Completed Round A: 3,948 configs evaluated; v1/v2 and widget preserved.

Nested selection returned −0.03896% equity/day. No validated edge claimed. Final data remains sealed; prior BTC/SOL contamination is documented.

Validation: 53 tests and 72 replay checks passed.

Deliverables: Exploration report · Full leaderboard

Review complete. Both deliverables are written and verified against the tree, and no project source was changed.

Outcome. Round A's headline is sound: all ten frozen finalists reproduce exactly from the committed engine, a future-perturbation probe found no lookahead in features, prequential logistic or execution, and the 28 v3 unit tests pass. The defects are in the final-check design, which matters because lab.final_v3 can only be run once.

Findings, ranked (7 total in .imd-findings.json, full write-up in artifacts/review_v3a.md):

• High: 2024 ETH segment cannot warm the 4h EMA200. 44,640 minutes is 186 four-hour bars, fewer than the 200 needed. The higher-timeframe state is zero throughout, so trend and session_trend emit nothing and regime_switch loses its trend leg. The one-shot check is silently void for 3 of 10 finalists.
• High: frozen logistic model leaks into the BTC/SOL transfer check. It is trained on ETH labels over exactly the same calendar window that BTC and SOL are scored on, with hour-of-day and trend features. Those checks for the logistic and meta_fade finalists are in-sample in time.
• Medium: maker fills credited on a 0.045 bp trade-through. Requiring 1.3 bp through cuts the logistic maker finalist from 0.099 to 0.039 % equity/day. The maker premium is non-monotone across tick sizes and smaller than the taker variant's result on more trades.
• Medium: the post-v2 ETH check is empty by construction. 110 sealed minutes can never hold a 48-hour trade, and the seal refuses refresh.
• Low: "positive outer folds" column is in-sample. Top-1 return has 55 percent in three trades plus six 50x liquidations, and the gap pre-check reads future bars.
• Info: breach sigma, momentum horizon and the logistic label horizon differ from the spec and report text.

Where edge is most likely. The 60m-signal, 48h-hold fade cluster is one hypothesis sharing 40 trades across three families, with gross well above modeled costs. Pre-register a single fade rule at 20x with taker execution and test it once, pooled across BTC, SOL and a properly warmed 2024 segment, to reach 100 trades. Drop the maker premium and the 50x result from any claim.

Addressed all seven findings in .imd-responses.json.

• Final scoring now rejects invalid warm-up, ML-calendar and short-window checks.
• Added 12 maker-fill sensitivities and corrected reporting.
• 60 tests and 72 configuration replays pass.
• v1/v2 and widget unchanged; final seals remain unopened.

Regenerated both required artifacts. No validated trading edge is claimed.

All seven findings from my first review are settled, and one residual defect of the same kind as the first high finding remains, rated medium. The review is in artifacts/review_v3a.md and the finding is in .imd-findings.json. No tracked file was changed, no sealed final data was opened, and the final-opening lock was not created.

Earlier findings

Finding	Status	How I checked
[high] 2024 segment cannot warm the 4h EMA200	Fixed	With 186 4h bars, the three trend-dependent finalists return not_scored; 200 bars are rejected and 201 accepted.
[high] Frozen logistic trained over the BTC/SOL calendar	Fixed	logistic and meta_fade are refused on the 2024 segment and on BTC/SOL; frozen code, seal and grid hashes match the tree.
[medium] Maker fills on a 0.01 trade-through	Fixed (disclosed)	I recomputed all 12 sensitivity rows and they match the stored ones exactly; the trial counts reconcile (3,960 v3 trials, 4,872 White vectors, 5,880 DSR arms).
[medium] 110-minute post-v2 seal	Fixed (declared void)	All ten finalists return not_scored there; the task's check (a) stays unmet, and the report now says so.
[low] In-sample "positive outer folds"	Fixed	Fields are renamed and labelled as in-sample.
[low] Top-1 concentration	Fixed (disclosed)	My replay gives +13.79% total, 55% of it in three trades, six liquidations — same as the report.
[info] Feature definitions	Fixed (documented)	The report and summary.json now state the actual definitions.

Remaining defect (medium)

The flow_follow finalist (6656a86c82bd55ce) is still scored as a silent zero on the 2024 ETH segment and on BTC and SOL. lab/final_v3.py:84 gives it no taker-flow input there, so it can emit no signals, yet the eligibility guard marks it eligible.

• Reproduction: on development bars it emits 125 events with flow and 0 without. Running the one-shot driver on synthetic bars returns eligible, n=0, 0.0%/day for it on all three checks.
• Why it matters: it contradicts the author's "never a silent zero" answer, and the unrepeatable 2024 opening would be spent on it for nothing.
• Fix within the design: add a flow guard to the eligibility check, or fetch and seal Binance 5m klines for January 2024 before opening.

Top-10 re-run

All ten finalists replay bit-exactly against the stored checkpoints, and the 49 Python tests pass. The "no validated edge" verdict is supported: the nested next-fold policy is −0.039% equity/day on 52 trades, and White p is 0.92 against FLAT.

• Final checks have little power left: only four finalists get a genuinely untouched check (the 31-day 2024 segment), and none can reach 100 trades there. The two ML finalists have no final check at all.
• Concentration: finalists 5–10 are negative without their three best trades.
• Duplicate bet: meta_fade and sigma_fade daily returns correlate 0.98.
• Maker adverse selection: for four of the five maker finalists, missed fills had higher gross than filled ones.
• Leverage: returns scale almost linearly with it; 50x adds 1–22 liquidations per finalist and 20x has none.

The review file ends with suggestions on where edge is most likely; the main one is that the 1h low-width fade cluster with 48h trailing exits is short of statistical power rather than beaten by costs, so the next step is more untouched history, not another search.

Fixed the reproduced missing-flow finding. Flow-dependent checks now return not_scored with null metrics when matching flow coverage is missing, stale or partial.

All 62 Python tests pass. Frozen integrity checks pass; no production final data was opened.

Created the three required outputs. Existing research, models and widget remain unchanged. The broader Round B search and live-module work remain incomplete, explicitly disclosed in the reports.

Grade: no edge. And the final check was never opened, so there are no final-check results to confirm. The project is still at "Round A complete; FINAL DATA STILL SEALED" (configs/frozen-v3.json:227). There is no lab/runs_v3/final_opening.json or final_results.json, and both the seal and the summary say finals_opened: false. Every headline number is from development.

I wrote the review to artifacts/review_v3b.md and the four findings to .imd-findings.json. No tracked files changed.

What I checked (on a clean clone of 869654f; the machine had no Python, so I installed 3.12 in a temp folder):

• Frozen state: the freeze check passes — code, grid and seal hashes all match — so the one-time opening can still be done cleanly.
• Tests: tests.test_final_v3 and tests.test_v3 pass, but they only use made-up seal data.
• Deflated Sharpe: I recomputed it by hand and it matches. The strongest development candidate, 8c6176fab4c83b51, has p = 0.020 after counting all 5,880 configs from v1, v2 and v3.
• White reality check against flat: p = 0.92; the best result, 0.189 %/day, is below the 0.426 %/day the bootstrap reaches 5 % of the time by chance. Counting the 1,008 configs with no saved daily returns, the overall p is capped at 1.0.
• Nested walk-forward: −0.039 %/day over 52 trades.

Findings, most severe first:

1. High: the final round never ran. Round B and the one-time final opening were not done, so the rules on final checks, at least 100 final trades and the final results distribution are unmet.
2. High: the untouched final checks can't support any edge claim even once opened. The 2024 check (lab/v3data.py:66) is only January 2024: 186 four-hour bars, below the 201 needed to warm up the long moving average.
   • Of the 10 finalists, 6 are excluded there: the three trend-type families (warm-up too short), the two ML families (model from a later period) and the flow family (no 2024 flow data).
   • The 4 that remain would make about 7–15 trades at their development rates, far short of 100.
   • The newest-ETH check is only 110 minutes, shorter than the 2-day maximum hold, so it can't score any finalist.
   • This can still be fixed by sealing a longer untouched period before the opening.
3. Medium: the BTC/SOL check doesn't use their full histories (lab/v3data.py:78). It reuses the June–November 2026 data from v1 and v2, the same dates as development. As a result the ML families, including the strongest development result, can't be tested on any final check.
4. Info: the deflated Sharpe, the reality check and the nested walk-forward code look correct, and the project's own descriptions of their limits are accurate.

I did not run the final opening (python -m lab.final_v3): it permanently uses up the sealed data, and when to do that is your one-time decision. I also didn't re-run the full 3,888-config search for lack of time; I relied on the hash checks and the project's stored check, which replays 72 sampled configs.

Repaired the upload failure: changed source is now 3.03 MB, below 8 MiB. All three caches are preserved byte-for-byte in the separately delivered results JSON, with a documented offline restore command.

Frozen hashes and final results are unchanged; finals were not reopened. All 71 Python tests and 182 TypeScript tests pass. The verdict remains no validated trading edge.

All three issues I raised last round are fixed, I found nothing new, and I grade the claim no edge. The review is in artifacts/review_v3b.md, and .imd-findings.json is an empty list. The repository itself is unchanged.

What I re-ran: I made a fresh copy of HEAD b3fac73 in a scratch directory and restored the three sealed 2024 data files, which match their sealed hashes. In that copy only, I removed the one-time lock and ran the final opening again. The output matches the committed final_results.json exactly, for all 5 finalists, 4 checks and 3 leverages. The tests and scripts/check-round-b.py pass, and the hashes match completion.json.

Order of events: the data was sealed at 10:25:56, the finalists were frozen at 10:27:23, and the finals were opened at 10:27:26. There is one lock, and the code refuses to open a second time.

Earlier findings:

• Round B and final opening never run (high): fixed. Round B added 1,296 configurations, five finalists were frozen, and the opening lock and per-check results now exist.
• 2024 check only one month long (high): fixed. It now covers all of 2024 plus 42 days of warm-up, with no missing minutes. All five finalists can be scored on it. The best rule makes 171–177 trades; rsi_fade makes only 76–82 and is correctly flagged as under 100.
• BTC/SOL reused the burned 2026 data (medium): fixed. Both now use the same untouched 2024 year. The authors say plainly that this is not the symbols' full history, which the task allows as a substitute era.
• Multiple-testing note (info): withdrawn. It reported no defect and its line reference didn't resolve.

Headline numbers: they all match summary_v3.md. For the best frozen rule (regime_switch, 60m), net equity per day is:

Check	10x	50x
ETH 2024	−0.005%	+0.004%
BTC 2024	−0.012%	−0.052%
SOL 2024	+0.017%	+0.050%

• Every 95% interval at 50x includes zero or lies entirely below it.
• No finalist has the same sign on all three symbols.
• The post-v2 ETH window is 175 minutes, too short to score anything.

Multiple testing:

• Deflated Sharpe: I recomputed the null threshold by hand over 7,176 trials and got 0.4506, which matches the stored value. The deflated-Sharpe probability is 0.018 (0.023 in my simpler approximation), far below the 0.95 needed.
• White test against not trading: p = 0.929, and conservatively 1.0 because 1,008 earlier-round result series are missing.

The authors' own conclusion, no validated edge, matches the evidence and is not overstated.

The supplied flow finding was already fixed: missing input returns not_scored with null metrics. Runtime code remains unchanged.

Added integration regression coverage; 72 tests pass. Regenerated all three v3 outputs and added the review disposition atop the summary. Frozen evidence is unchanged; no final data was reopened.

Details are in REVIEW-ANSWERS.md. artifacts/review_v3b.md was absent, so validation covered the supplied finding.

Implemented the frozen v3 finalist in the compact widget and replay, preserving v1/v2. Added evidence views, downloads, refreshed screenshots, and complete static export.

Validation passed: production build, typecheck, 251 TypeScript tests, 72 Python tests, and 34 browser checks. Submission is 6.93 MB, below 8 MiB.

The evidence supports no validated trading edge. WebKit remains unverified due to missing system libraries.

Setup and limitations are documented in README.md and validation.

Area	Count / result
Brief requirements	24 reviewed: 19 met, 3 missing, 0 faked, 2 need requester input
Broken features	1 high
Product and action on first screen	Yes, in supplied 390px screenshot
taste@1	Fresh run unavailable; supplied result: 14/16
Copy findings	0
Share card	Present

Recorded one finding in .imd-findings.json:

High — bug: expanding the source hashes breaks the phone layout. At 390px and 320px, open Research → “Data coverage, freeze and reproducibility.” Each unwrapped hash measures 422.4px in the shipped font, exceeding its 340px row before the label is added. Allow the values to wrap while preserving the complete selectable hashes.

DOM checks passed for both replay engines, stepping, next signal, rewind, research filters and all 16 gallery states. Bundled v3 evidence matches the pinned inputs. Fresh Chromium/WebKit verification was unavailable; the layout finding uses source, rendered DOM and measured font metrics. Application files were unchanged.

Brief requirement	Status and evidence
“Keep the v1/v2 code, scorer, data cache, tests and widget”	Met — preserved modules, caches, tests and v2 engine selector.
“Prove” a cost-beating edge or convincingly establish none	Missing — the saved verdict explicitly establishes neither.
“Experiment a lot, follow the evidence”	Met — 5,256 v3 trials; discovery and final evidence distinguished.
Timeframes, holding periods and multi-timeframe stacks	Met — search configurations and frozen 1h/4h rule.
Conservative maker fills, adverse selection, fill rate and missed returns	Met — venue model and trial diagnostics.
Specified maker/taker fees and taker slippage	Met — venue model and “Costs before confidence.”
Leverage decision, 50×/20×/10× reporting, volatility sizing and liquidation	Met — execution configurations and research leverage controls.
Existing features and public historical inputs with stated coverage	Met — preserved feature modules and data-coverage disclosure.
Permitted rule, regime, ML, meta-label and exit experiments	Met — recorded family/configuration distribution; unused freedoms disclosed.
Nested walk-forward, earlier-fold selection, purging and one-day embargo	Met — nested results and validation implementation; adaptivity disclosed.
Burned v1/v2 holdouts used only for development	Met — research limitations state this explicitly.
Freeze finalists before one-time final opening	Met — matching freeze and opening records.
Final check on every newer ETH bar	Needs the requester — adequate untouched forward history; supplied 175-minute window is unscored.
SOL over its full period	Missing — an untouched 2024 era substitutes for full history.
BTC over its full period	Missing — an untouched 2024 era substitutes for full history.
Fetch and seal an untouched 2024 ETH segment	Met — full-year scoring era plus warm-up in seal evidence.
Every historical trial in DSR and White/SPA versus FLAT	Needs the requester — complete legacy records/vectors; current counts are lower bounds and global White bound is conservatively 1.
At least 100 final trades and stability before claiming an edge	Met — no edge claimed; grade remains U.
Strongest result, full distribution and what killed promising ideas	Met — Research disclosures and complete trial download.
CompositeWidget and replay show the best frozen v3 finalist	Met — frozen regime-switch ID; unsupported rows explicitly OFF.
Desktop/phone budgets and seven shared RSI tiles	Met — fixed widget dimensions and supplied state captures.
Per-TF score, state, signal age, execution, leverage and confidence mark	Met — widget rows and

Fixed the reproduced overflow. Complete hashes now wrap and remain selectable at 320, 390 and 1440 px.

Rebuilt dist/, added regression checks and screenshots, and recorded the finding as fixed in .imd-responses.json.

Build, typecheck, 251 tests and 37 browser checks passed. Probe and platform limitations are documented.

Area	Previous review	This review
Requirements	Counts not supplied	16 met · 2 missing · 2 need requester · 0 faked
Broken features	1 high	0 reproduced
Product on first screen, 390px	Not supplied	Yes
taste@1	Not supplied	12/16 raw
Copy findings	Not supplied	0 new
Share card	Not supplied	5/5 metadata present

The reported hash-overflow bug is fixed. At 320, 390 and 1440px, expanding the disclosure leaves page width equal to viewport width. Both complete 64-character hashes wrap onto three lines and select exactly. Enter/Space toggling and Chromium replay interactions pass. The fix is present in web/style.css:1060.

.imd-findings.json contains {"findings":[]}. No product files changed.

WebKit remains unverified: its renderer aborted before loading the page because it could not create an EGL display. The probe’s four flags concern transition contrast, label heuristics, the preserved monospace font and a loading placeholder; they do not substantiate another defect in this revision.

The original brief mapping below reflects delivered records; this round did not rerun research or reopen previously disclosed limitations.

Explicit requirement	Status	Evidence
“keep the v1/v2 code, scorer, data cache, tests and widget”	Met	Retained lab/, data/, tests/, src/widget/
Prove a cost-beating edge or convincingly establish none	Missing, disclosed	Summary explicitly claims neither
Maker trade-through, adverse-selection check, fill rate and missed-fill returns	Met	[Execution engine][venue], saved results
Taker/maker fees and 1bp taker slippage	Met	[Execution engine][venue]
Results at 50×/20×/10×, sizing and liquidation	Met	[Research][research], saved settings
Historical source coverage stated	Met	[Coverage disclosure][research]
Nested earlier-fold selection, purging and one-day embargo	Met	[Walk-forward implementation][walkforward]
v1/v2 holdouts treated as burned development data	Met	[Research][research]
Final checks frozen before one-time opening	Met	[Opening record][opening]
Every newer ETH bar scored as a final check	Needs requester: longer forward window	[Research][research]: available 175 minutes cannot support 48h
Untouched BTC/SOL checks over full periods	Needs requester: approve substituted eras	[Research][research]: untouched 2024 substitutes disclosed
Untouched 2024 ETH segment fetched and sealed	Met	[Seal][seal], summary
Every historical configuration counted in DSR and White/SPA	Missing, disclosed	[Research][research]: unlogged trials and missing legacy vectors
≥100 final trades and stability before any edge claim	Met	[Research][research]; no validated-edge claim
Strongest result, full distribution and failed ideas reported	Met	[Research][research], trial download
CompositeWidget and replay show the frozen v3 finalist	Met	[Widget][widget], [replay][replay]
Desktop/phone size budget and seven RSI tiles	Met	Browser measurements and retained layout
Per-TF score, state, last signal/age, execution, leverage and grade	Met	[Widget][widget]
Preserve palette/font; closed-bar signals without flicker	Met	Styles; Chromium stability check
Refresh desktop/phone screenshots of all states	Met	[Screenshot collection][screenshots]

The site is an ETH research replay terminal carried by compact dark charts and a monospace signal grid.

[research