Internal working note. This is the loop-4 accountability pass for the week ending 2026-09-18, written on the Monday 09-21 corpus. The Miss Autopsy is the one place the desk may read its own books, so paper ledgers are cited below. It covers the two misses the user named, memory's rebound and shipping's run, as deep dives. The mechanical miss-screen table is at the end. No autopsy was written for the weeks ending 09-11 or 09-18 until this one. The ritual lapsed after 09-08, which is itself one of the findings.

All prices are settled closes through 2026-09-18, computed in code from the engine's daily bars. The shipping funds (BWET, BDRY, BOAT, SEA) had no coverage; their bars were first fetched on 09-21, and the 09-18 bar is used here.

The short version

The common thread: in both cases the desk's reading was right, early. The failures came afterwards:

• A pass or deferral carried no condition for reopening it.
• A condition written in prose was never turned into a trigger.
• The thesis was judged on the wrong instrument: the cohort's structure for memory, crude's price for shipping.

This is the same finding as the 09-08 autopsy: "the desk's passes are wrong three times in four." This week adds two more cases.

---

Miss 1: the memory rebound

What the rebound was: one 13-session V, then a month of chop

The whole cohort bottomed on 2026-07-29, the day of the Korea margin cascade and the Situational Awareness forced unwind.

• The whole gain came in the first 13 sessions (07-29 → 08-17). Since then the cohort has gone sideways: pullback lows on 08-24/25 (MU −10%, SNDK −17%), a second dip on 09-15, and a 09-18 close right at the leg-1 high.
• Every breakout above the leg-1 high has failed or is one day old. MU cleared it on 09-04 and was −8.7% by 09-15. DRAM cleared it on 09-08. SNDK cleared it on 09-18, by 0.3%.
• So a bottom → pop → flag → breakout entry buys at the leg-1 high, which is where prices still are. The rebound the user expected is real, but it is behind us. What is left is whether the breakout holds, and Micron's 09-30 results are the next test.

What the desk said and did

No paper book bought any memory name between 07-29 and 09-18 (all research/trading/paper/*/LEDGER.json scanned). The only trades in the window were the conviction stop-out at the low, and autotrade's DRAM exit on 07-20 (price below its 20-day).

Why we missed it

1. Dominant: a conviction miss built into the rules. Every entry the desk wrote required a formed base, a 20-day reclaim plus a base, or a flag plus a held base. The rebound was a V whose whole gain came in 13 sessions, so the structure entry could only fire near the leg-1 high. There were five explicit "base unproven" passes. This is the user's own trend-hold style working as designed: it does not buy the pop. The miss is the known cost of that style, and until now it hadn't been written down anywhere.
2. The exits worked at the capitulation; nothing was built to enter it. The conviction book bought MU into the fall, and its default −15% stop sold both lots at the exact low. The void rungs crossed the same day. Then nothing re-armed on the reversal. The machinery is complete for "get out" and blank for "the flush just ended".
3. A condition written in prose was never coded. On 07-31 the desk wrote the right test, "whether this cohort builds higher lows". It was met on 08-04, still 9–15% under the leg-1 high. Only price levels (confirm and void rungs) were machine-watched, so the prose condition was never evaluated.
4. A known gap wasn't built in time. The early-turn brief (07-27) named this exact blind spot two days before the low. turning[] shipped on 08-24, after leg 1, and the breadth-thrust gauge (TASKS-ENGINE.md:1140) is still blocked on research.
5. Later triggers were set at levels already crossed. The 09-07 SNDK line (1,740) and the 09-15 MU trigger (989.96) both sit below the 08-17 close. In effect they say "buy the V's top".

Not a coverage miss. Memory is the P1 primary thesis: a 10-name watchlist, a perspective updated three times on 07-30, and daily lane reads. The thesis was right: the perspective never flipped (HOLD, then INTACT at the 09-06 Campaign 1 review). The 13F mirror books (tepper, druckenmiller, aschenbrenner, coatue) held MU and SNDK through the whole drawdown and rebound; that is passive exposure, not a catch.

What would have caught it at T−1

A cohort capitulation-reversal flag, checked at the settled close. It fires when all four hold on the memory cohort (MU, SNDK, STX, WDC, SIMO, EWY):

• median drawdown from the 252-day high is −25% or worse
• median RSI14 on the prior day is 40 or below
• median one-day change is +5% or more
• at least 80% of the names are green

Over 401 sessions (2025-02-13 → 2026-09-18), the strict version fired twice:

• 2025-04-09 (the tariff-pause reversal): cohort median +4.5% over 10 days, +15.7% over 20.
• 2026-07-30: +8.4% over 10 days, +3.5% over 20. From the 07-30 close to the 08-17 high, MU gained +15.7% and SNDK +39.6%.

A looser version (drawdown −20%, RSI 45, up 4%) fires four times, adding a false fire on 2026-07-21 (−4.3% over 10 days).

Two fires is not evidence of an edge, and a rule fitted after seeing the miss is exactly the hindsight this note should distrust. So this is a flag that sends an analyst to read the reversal day, not an entry rule. If it ever becomes an entry, it needs its own stop at the capitulation low, not the flat −15% that sold the low. That is a separate style decision (see "Proposed", item 5).

---

Miss 2: shipping and freight

What ran

• Not covered at all: TNK, NAT, TRMD, IMPP and ZIM have no bars and are on no watchlist. SBLK, GOGL, ASC and NVGS were named in the 04-12 trace and are also absent.
• The moves began in January, before the war reached Hormuz. The first Hormuz capture is 03-02.
• These were steady climbs, not spikes. The tanker stocks almost never cleared +30% in 30 days on an autopsy date, so a 30-day screen couldn't see them even where we had coverage.
• BWET's price is real. There is no split artifact: the biggest single day was +28%, and daily volume grew from hundreds of shares to about 200k. Press coverage matches, putting it at +3,600% YTD in early September and the best non-leveraged US fund this year (CNBC 09-13, TheStreet).

What the desk said and did

Why we missed it

1. Dominant: a conviction miss, judged on the wrong instrument. The desk found the vector twice (04-12, 07-01) and deferred because crude had faded at the truce. Freight had already decoupled from crude; it was pricing routing and insurance, not barrels, and it kept running. The "spike-and-fade, not trend-hold" objection was also wrong by the desk's own mechanical definition: FRO and INSW made 52-week highs above their 200-day averages on 01-13 and trended for eight months.
2. The deferral had no reopen condition. Four tanker names later made new 52-week highs, and nothing brought the 07-01 question back.
3. The purest instrument was invisible to every scanner.
   • monster-discover screens Finviz stocks with market caps of $2B and up (python/src/discover.py:49-65). Funds carry no market cap, so ETFs can never pass.
   • monster-discover:etf scores all US ETFs, but only those we already hold bars for (208 of 5,641 on 09-21). It isn't on the cadence and has never saved a report.
   • monster-etf-scan is a curated list of 51 funds with no shipping funds.
   • miss-screen reads only covered names.
   • BWET cleared the $10M/day liquidity bar in March ($12.6M/day), and was over $100M/day by August.
4. Intake had the evidence and never looked. Five captures came from four followed accounts, including one (@Vmaxpax, 08-18) that stated the freight-not-crude thesis a month before the desk did. The queues went unreviewed. The review rule, "novel symbols: quarantine, zero adds", treats fetching bars (free) the same as adding to a watchlist (a real decision). The 09-06 review also archived the Hormuz cluster as misbucketed.
5. The ritual lapsed. No autopsy after 09-08. This wouldn't have changed BWET, which was never covered, but ECO (+32% over 30 days, unclaimed) is on this week's screen and nobody read it.

"Not our game" doesn't hold. BWET is not on config/avoid-instruments.json, and it isn't leveraged or inverse. It is a freight-futures commodity pool, which is a plain US-listed fund. Two questions are genuinely open before anyone buys it: its tax form (a commodity pool may issue a K-1) and whether the Robinhood account can trade it. Had it been covered, it would have been a MISS row on every autopsy date from 07-31 to 09-18 (+44% to +113% over 30 days).

What would have caught it at T−1

---

What today's reading says about these misses

The user shared a set of trading links on 09-21 (ops/workshop/ai-trading-stack-roundup/, ops/workshop/horizon/, and the Jev study's _trading/ cluster). None is worth copying as code. Several of them name the exact mechanisms behind these two misses:

• The memory miss is a textbook failure regime. The "Quant Canon" table's failure column lists "reversals after market bottoms" for cross-sectional momentum and "panic-rebound reversals" for the 52-week-high effect. That is the memory rebound, and it is the style this desk trades (trend-hold entries, relative-strength ranking, new-high breakouts). The table's own argument applies: these failure regimes are "the compensation structure of the premium", not defects to engineer away. So the honest options are to accept and document it, or to add a separate, bounded arm for this regime. Don't bend the trend rules to catch Vs.
• Daniel & Moskowitz (2016) described the state we missed. Momentum crashes happen when past losers rip in rebounds after panic states (NBER w20439 abstract). For a long-only book this isn't a crash, because we hold no short side. But the memory cohort were the losers (−34 to −57% off their highs) and they ripped +35–76% in 13 sessions. Their panic state is defined at market level (a falling market plus high volatility); our cohort flag above is a sector-level adaptation and needs its own test. Where possible, take the state variables from the paper instead of inventing thresholds.
• Short-term reversal (Jegadeesh 1990) is the anomaly that buys this regime. In the table: "long prior-week losers … fails on transaction costs at scale". As a cohort-level flag that routes one analyst read, the turnover cost doesn't arise. That is how to read it for us.
• Freight was a universe miss, not a signal miss. A plain time-series momentum rule (Moskowitz, Ooi and Pedersen 2012, already in R0) would have been long BWET from about February. The rule can't trade what isn't in its universe. Horizon's pitch, "the research cycle is the edge", is right about this: our cycle broke at its first stage, deciding what to look at, not at the backtest.
• Jev's best trading use case matches the freight chain. In the Jev study, Abol's thread asks about indirect impact (suppliers, customers, competitors, geography): "around 84% of impactful judgements were on stocks the article never directly mentioned." Hormuz → tanker rates → BWET is that chain. But the desk did derive it (04-12, 07-01), so a router adds little at derivation. Where it would have helped is intake: ranking the five BWET captures into an analyst's hands instead of leaving them in a 1,655-row unreviewed queue. That supports J2's ranked-inbox half, not a new derivation tool.
• "One threshold per action, scaled to what being wrong costs" (the RohOnChain blueprint's one good line). Intake applies one rule to two actions that cost very different amounts: fetching bars (free) and adding to a watchlist (a decision). T2 above is that rule applied.
• Hindsight is lookahead (the ai-hedge-fund study). Every trigger here was fitted after seeing the miss. T1 is the only one tested across a universe (≈0.8 pings/week, 6 of 28 real winners). The memory flag has n=2. Treat all of them as reasons to look, and grade them forward before any becomes an entry.

Proposed (not filed; the user's call)

Ranked by what they fix per unit of new machinery. The desk's standing rule is to delete rather than add, so the first three need no new code.

1. Restore the Monday autopsy (cadence loop-4). No code. It lapsed after 09-08, and this week's screen already has an unclaimed tanker (ECO).
2. Every defer or pass carries a reopen condition as a trigger-check row (T3). This reuses the existing trigger machinery. It would have reopened the 07-01 tanker deferral on 07-21, and the SNDK and MU passes at their 08-04 higher lows.
3. A condition written in a brief is a trigger-check row, or it doesn't count. The 07-31 "builds higher lows" test was met on 08-04 and never evaluated. Same machinery as item 2. A condition that can only be checked against text, such as "freight rates keep climbing", goes to Jev instead. Each night it becomes one yes/no question asked of each new capture. The user's idea (2026-09-21): "if we're able to set our rules into json then jev can just pattern match." That match would have fired on @Vmaxpax's 08-18 freight post, a month before the desk's 09-18 vertical. The design is J2's planned campaign-condition questions, extended to deferrals and brief conditions (in the Jev study's "Our candidates" as "watch our own written conditions"). Numeric conditions stay in code.
4. ETF tape coverage (T1) through S0, which is already planned. S0 archives Polygon's grouped-daily file (the whole US tape in one request per session). Point monster-discover:etf at it so its universe is all 5,641 ETFs instead of 208, and run it with the weekly monster-etf-scan. No new data source.
5. A style decision on V-rebounds. Either write it into the operating book as a known cost ("trend-hold does not buy capitulation reversals"), or add a bounded capitulation-reversal flag that routes one read. A flag is not an entry. If it ever becomes an entry, it gets its own stop at the capitulation low, never the flat −15% that sold MU's low.
6. Intake: fetch, don't quarantine (T2). A novel symbol mentioned by a followed account gets its bars fetched; the quarantine stays for watchlist adds. Or let J2's ranked inbox do it.
7. Open checks that fell out of this autopsy:
   • BWET tradability: tax form and whether Robinhood can trade it.
   • ECO (+37% over 30 days): claimed by no book or thesis.
   • FRO / INSW / DHT: the 07-01 recommendation to add them was only partly done (07-29).
   • gottheimer-trades inconsistency: its ledger holds MU and SNDK lots that decisions/LEDGER.json records as sold on 07-09.

This week's miss-screen (the mechanical half)

deno task market -- miss-screen, run 09-21 on settled summaries (as of 09-18): +30% or more over 30 days with $10M+/day dollar volume. 1,116 symbols covered; 15 movers; 14 unclaimed.

The per-name lines for the last two rows are still owed. This note went deep on the two misses the user named instead.

Sources

• Price data: research/market-engine/data/stocks/{MU,SNDK,DRAM,SKHY,STX,WDC,SIMO,EWY,SMH,SOXX,BWET,BDRY,BOAT,SEA,FRO,INSW,DHT,STNG,ECO,HAFN,LPG,MATX,DAC,USO,SPY}/ohlc/max.json. BWET, BDRY, BOAT and SEA were fetched 09-21 via the managed fetcher.
• Memory: research/perspectives/2026-05-19-memory-supercycle/{README,log}.md; research/issues/2026-07-18-ai-semi-derate-turn-signature-armed.md; research/issues/2026-08-08-vertical-memory-storage-the-group-splits.md; research/issues/2026-09-07-vertical-memory-storage-customer-consumption-test.md; market briefs 07-29 → 09-18; research/trading/paper/conviction/LEDGER.json and all paper/*/LEDGER.json; research/trading/decisions/{LEDGER,CALL-FLAGS}.json; .brief/early-turn-detection.md; ops/tasks/TASKS-ENGINE.md:1138-1150.
• Shipping: research/scans/2026-03-10-geopolitical-risk-archive.md; research/scans/2026-03-12-monster-scan-archive.md; research/projects/autonomous-research/traces/2026-04-12-iran-war-second-and-third-order-economic.md; research/investigations/2026-07-01-iran-war-second-and-third-order-effects-beyond-oil.md; research/issues/2026-09-18-vertical-energy-oil-freight-not-food.md; research/targets/2026-09-18-dba-target-desk-call.md; research/capture/tweets/*bwet* and 2026-07-13-TheETFTracker-*; research/capture/triage/_archive/intake-candidates-2026-{07-13,09-04,09-11}.md; research/capture/triage/reviews/2026-09-{06,14}.md; research/market-engine/python/src/discover.py:49-65; research/market-engine/cli/synthesis/miss-screen.ts; research/market-engine/config/avoid-instruments.json.
• External: CNBC, "Up 3,600%, this freight fund has posted the biggest gains of all on Iran war oil shock" (2026-09-13); TheStreet on BWET; Daniel & Moskowitz, "Momentum Crashes", NBER w20439 (abstract).
• Today's reading: ops/workshop/ai-trading-stack-roundup/{notes,what-to-take}.md; ops/workshop/horizon/notes.md; the Jev study's _trading/ cluster.