Two Mondays passed without this note being written. It is being written now, on the September 8 settled close, because the thing it exists to catch has been running the whole time and the ledger has been keeping score.

The finding that matters: the desk's passes are wrong three times in four

The decision scoreboard grades passes, not just buys. For the claude:conviction actor it currently reads: 43 decisions, 28 executed, 15 passed, PASS_MISS 0.75, PASS_COST +1.26. Three out of four names this desk declined went on to beat the benchmark, at an average cost of 1.26 points each. Its executed hit rate over the same period is 0.18 — the lowest of any actor on the board carrying a meaningful sample.

The conditional split is worse and more specific. The scoreboard tracks a condition called "pass in uptrend": four such decisions, two scored, and WIN1m on the miss measure reads 1.00. Every single scored pass on a healthy uptrend was a mistake.

This is not an abstraction. On August 25 the desk passed MU with the reason "trigger met by falling into it, not by basing: MU -5.83% today and -10.04% on the week." MU closed September 8 at $1000.26, +14.0% over thirty days and 63.9% above its long average, and is now the single cleanest trend-hold name on the board. The same session it passed SIL for "chase risk, not a fresh break — $100.37 is 16% above the $86.39 trigger." SIL closed September 8 at $98.50, still inside its zone.

And the desk did it three more times today

On the September 8 close, ten entry levels were crossed. The desk recorded in-zone verdicts on all of them and converted none, then logged three explicit passes — COHR, GLW and IREN — with a reason of the same shape as the August 25 MU pass: the price leg cleared but the written condition wants a reclaim, a held flag and business evidence, and today is the pop rather than the flag.

Each of those individual judgments is defensible in isolation. The optics and neocloud complex is up double digits on the week while still negative or barely positive over thirty days; COHR is -20.4% on the month; SOXX cleared by one percent. Waiting for a flag is the written rule and the rule was followed.

The autopsy's job is to notice that "defensible in isolation" is exactly what the previous fifteen passes also were, and the ledger says the aggregate of those defensible judgments loses 1.26 points a name. A rule that is individually reasonable and collectively wrong is the most expensive kind of rule to have, because nothing in the daily loop ever argues against it.

Coverage miss vs conviction miss

This is a conviction miss, not a coverage miss — and that distinction is the one thing that has improved. The desk saw every one of these names. They were on the board, they had written entry conditions with numeric triggers, the triggers fired, and the firing was detected and verdicted the same session. The pipeline from tape to alert is working.

What is not working is the step from alert to position. Fifteen passes, zero conversions from this actor in the scored window, and a scoreboard that says the passes were wrong. The gap identified in the August 25 autopsy — that the conviction pipeline is jammed shut — has not moved; it has simply accumulated four more weeks of evidence.

Calls graded, both directions

The desk was not wrong about everything, and the autopsy should say so.

• The exit side reads well. The three names the board carried as exits into this week — SYK, LULU, ISRG — all reconcile as genuine long-trend breakdowns rather than pullbacks: SYK -29.7% on the year and 18.2% below its long average, LULU -38.6% and 32.3% below, ISRG -25.5% and 24.3% below. Calling those broken was correct.
• The quantum vetoes were correct and remain correct. The desk passed the whole quantum cohort on August 27 and again on August 28 for being lottery-shaped, and the September 8 board read vetoed the same names again on the same grounds: QBTS 62% off its high after a +2004% two-year move, RGTI 73% off after +2191%, QUBT 68% off after +1293%. All four bounced this week and none of them repaired anything.
• The downgrade of pass discipline on healthy names is the one direction where the record is bad, and it is bad consistently rather than randomly.

Verdict + reasoning

The desk's weakness this window is not detection and not risk control. It is that the conviction actor's written entry conditions are strict enough that a name in a confirmed uptrend can satisfy the numeric trigger and still be declined indefinitely, and the ledger now has enough scored passes to say that this costs money rather than saves it.

Two things follow, and neither is a change to make tonight:

1. The "wait for the flag" rule needs a measured expiry. Right now a name can cross its trigger, get an in-zone verdict, and be passed on "no flag yet" every session forever, with no mechanism that ever notices the name kept going. A pass on a name that then clears a second time, or a third, is a different decision from the first pass and is not currently distinguished from it.
2. PASS_COST belongs where the desk reads it daily. The number exists and is computed; the daily loop does not surface it. A desk that grades bad buys and not bad passes will keep making the trade it is not being charged for.

Both are proposals for the money-loop pass, not unilateral rule changes. The rule was followed correctly today; the question the scoreboard raises is whether the rule is right, and that is a decision to take deliberately rather than mid-session.

Sources

• Decision scoreboard, actor and condition splits: deno task market -- decision board (September 8 run)
• Pass ledger detail: deno task market -- decision list --action passed
• September 8 settled tape: research/market-engine/data/summaries/*.json
• Prior autopsy and the jammed-pipeline finding: research/investigations/2026-08-25-miss-autopsy-week-ending-2026-08-21.md
• Today's in-zone verdicts and passes: research/issues/2026-09-08-market-brief.md