Internal working note. The loop-4 accountability pass, run a day late: the Monday flagship itself shipped on time (2026-08-24-market-brief, edition flagship), but this autopsy — its required half — was never written, and cadence:check reported the whole loop overdue as a result. Miss-screen basis: settled 2026-08-25 summaries, ≥+30%/30d movers with $10M+ dollar volume — 88 movers, 74 of them unclaimed, 0 stale skipped. Conviction basis: today's conviction:execute dry-run. Ledger basis: the nightly call-check inside the 2026-08-25 batch.

The finding that matters: the conviction pipeline is jammed shut

Last week's autopsy diagnosed two unhandled in-zone flags (CBRS, ABCL) as a human-loop failure and filed one fix: read the conviction dry-run in every daily. That fix is now provably insufficient, because reading it does nothing.

Every row in today's dry-run is skipped. Twenty-four rungs, zero convertible. The breakdown:

• 14 rows — agent verification required (latest in-zone flag is code-generated). The oldest is NBIS, flagged 2026-07-12 — six weeks unverified. NVDA's rung has sat since 2026-07-22, five weeks, and NVDA reports tomorrow. Also queued: MSFT, CBRS, ABCL, MU, GLD, TLT, STX, CEG, ARM, LEU, SKHY, TER, ZM.
• 8 rows — latest flag is invalidation (TLN, KGS, SMH, WULF, META, ELAN, VST, NTLA). Correctly parked.
• 2 rows — declined (pass 2026-08-25) (MU, SIL). These are the only two that moved, and they moved today.

So the pipeline's throughput last week was two decisions out of twenty-four open rungs, and both landed after the week closed. Last week we called CBRS/ABCL a discipline smell on the part of the operator. That was too generous to the machine: a code-generated in-zone flag can never convert without a human verdict, nothing schedules that verdict, and so the queue fills monotonically. The desk did not fail to look. It has no mechanism that converts looking into a decision.

This is the same shape as the miss-screen problem below — detection works, adjudication has no owner — and it is now the desk's most expensive structural gap.

The misses, triaged

• The software re-rate is STILL unowned, one week after we named it. Last week: seven unclaimed movers in one neighbourhood, verdict "a LANE gap," no action taken. This week the same cluster is EIGHT names and every one is still MISS: TEAM +91.6%/30d (was +70% — it ran another 22 points, now +54.2% vs its 200-day at RSI 70.9), U +57.2%, PATH +53.7% (RSI 71.0), WDAY +43.7%, ESTC +37.0%, SAP +35.6% (RSI 73.0), MDB +35.8%, GTLB +33.2%. All sit in cloud-saas / cloud-etf-holdings / tech-insider-buys — watchlists we already own. Naming a gap is not owning it. The standing decision is now overdue by a full week: stand up an application-software re-rate lane, or record an explicit decline with a reason. Silence for a second week is the finding.
• Precious metals — same failure, louder. Last week: "GDXJ +31%, AEM +38% — the gold-miner echo… not our lane today, and saying so out loud is the point of this line." Saying it out loud did not make it a decision. This week: AEM +54%/30d at RSI 84, AU +55% (RSI 81), IAG +51% (RSI 77), WPM +47% (RSI 81), GDXJ +39%, SIL +36%, AGI +32%, AG +30% — eight names, all MISS. GDX is benched under gold-crash, so the desk owns the wrapper and none of the miners. A bench seat on the ETF while eight constituents run unclaimed is not coverage.
• Biotech — the desk is half-right and it is worth saying which half. Owned and paying: MRNA +193%/30d HELD, ABCL +135% HELD, TWST +68% HELD, RGEN +38% BENCH:biotech-capital-cycle. Unclaimed in the same melt-up: AMLX +103%, IOVA +69%, TEM +61%, GENB +52%, TXG +40%, CDNA +37%, HALO +33%, ARKG +32%, CRL +31%, CRSP +31%. TXG is the one that stings — it is in the biotech-capital-cycle watchlist and was named in last week's conviction discussion, and it still carries no position, no bench seat, and no recorded pass.
• AAOI — the tape closed this question for us, and we should not take credit. Carried open last week as "does AAOI earn a bench seat in optical or a recorded pass." Today: $113.15, −13.98% on the week, −36.30% over three months, RSI 45.3, weak-down. The run is over. Declining the seat now is correct — but the desk never made a decision, it simply ran out the clock and was rescued by mean reversion. An unrecorded non-decision that happens to work is still a process failure; next time the clock runs the other way.
• Nuclear fuel cycle — the week's genuine win, and the control case. UEC +40%/30d and UUUU +41%/30d both tagged BOOK:nuclear-fuel-cycle. This is the ONLY cluster in the screen where a stood-up lane captured the move instead of watching it. Same tape conditions as the software cluster; the difference is that somebody wrote the thesis down. That is the whole argument for closing the two gaps above.

Coverage miss vs conviction miss

Coverage misses this week: zero, again. Every unclaimed mover was already inside a watchlist or screener bucket before it ran. Two consecutive weeks of this says the acquisition and screening layers are not the constraint and should stop being tuned as if they were. The losses are entirely in adjudication: an unowned lane named but not stood up (software), a lane declined by silence rather than by decision (miners), a conversion queue that cannot discharge (24 rungs, 2 decisions), and a question answered by price decay rather than by judgment (AAOI).

Filed this session: ops/tasks/TASKS-ENGINE.md — miss-screen returns a feed, not a finding; cluster MISS rows by shared watchlist and demand own-or-decline, with the prior week's verdict carried forward so a repeat reads as a repeat. Threshold tuning was measured and rejected (at +50%/30d the software cluster degrades to three names and the pattern disappears).

Calls graded, both directions

• What worked. SPY $765.91 (+3.6%/30d, RSI 54.2) — trend intact, no new entry called, none needed. TLT's 08-19 in-zone reclaim is holding: $83.47, +2.23% on the week, no new low, which is precisely the condition the call said would confirm it.
• A STALE-BASIS trap in our own invalidation, caught this session and worth more than the rest of this note. HYG's call reads "a close below the 200-day ($78.35)." The parenthetical was the 200-day on 2026-08-18. The average has since risen to $80.15, and HYG closed today at $79.92 — 0.30% below the live 200-day. Read against the number written in the call, the tell is comfortably intact; read against the condition the call actually states, the invalidation has just been met. A hard-coded illustrative value inside a moving-average condition silently converts a live trigger into a frozen one, and the frozen version can never fire as the average climbs. Nothing flagged this — the board still lists HYG as "a real base is forming." Two caveats against over-reading it: the breach is 0.30% on day one, and the surrounding tape (RSI 59.5, weak-up, +0.87%/30d) looks nothing like credit cracking. Recorded as an in-zone/invalidation flag this session; the class belongs in ERROR-PATTERNS.md as a variant of unlabeled-basis, since every other prose call naming "($X)" against a moving average carries the same defect.
• What did not. GDX's call said don't chase RSI 70 after a +21% week; wait for a pullback that holds the 200-day. GDX is now $105.52, RSI 77.2, +40.3% over 30 days, and the pullback never came. The discipline was defensible and the cost was the entire move — this is the same judgment that left eight miners unclaimed above, and the two should be graded as one decision, not two.
• A flag-quality bug, not a call miss. KRE carries a code invalidation fired 2026-08-20 on last 74.71 <= trigger 75. But the call's written condition is a two-leg test: a close below the August low on the same day HYG loses its 200-day. HYG never lost it — it is $79.92 today. So the numeric trigger fired on one leg of a two-leg condition and the ledger now carries an invalidation the prose does not support. The single-leg trigger is the defect; the call itself is intact.

Verdict + reasoning

Open, carried to next Monday, with last week's list deliberately unchanged because none of it was discharged:

1. Conviction queue throughput — the new headline. Either code-generated in-zone flags get a scheduled verification slot with a bounded queue age, or the flag→convert path is decorative. A rung unverified for five weeks on a name reporting tomorrow (NVDA) is the proof.
2. Software re-rate lane — stand it up or decline it in writing. Second week overdue.
3. The miners — GDX benched while eight constituents run is a coverage fiction. Decide.
4. AAOI — close it as a recorded decline, and note in the record that the tape, not the desk, resolved it.

The screen's own scoreboard stays the referee.

Sources

• research/market-engine/data/summaries/*.json (settled 2026-08-25)
• miss-screen, conviction:execute, call-check --dossier, trip-report — this session (runlog 2026-08-25)
• research/issues/2026-08-24-market-brief.md (the Monday flagship this autopsy accompanies)
• research/investigations/2026-08-17-miss-autopsy-week-ending-2026-08-14.md (the carried-open list)
• research/targets/2026-08-18-*.md, research/targets/2026-08-19-*.md (the week's desk-call slate)
• ops/tasks/TASKS-ENGINE.md (miss-screen clustering row filed this session)