Thesis-tag reopen sweep + omission-loop backfill (overnight build, 2026-07-19)

Question: the omission feedback loop (thesis-tag-audit + thesis-verdict-scoreboard, built 2026-06-24) existed but was starved and unconsumed — 51 tags all sourced from profile refreshes, 5 reopen prompts sitting unexamined, ~30 deep-dive verdicts never tagged, and no weekly ritual reading the reports. User call 2026-07-19: build out the open follow-ups overnight.

What shipped tonight (all four parts):

1. All 5 fired reopen triggers re-examined with receipts (below), tags updated, bands refined so nothing on the board insta-fires.
2. 30 deep-dive verdicts backfilled into config/thesis-tags.json (store 51 → 81 tags), as_of = each artifact's own date so returns-since-verdict are honest. Reopen bands use ONLY artifact-stated levels; where an artifact had none the band is tape-derived and marked auto-derived. Lint: 0 errors / 0 warnings.
3. Pass-edge attribution by source stream added to thesis-verdict-scoreboard (which stream's passes run away from us), with self-check coverage.
4. Weekly consumption wired (user-ratified): both reports + two standing questions (rotating bear-case attack, book-concentration read) are now steps in the weekly money loop — ops/docs/RUN-SCHEDULE.md § Weekly money loop, mirrored in .brief/money-loop.md and PRODUCT-LOOP.md.

The five re-examinations

The backfill (30 tags, verbatim-level bands)

Sources: 28 June–July deep-dives (extracted by 4 read-only analyst agents, desk-judged) plus the two 2026-07-18 copper dives. States as the artifacts rendered them: intact ×14 (GENB MEDP SDGR CGEH LAES ARGX CBRS EWY PL RKLB STX WDC FCX SCCO), intact-but-priced ×14 (CIEN ABCL ELAN RGEN INFQ IONQ QBTS RGTI KLIC OUST ILMN NTRA RVMD VPG), broken ×1 (QUBT), no-view ×1 (ALMU).

Found at backfill — six constructive June verdicts already stopped out. STX, WDC, EWY, PL, RKLB, CBRS all sit BELOW their artifact stop/support levels (e.g. STX 787.66 vs stop ~810; EWY 162.54 vs stop ~184; RKLB 67.62 vs ~70). The June trend-hold entries never fired (each required an SMA20 reclaim that never came), so the entry discipline kept us out of the drawdown — but the constructive verdicts are tape-contradicted and each tag carries a "re-exam owed" note. Consolidated follow-up row in TASKS-RESEARCH.

Twelve below-legs that would have insta-fired at the current price were intentionally not armed (a permanently-firing leg is anti-signal — the A222 rule); the breaches are recorded in notes instead.

First honest read of the loop (post-backfill, data through 2026-07-17 close)

Scoreboard: 27 cautious tags priced; avg return of passed names −8.7% → passing SAVED +8.7% per pass; 10 correct avoids (<−10%), 0 passes ran >+25%; worst miss RGEN +16.4%. By stream: deep-dives 16 passes avg −14.6% (deep-dive caution has been excellent), company-profiles/investigations ≈ flat (freshly re-affirmed clocks). The DELL-class error is currently ABSENT — the June caution was right in aggregate.

Audit board: 7 prompts, all genuine (0 insta-fires). For the Monday pass:

• intact-broke ×3: RKLB −31.0%, PL −28.2%, WDC −26.8% since 2026-06-29 — covered by the stopped-out re-exam row above.
• caution-overridden ×4: ELAN (+4.5%, −6.8% from high, strong-up — the breakout branch of the livestock pick is firing while we sit tagged "priced"; Q2 print ~early Aug is the decision point), ILMN, NTRA, RVMD (all strong-up within 6% of highs). These are the exact "wait for pullback silently becomes missed the confirmation" candidates the loop was built to catch — each needs a desk verdict Monday.

Verdict + reasoning

Resolved as built: the omission loop is no longer write-and-forget. Verdicts from every stream now enter the store when made, fired triggers get re-examined with receipts, the scoreboard attributes pass-edge by stream, and the weekly money loop consumes all of it plus the two standing questions. The board's 7 open prompts are Monday's queue, not tonight's — flips from price alone would repeat the error class this loop exists to prevent.

Sources

• research/market-engine/config/thesis-tags.json (51 → 81 tags)
• research/market-engine/data/summaries/ (2026-07-17 close)
• research/market-engine/data/stocks/{BRBR,HUBS}/edgar/key-financials.json (pulled 2026-07-19)
• research/deep-dives/ 2026-06-01 → 2026-07-18 (30 verdict artifacts, listed in tag sources)
• ops/docs/RUN-SCHEDULE.md § Weekly money loop; .brief/money-loop.md; PRODUCT-LOOP.md