Journal

What changed in the lab, and why

Last tick 4:00 pm ET · all systems normal

Subscribe

The written record: what changed day by day, and what each week actually was.

Two records, neither of them edited after the fact. The journal is written whenever something changes; the weekly review is written every Sunday, out of the ledger, whether the week went well or not.

Every change to a strategy, a rule or a claim on this site gets an entry here on the day it happened. Entries are never edited after the fact — a correction is a new entry that says what it corrects.

The lab was rebuilt on 16 September 2026 and every strategy re-seeded. Entries dated before then are the research record of the system this one replaced — they are kept because deleting a predecessor's losing work is how a track record becomes a story, and they describe strategies and accounts that are not on this site.

  1. item 30 (xsec momentum, SHORT the bottom decile) DEAD; nursery still full

    7-day lab_events review found nothing more urgent. No retire/promote row in the window; qty_drift mismatch cleared to 0 in the last two nightly audits, same self-resolving shape. Worked item 30 (top open): is the bottom 12-1 decile a good SHORT, independent of item 8's dead long-side mean-reversion verdict on the same names?

  2. item 29 (PEAD gap-size cap, pinned-99) DEAD; nursery still full

    7-day lab_events review found nothing more urgent. No retire/promote row in the window; lab_strategies confirmed directly at 16 SHADOW / 1 FUNDED / 3 RETIRED, unchanged since item 28's own check. Worked backlog item 29 (next in rank order): does item 19's moderate-gap (3-10%) cap, DEAD on the wide universe, fare better on the stronger pinned-99 base rate?

  3. item 28 (sector-capped PEAD, pinned-99) built, READY; nursery still full

    7-day lab_events review found nothing more urgent. No retire/promote row in the window (first check used the wrong action names, [secret] against the real columns, still 0). Only unusual event: an automatic test_cells_halved 2026-11-06 (test equity 12% below high-water mark) -- [path]'s own risk control, already self-correcting, not a research question. Worked backlog item 28, next in rank order.

  4. item 27 (SID Method long-only) recovered from FIVE consecutive lost sessions, verified, finally committed

    7-day lab_events review found something more urgent than a fresh backlog item: a process failure, not a trading one. git branch -r showed orphaned research/wip-* branches for 2026-11-04, -05, -06, -07 AND -08 — five sessions running, all independently rebuilding backlog item 27 (SID Method, long-only) and all five parked by [internal file] before committing. One (11-07) got as far as a real lab.backfill replay before being parked; Postgres kept the evidence (sid_method_long backfill closed=852 pf=1.232 net=1689.78 dd=12.93%) even though the code never landed. This is the worst repeat yet of the item 10/14/21/25/26 loss shape — the same hypothesis lost and [secret] five times running is pure waste, not research.

  5. earnings_gap_continuation retired live; seat freed, item 21 (PEAD pinned-99) ADMITTED over the queue's single-session sibling

    **7-day lab_events review found something more urgent than backlog item 27 (top open): earnings_gap_continuation (single-session hold on the reaction-pop trigger, admitted 2026-09-21) retired 2026-10-31 (shadow_failed(n=40,pf=0.75,age=40d)); the freed seat sat unused two days. Pulled its 40 live lab_trades directly: 14W/26L, net -$99.38, PF 0.747 — checks out, and at n=40 reads as variance around a modest backtested edge (PF 1.12-1.14), not a decisive live refutation. Retirement is [path]'s own call, not research's to second-guess.

  6. item 26 (PEAD outsized-gap floor) recovered from lost WIP, ADMITTED; sid_method_pattern retired, seat already spent by the lost session

    7-day lab_events check found something more urgent than a fresh backlog item, again. sid_method_pattern retired 2026-10-29 00:20 UTC (shadow_failed(n=40,pf=0.85,age=38d)) — a seat freed — but a session had already run against it at 02:48 UTC the same night and left its work uncommitted: git branch -r showed research/wip-20261029T024859Z, parent exactly current master, the same loss shape as items 10/14/21/25. Its tree built [path] (item 26: PEAD's reaction-day-pop/3-day-hold rule restricted to pops >10%, the complement of item 19's dead moderate-band cap) + 7 tests + a sensitivity report, and had already run python -m lab.admit pead_earnings_drift_outsized successfully — the DB write (SHADOW row, lab_strategies) survived in Postgres even though the git commit never landed, confirmed directly: lab_strategies shows it born 2026-10-28, 16 SHADOW / 1 FUNDED, sid_method_pattern absent.

  7. nursery seat freed by sid_method's retirement; item 23 (sector-relative short-term reversal) ADMITTED

    7-day lab_events check found something more urgent than the top backlog item. sid_method retired 2026-10-24 (shadow_failed(n=41,pf=0.13,age=33d), [path]'s own call) — confirmed against lab_strategies directly: 15 SHADOW / 1 FUNDED, no admission attempted against the freed seat since. Four READY candidates (items 20/21/22/23) have been queued behind capacity since 2026-10-15; per the 2026-10-19 session's own comparison, item 23 (sector_reversal_1m) carries the strongest evidence of the four — individually significant IS (t=2.617) and OOS (t=2.175) even uncorrected, where the others lean on lab.admit's portfolio bar rather than a per-trade t-stat. Worked the capacity question instead of a fresh backlog item, the same precedent as the 2026-10-15/-16 sessions.

  8. item 25 (sector-capped earnings-gap continuation) recovered twice-lost, DEAD; a 4th qty_drift mismatch (CCI) checked, owner already responding

    7-day lab_events check first. qty_drift fired a 4th time (CCI, pead_earnings_drift, ledger=20 broker=27, id 3511, today) — not a repeat of last session's unaddressed finding, though: the runner now has two new mechanisms absent a week ago, orphan_flattened (ARM's 10-21 drift was auto-flattened today 13:45 UTC) and entry_refused:symbol_owned_by — the owner is visibly already acting on the growing-drift flag; CCI is one new instance of the still-open mechanism, not an escalation. No research action possible (protected files). Noted, not [secret] at length.

  9. qty_drift is growing, not a one-off: 3 unresolved overfills, two distinct mechanisms; no backlog item taken

    7-day review overrides the backlog, per contract. Last session's single ARM mismatch (10-21) became mismatch=3/new=2 the very next audit (lab_flags. last_audit, 10-23 00:40 UTC): pead_earnings_drift SPCX ledger=4 broker=11, reversal_fade_1d_breadth T ledger=78 broker=157, plus the still-unresolved ARM. All three positions are open right now (lab_trades 76518/76522/ 75842, no exit_ts) — live excess exposure, not closed history.

  10. item 24 (vol-spike early exit) built, tested, READY; live qty_drift mismatch found, left for the owner

    Urgent check first: tonight's fills audit flagged the test account's FIRST-EVER mismatch (lab_events 2951): reversal_fade_1d_breadth ARM, ledger=7 broker=14, traced to qty_drift event 2348 — the runner's own I5 invariant already paged the owner and took no action, by design. One ledger row exists for this position; the broker holding double is an order- submission/fill-adoption question inside [path]'s pending-order lifecycle (protected, not a strategy-file bug — strategies only emit Intent signals). No research-side fix exists; flagged for the runner's owner. Nursery unchanged: 16 SHADOW / 1 FUNDED, items 20/21/22/23 still queued.

  11. sector-relative short-term reversal (item 23) built, READY, individually significant both windows; backlog refilled to the 8-open floor; nursery still full, no admission attempted

    7-day review: no retirements, no halts beyond supply_demand_retest's 2026-10-16 promotion (already known). lab_strategies confirmed directly: still 16 SHADOW / 1 FUNDED, no seat freed since. Routine entry_refused (symbol-ownership contention) and two exit_failed (PEP/STZ, qty already held for another order, test account) rows in the window — mechanical, not mine to act on. One housekeeping find: [internal file] was already below the 8-open floor BEFORE this session touched anything (confirmed via git stash — test_backlog_has_at_least_eight_open_items_and_a_dead_section fails on clean master too, 7 open items, not 8) — the last two sessions added items but never refilled after taking one. Fixed this session (see below).

  12. item 22 (MACD basket) built, tested, READY; nursery still full, no admission attempted

    7-day lab_events review: nothing more urgent than the backlog's own queue — audits clean since 2026-10-15; lab_strategies confirmed 16/16 SHADOW / 1 FUNDED (already known from c83128c). Saturday: outside the Mon-Thu admission window too.

  13. item 13 ADMITTED overnight; item 21's lost session recovered from an orphaned `run.sh` WIP branch

    (1) Item 13, bookkeeping only. supply_demand_retest promoted SHADOW→FUNDED 2026-10-16 00:20 UTC ([path]'s call), freeing a seat the nightly pass gave to reversal_fade_1d_breadth 2 min later: admitted, 547 trades, PF 1.29, dd 13%, net +$1,416 (matches the 2026-10-09/-15 research numbers). Seat refilled instantly, so earnings_gap_continuation_99 (item 20) hit nursery_full again 7s later. BACKLOG item 13 → admitted.

  14. item 20 built and READY; still blocked on nursery capacity, not evidence

    7-day review: no retirements, no halts. One recurring audit mismatch (xsec_momentum_12_1 TWST, first paged 10-08) still rolling through the window, new=0 since — no action needed. Confirmed nursery state directly against lab_strategies rather than trusting last session's note: still 16/16 MAX_TEST SHADOW, sector_low_vol still the 16th seat, still no retirement. The automated nightly attempt on item 13 (reversal_fade_1d_breadth) refused nursery_full again at 2026-10-13 02:32 UTC — no seat freed since Monday. Per the prior session's own pointer, resumed at item 20 rather than re-spend an admission attempt on item 13's guaranteed capacity refusal.

  15. nursery hits capacity; item 14 recovered from parked WIP; item 18 blocked; item 19 DEAD

    7-day review: one recurring audit mismatch (xsec_momentum_12_1 TWST, paged 10-08, still rolling through the window, no new action). No retirements.

  16. breadth filter rescues short-term reversal from its drawdown refusal (item 13)

    Checked lab_events first: no retirement/halt. One audit mismatch (xsec_momentum_12_1 TWST sell 2026-10-07, ledger $159.29 vs broker $158.43) — already paged once on 10-08 (new=0 since), just rolling through the 7-day window; no action needed. Noted in passing, not investigated further: sector_low_vol (item 12, DEAD this morning on 4 independent nightly replay_failed windows) flipped to test_route replay_failed → bar overnight per lab_strategies/lab.select's own rolling replay. That table/route is [path]'s alone per the contract — not [secret] here, flagging only so a future session isn't surprised by it.

  17. Sector-relative low-vol tilt DEAD on an IS-only artifact (item 12); backlog refilled

    7-day review: one audit mismatch, explained not ignored: xsec_momentum_12_1 TWST sell on 2026-10-07, ledger $159.29 vs broker $158.4278, paged the owner 2026-10-08 (new=1), now a repeat. tsla_orb15_holdeod locked out of TSLA by tsla_macd_trend all week (symbol_owned_by). tsla_macd_trend flipped bar↔replay_failed on select.py's own [secret] as it crossed the 40-trade floor — expected, not mine. No retirements/halts. Positive control: sector_momentum_12_1 2yr backfill: PF 2.748, 91 trades — known range, trusted.

  18. QQQ MACD trend DEAD on a trade-count floor, not its PF; backlog item 6 mislabel corrected; refilled to the 8-open floor

    Autopsy: lab_events last 7 days show nothing urgent beyond what the backlog already carries — the 2026-09-27 test_route_changed flips (bar → replay_failed) on earnings_gap_continuation/rvol_fade_5d/tsla_macd_trend corroborate item 15's already-diagnosed dilution finding, not a new fire; one mismatch=1 audit blip on 2026-10-01 self-resolved by 12:22 the same day (drift_flattened, CCL). Proceeded with the backlog's own rank order.

  19. Earnings-gap continuation's edge does not survive the widened universe (item 15 DEAD)

    7-day review: audits clean except one self-resolving mismatch (test acct, 2026-10-01: pead_earnings_drift/earnings_gap_continuation both missed a CCL fill behind the 15s poll on 09-30, [secret]/double-bought, flattened by drift_flattened at 12:22 the next session — [path] territory, already landed in 0a80564/b7f703c, not mine). Also netting events (sector_momentum_12_1/sid_method sharing MRNA, xsec_momentum_6_1/ sid_method sharing ASST) — normal symbol-contention, resolved by netting_ended same day. No retirements/halts. Positive control: tsla_macd_trend 2yr backfill PF 1.622, 23 trades, net +$2,356 — known range, trusted.

  20. two sessions' work found lost, not re-lost; dual sector momentum DEAD (no-op)

    7-day review: audits clean, no halts. select.py's automatic [secret] (v5 §2.9/2.10) sized 3 books down to override on 2026-09-27 (test_route_changed: bar → replay_failed, current replay stopped clearing the bar — tsla_macd_trend's ~23 trades/2yr is just under the 40-trade floor, a grandfathered seed, not new) — select.py working as designed, not mine to fix. Positive control: re-ran tsla_macd_trend, PF 1.593/23 trades/net +$2,287.56 — known range, trusted.

  21. sector_momentum_6_1 ready, blocked on the Mon-Thu window; item 17 is genuinely infra-blocked

    7-day review: one audit mismatch, explained not ignored: mismatch|NTRA| buy/sell|2026-09-25 (test acct, already seen once, not repaging). Traced to sid_method_pattern's NTRA: a stop rearm was already through the market twice, the exit retry came back not_filled, orphan_flattened closed it at market minutes later — disagreeing with the ledger's own exit. Commits db176a1/ 0859298 (16:53/17:35 ET 09-25, broker-truth-before-stops) landed AFTER the 13:30 incident that caused it: stops a repeat, doesn't retro-fix one row. Not my writable surface. No retirements/halts. tsla_orb15_holdeod still locked out of TSLA by tsla_macd_trend (6th session, not actionable). wm_neckline/ bollinger_pullback*: 0 closed SHADOW trades in 5 days (pending-order books arm/cancel on invalidation, not a bug). Positive control: tsla_macd_trend 2yr backfill PF 1.703/23 trades/net +$2,565.67 — trusted.

  22. 6-1 cross-sectional momentum ADMITTED (owner priority: item 16)

    7-day review: audits clean (one transient unrecorded=1/new=1 test-account blip, 09-24 00:40, self-resolved next run — not a mismatch). tsla_orb15_holdeod still locked out of TSLA by tsla_macd_trend, 4th session flagged, not actionable here. No retirements/halts. Positive control: tsla_macd_trend backfill: PF 1.797, 23 trades, net +$2,803 — known range, trusted.

  23. Long-the-losers reversal DEAD; item 6's data gap now grounded in a real API check

    7-day review: audits clean. No retirements/halts. tsla_orb15_holdeod still locked out of TSLA by tsla_macd_trend (symbol_owned_by, continuous since 09-21, third session flagged, not actionable from research). review_push_failed (09-21), outside my surface. Nothing more urgent than the backlog. Positive control: lab.backfill --strategy tsla_macd_trend --years 2: PF 1.792, 23 trades, net +$2,787 — matches last week's range. Trusted.

  24. ORB on QQQ direct: passes the literal admission window, dead on the longer history

    7-day review: all audits clean. No retirements or halts. tsla_orb15_holdeod locked out of TSLA all week by tsla_macd_trend (symbol_owned_by, same open symbol-contention question flagged last session, not actionable from research). One review_push_failed event (09-21), outside the researcher's writable surface. Nothing more urgent than the backlog. Positive control: tsla_macd_trend reproduced its known range (PF 1.779, 23 trades, net +$2,740). Trusted.

  25. Reconciling a day of direct owner activity; low-vol tilt DEAD

    7-day events / autopsy: item 2 was already admitted before this session started — the owner ran python -m lab.admit earnings_gap_continuation directly today (Mon-Thu ET window, exactly as last session flagged): 397 trades, PF 1.12, max dd 5% of cell, net +$271. The owner also built two new book-chapter books (bollinger_pullback, bollinger_pullback_trail) and ran the sensitivity report on every registered strategy. Four already-refused books (wm_neckline, supply_demand_retest, sid_method, sid_method_pattern) plus the two new bollinger books were then admitted to SHADOW via the documented --override path — a human decision, in writing, published on each lab_strategies.note ("a live audition on the test account regardless of the replay; the retirement rules decide"). The replay verdicts themselves did not change; only their live-tradability did. This session ran no override — recording it here so the log is honest about who decided what. Also noted, not actionable from research: tsla_orb15_ holdeod was locked out of TSLA all day (entry_refused: symbol_owned_by:tsla_macd_ trend, dozens of 5-min ticks) — a symbol-contention question for the runner's owner.

  26. Gap continuation after earnings: built, tested, admission blocked by the (new) weekend gate, not the backtest

    7-day review: SID Method refused admission this week (dead, 6ad6680). Two admissions already used by an earlier session (xsec_momentum_12_1, sector_momentum_12_1, both admitted 09-18). Audit self-resolved; 09-16 orphans already cleaned up in the v4 cutover. Nothing more urgent than the backlog. Positive control: today's automatic tsla_macd_trend backfill reproduced its known range (PF 1.874, +$3,098.59, 23 trades) — trusted.

  27. Audit false-positive diagnosed (unfixable here); short-term reversal DEAD on drawdown

    Positive control: re-ran tsla_macd_trend's 2yr backfill fresh, reproduced PF 1.877/23 trades exactly. Trusted the session after that.

  28. Sector-relative momentum ADMITTED; three

    backlog items reclassified BLOCKED on real architecture limits

  29. First swarm gate_candidates arrive and decisively FAIL family-wide verification; a retired-genome sizing bug traded $0-cell mp for two days (already fixed same-day by deploy); firstbar_momo two days from its kill line

    Autopsy (watermark 2026-09-13; trading days since: 09-14, 09-15, 09-16 partial).