never inferred.
7. End each episode with one verdict: mechanism-consistent, data/execution defect, unexplained, or would have killed. The section is narrative evidence, not an eleventh endpoint.

Evidence boundary

• STATUS: SOURCE-BOUND · DATE: 29 August 2026. Candidate scores and tiers come from rank.json/rank.md; strategist priors are superseded where the judge amendments conflict.
• STATUS: LOCKBOX SEALED · DATE: 29 August 2026. No statistic here uses 2022+ returns.
• STATUS: NO AUTHORITY · DATE: 29 August 2026. This document authorizes no run, purchase, capital, broker action, policy amendment, or lockbox opening.

Critic round 1 — applied

• STATUS: APPLIED · DATE: 29 August 2026. Downgraded all three Tier-B outlines from pre-registration-ready, added the six conjunctive lockbox gates, and made each missing power/timing/benchmark/control precommitment explicit.
• STATUS: APPLIED · DATE: 29 August 2026. Added item-level STATUS/DATE to all eighteen DIAG rows and corrected iHML's judge point to 1.0% active.
• STATUS: APPLIED · DATE: 29 August 2026. Replaced the obsolete LL-4b sizing section with Round-4 v3/kernel/headroom facts, the exact ladder arithmetic, A/B/C/D construction options, and policy-v2 A/B/C owner authority.

Round 3 — applied

• STATUS: APPLIED · DATE: 29 August 2026. Replaced the provisional ladder amendment with v1.2-r2’s rank proof, bounded canary, exact prefix-N algorithm, prefix_n_zero refusal, trial 53/PBO contract, and Q13–Q16 authority boundary.
• STATUS: APPLIED · DATE: 29 August 2026. Added the two report-only survivor diagnostics, corrected arithmetic-mean-versus-CAGR labels, reproduced policy-v1 calibration, and narrated the ten distinct worst drawdowns without claiming causal attribution.
• STATUS: MUST-FIX · DATE: before owner ratification. Applied both independent amendment reviews and the policy review: exact ranks 1..D, canary go/no-go closure, approval-event/cutoff/retry bindings, canonical pseudocode bytes, v3 rehearsal compatibility, the corrected 66-line ledger pin, complete PBO definitions, split event phases, hash domains, placeholders, inherited state, downgrade refusal, and recomputation. The optional OPG-vs-MOC audit was absent, so no execution numbers were invented.