The studies, with the files

Strongest pre-registration in the bundle: PREREG-2026-09-16.md is byte-identical to the source and its sha256 4e921928ecd174744cb1d2970fce62d1c0305015e9308a8c4d2f3aa3242cfe32 matches the recorded .sha256 file ('frozen 2026-09-16 17:48:13 EDT'); prereg created 17:48:11, engine.mjs 17:50:33, first output 17:50:48. Arms = 26 tape cells (13 setups x 2 roots) + the two prop hypotheses H-P1/H-P2. Not counted as arms: 26 seeded coin-flip control streams, and the post-hoc size addendum (520 cells; 4 positive in both halves at list fees, 3 of them coin-flip controls). The site's strategyClaims.ts numbers (all 26 cells' net/t/n/verdict, overlay table, size rows, 520/4/3) match out/*.json exactly. CATALOG.md/SOURCES.md are the public claims being tested (third-party URLs). Run: node --max-old-space-size=12288 engine.mjs; node sim.mjs; SIZES=1,3,5,10,20 node sim.mjs. Trade-level outputs (out/trades-NQ.json, out/trades-ES.json) are not included.

Backtest of claimed 'profitable prop trader' strategies (13 setups x NQ/ES micros) plus a prop-account simulator

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ, ES 2018-2026); CPI/NFP/FOMC release timestamps from the fp-context event archive (<data-root>/fp-context: events/**/*.jsonl with ts_event/ts_published/subclass/region, events/earnings/_weights.json, features/context_daily.jsonl); built from FMP, Fed, BLS/BEA, Federal Register, GDELT, Alpaca and SEC EDGAR feeds; not redistributed and its builder is not included

Post hoc diagnosis, no pre-registration. Arms is a lower bound: 26 cells reported in the write-up (5 breakout variants, 4 target multiples, 1 breakeven-stop variant, 8 late-session-momentum cells, 8 overnight/day cells) plus about 27 conditioner splits printed by study2.mjs and not tabulated. The breakout replication covers 2026 YTD only; the multi-year evidence the write-up cites is the separate pn-001 study (orb-vxn-and-gex-regime) and an external paper. Three passages quoting the desk's live pick record were redacted under the 2026-09-10 disclosure rule. Scripts write a debug file to <out-dir>.

Why the tape-graded record loses: replication of the published level-breakout family, plus late-session momentum and overnight drift checks

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (2020-2026; the breakout scripts read 2026.csv only); QQQ dealer-gamma history <data-root>/fp-data/gex/history/*.json.gz from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE)

Both preregs were created 22-31 s before the scripts and about 4 minutes before RESULTS; no in-text times. Arms: pn-001 24 cells (2 symbols x 2 OR windows x 3 VXN terciles x 2 exits); pn-070 8 cells (2 symbols x 2 regimes x 2 tests). pn-070 H1 (breakouts on negative-GEX days) is a gamma-regime rule that was left 'PROMISING, accruing' at n=3/8 with a rerun due at about 60 classified days (late Oct 2026); it was never closed.

pn-001 opening-range breakout by VXN tercile (NQ, ES 2010-2026) and pn-070/071 GEX-regime fade vs breakout (2026 forward archive)

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ, ES 2010-2026); FRED VXNCLS.csv (public); QQQ/SPY GEX snapshots <data-root>/fp-data/gex/2026/ from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE) (pn-070 only)

320 cells were pre-counted by the study (8 instruments x 5 RVOL buckets x 2 entries x 4 frictions); 1 cell was confirmatory. Unconditional ORB (all trades) is also reported: +0.744 pt gross, -0.256 pt net (t -0.21). Timestamp defect: the prereg says 'Frozen 19:30 MDT' and its amendment says '19:45 MDT', but the file was created 21:26:52 EDT (19:26 MDT) and last written 21:27:37 EDT, before rvol_orb.py (21:28:24) and the first raw output (21:31:33). The file order supports pre-registration; the in-document times are wrong. The results honestly record that monotonicity criterion 2 was written too loosely and a mid-study bug fix (NQ results bit-identical).

Does relative-volume selection rescue the opening-range breakout on futures?

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (8 roots, 2010-2026)

Included because Parts C and D of this prereg define the gamma-wall rules tested in gamma-wall-backfill, and crown_policies.py/open_factors.py are code dependencies of every gamma study. Part A (7 policies) prereg created 11:57:36, before crown_policies.py (11:58:29) and its results (11:58:53). Partial because Parts B-D were appended to the same file later (final write 12:34:38) and the Part B grid results were written at 12:14:16, so the file system cannot show Part B's text predates its results. Arms = 7 (Part A) + 189 (Part B grid: 7 policies x 27 geometries). P6 is a 15-minute opening-range-break control (-0.003 R/trigger, n=3,497). One sentence quoting the live crown-play record was redacted under the disclosure rule. open_factors.py also runs the 2026-09-08 open-factor study, which is not part of this bundle.

Mechanical crown-play policies on NQ (Part A), exit-geometry grid (Part B), and the definitions of the gamma-wall policies (Parts C, D)

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ 2011-2026; open_factors.py builds the data/NQ_1m_et.parquet cache that crown_policies.py requires); FRED VXN

The original prereg file (created 12:55:44) predates the data pull (12:56:04 to 12:57:38). Partial because three disclosed changes followed partial output: Amendment 2 (proxy spot fix, written after descriptive policy figures had printed), a walk correction that removed a look-ahead (which had produced 62-64% hit rates), and Part D2 (added after Part D was refuted). Timestamp defect: those sections are stamped 13:05, 13:12 and 13:20, but the prereg file's last write is 13:03:53 and the results file (with D2 results) was created at 12:57:48 and last written 13:06:46. Arms = P8, P9, D, D2 (P10 is descriptive). One sentence giving the live shadow-ledger tally was redacted under the disclosure rule.

Gamma-wall history backfill: wall-fade and wall-break policies on 730 days of reconstructed dealer positioning

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ); QQQ/NDX end-of-day strike gamma exposure and QQQ daily OHLC per date from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE) (pulled with backfill_history.mjs and backfill_ndx_close.mjs into <data-root>/fp-data/gex/history); live 09:10 QQQ snapshots <data-root>/fp-data/gex/2026/QQQ_*T0910.json.gz for the proxy validation (forward archive, not included)

Near miss recorded honestly: AUC 0.658 and top tercile 82.0% pass, bottom tercile 56.4% fails the 55% bar; width alone gives AUC 0.660. Arms = 1 classifier + 2 conditional rules (C1, C2) that were only to be evaluated if the classifier validated. Timestamp defect: the prereg says 'frozen 2026-09-09 14:20', but the file was created 13:28:11 (same second as cage_classifier.py) and the results file 18 s later, headed 13:28. The file order supports pre-registration; the stated freeze time is wrong.

Cage-day classifier: can pre-09:30 factors predict a close inside the gamma walls, and does that make the walls tradable?

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ); <data-root>/fp-data/gex/history from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE); FRED VXNCLS.csv

The prereg file (13:33:24) predates the first flow data file (13:33:31), consistent with 'frozen before the bulk pull'. Arms = FL1-FL3 (8 descriptive factor/target rows carry no verdict). Timestamp defect: the prereg says 'frozen 2026-09-09 14:45'; the file system and the results header (13:35) show it was written about 70 minutes earlier than stated.

Options-flow direction factors: does net options premium carry NQ direction?

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ); per-minute QQQ net-premium ticks and market tide per date from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE) (backfill_flow.mjs writes <data-root>/fp-data/flow/*.json.gz)

All three policies (M1 wall drift, M2 tilt shift, M3 flip cross) were positive in both halves but none cleared one-sided p < 0.0167, so all failed the frozen rule. Timestamp defect: the prereg says 'frozen 2026-09-09 13:55', but it was created 13:10:11 (same second as wall_migration.py) and the results 8 s later, headed 13:10.

Gamma-wall migration at the 10:10 book refresh as an NQ entry signal

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ); <data-root>/fp-data/gex/history from Unusual Whales API data (requires your own subscription; the pull scripts here read the key from an env file named by UW_KEY_FILE)

Per claim: day session = pre-registered family D1 (8 roots, none pays; 5 of 8 negative); gap continuation = family E (2 arms, both refuted); turn-of-month = family C1 (refuted; flips sign between halves); 'Release days are wilder' ($2,880 vs $2,776 mean absolute NQ move, 2020+) is NOT a pre-registered test: the results label that table 'Computed after the fact and labelled descriptive — no test attached'. The NQ hour profile was pre-declared as descriptive. Partial because: 2 arms beyond the 27 frozen (declared); the prereg file was edited after the results (an 'ANSWERED' banner; no hash of the original body exists); and the release-day range figures are post hoc. Timestamp defect: the prereg says 'Frozen 2026-09-16 10:15 ET', but it was created 09:58:43 and the results JSON was written 10:02:33, so the stated freeze time is after the results.

What the desk can actually publish: a 29-arm strategy suite (overnight, timing, calendar, day session, gap follow-through)

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (8 roots, 2010-2026); the fp-context event archive (<data-root>/fp-context: events/**/*.jsonl with ts_event/ts_published/subclass/region, events/earnings/_weights.json, features/context_daily.jsonl); built from FMP, Fed, BLS/BEA, Federal Register, GDELT, Alpaca and SEC EDGAR feeds; not redistributed and its builder is not included

Arms = 19 per root x 3 roots (15 hypotheses Bonferroni-counted on NQ, p < 0.0033). Attention (B1): the narrower attention-arm prereg (created 05:53:05, about 45 min before the run) and B1 here were both written before results; B1 = +$281.58/night, t 1.57, p 0.117, refuted. CAUTION on the placebo: with attention dates shuffled, the same arm returns +$100.11 (t 0.70), about the $98.00 baseline. Shuffling therefore removes the effect, it does not reproduce it; both write-ups describe this backwards. The refutation stands on p = 0.117 and random-night placebo p = 0.188. Pre-FOMC (N1): partial. The prereg itself quotes the answer (+$973 net, n=53, t=3.12) from the event-nights run hours earlier, and N1 re-measures those same 53 nights, which lie inside the 2016-2026 window where the effect was first found (macro-events-fundamentals). It is a re-measurement, not an independent confirmation. 'On a longer sample' in the results is not true for this arm. Timestamp defect: the prereg says 'Frozen 2026-09-16 09:45 ET', but the file was created 06:37:45 and the results 06:38:58.

The overnight drift under three layers (technical, event, behavioural), including the attention arm B1 and the pre-FOMC arm N1

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ, ES, GC 2016-2026); VIX daily <data-root>/fp-data/fmp/vix-daily.json (FMP); the fp-context event archive (<data-root>/fp-context: events/**/*.jsonl with ts_event/ts_published/subclass/region, events/earnings/_weights.json, features/context_daily.jsonl); built from FMP, Fed, BLS/BEA, Federal Register, GDELT, Alpaca and SEC EDGAR feeds; not redistributed and its builder is not included; Wikipedia pageview attention series (fp-context attention/wikipedia, public API)

Prereg created 22:42:14, runner 22:54:20, first results 23:26:19 (re-run 05:32 after the news wire was filled; verdicts unchanged). Arms = 7 (NQ) + 7 (ES) + 6 (GC) primary cells. Pre-FOMC was pre-registered as a pipeline POSITIVE CONTROL for an effect already known from the 2026-08-25 study, not as a new hypothesis.

Does the overnight drift concentrate in identifiable event nights beyond pre-FOMC?

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ, ES, GC 2020-2026); the fp-context event archive (<data-root>/fp-context: events/**/*.jsonl with ts_event/ts_published/subclass/region, events/earnings/_weights.json, features/context_daily.jsonl); built from FMP, Fed, BLS/BEA, Federal Register, GDELT, Alpaca and SEC EDGAR feeds; not redistributed and its builder is not included

The pre-FOMC drift was a pre-specified SECONDARY test ('reported, never confirmatory'), measured 18:00 ET prior day to 09:30 ET open: +25.38 bps gross, t +3.61 (corrected), one cell among about 176 in this study and about 544 pre-counted across that day's three studies (the study notes the expected max |t| under the null is about 3.3). Prereg created 00:36:30 EDT, before fundamentals_nq.py (00:37:36) and the first output (00:37:41); its in-text 'Frozen 20:25 MDT' (22:25 EDT) is two hours earlier than the file and cannot be verified. Results carry two dated corrections (2026-08-26: calendar repair; jobless-claims result refuted). Arms = about 176 cells as pre-counted by the study.

Do macro events, surprises and administrations predict NQ? (where the pre-FOMC drift was first found)

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ 2010-2026); FMP economic calendar 2016-2026 JSON (<data-root>/fp-data/fmp/econ-calendar-*.json; FMP subscription)

This is the only genuinely out-of-sample test of the pre-FOMC rule, and it has 1 of the 8 forward observations its own frozen bar requires. The rule graded here enters at 18:00 ET (after the daily halt) and exits 09:30 ET. The forward track file (prefomc-shadow-track.jsonl, per-date entry/exit prices) is not included. The grader (prefomc_shadow.py, created 07:12:11) predates the spec (07:20:57); the spec froze the promotion bar before any forward observation existed.

Pre-FOMC overnight drift: forward shadow-grading spec and grader

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ); FMP economic calendar (FOMC decision dates, 'Interest Rate Decision', country US)

Post hoc: about 25 hypothesis variants screened (18 in close_study.py), then the survivor put through robustness checks; no pre-registration exists. Survivor: |15:00 to 15:30 move| > 0.30%, follow it to the close: +5.79 bps/trade, t 2.17, bootstrap p 0.016, hit rate 51.5% (not significant). That t is below the t >= 3 bar the later pre-registered studies use. 2023-2026: the base rule is negative on all four index futures (NQ -2.3, ES -4.0, YM -0.9, RTY -0.6 bps, all n.s.); NQ longs +2.5 pt, shorts -8.9 pt. close_deep_dive.py needs pm_paths.parquet, which no script on disk generates. The calendar-study (OPEX, month-end, quarter-end) code was not found.

Close Read feasibility study: the 15:30 to 16:00 ET NQ window

Data needed to rerun: CME futures 1-minute OHLCV bars (Databento GLBX layout: ts_event,rtype,publisher_id,instrument_id,open,high,low,close,volume,symbol) at <data-root>/fp-data/glbx/<ROOT>/1m/<YEAR>.csv; licensed vendor data, not redistributed (NQ, ES, YM, RTY 2010-2026)