Question: blind rederivation spot check week ending 2026-10-09 (cadence loop-34; method .brief/correctness-rederivation-audit.md "Cadence").

Verdict: 3,393 numbers checked, 3,314 MATCH, 9 MISMATCH, 70 UNVERIFIABLE. All nine mismatches reached a reader (SHIPPED). They group into five mechanisms, and each mechanism gets one proposed DEFECTS row (below). The two machine-written scans (monster 09-30, washout 10-08) re-derive clean. Every mismatch sits in agent-written prose, an agent-built table, or a label, which is the pattern the brief predicts.

What we're asking

Do the numbers a reader saw this week survive an independent recomputation from L1, meaning the on-disk daily bars (data/stocks/<SYM>/ohlc/max.json through python/src/bars.py read_daily, as of each artifact's data date), EDGAR company-facts, Massive statements and the artifact's own cited on-disk sources?

Method

• Seed and picks. Candidates were all reader-facing files dated 2026-09-26 to 2026-10-08 in research/issues/*-market-brief.md, research/deep-dives/, research/company-profiles/2026-09-2*/2026-10-0*, research/scans/ and research/signals/, with audience: internal excluded and no 10-09 files: 139 candidates in all. random.Random(20261009).shuffle (seed 20261009) took the first five:

  1. research/company-profiles/2026-09-28-MPWR.md
  2. research/scans/2026-10-08-washout-scan.md
  3. research/scans/2026-10-02-ai-infrastructure.md
  4. research/company-profiles/2026-09-28-POWI.md
  5. research/scans/2026-09-30-monster-scan.md

• Desk calls (M=3). research/api-data/desk-calls.json has no conviction field. The ranking used (1) open calls with stance buy, because they are the only open calls that tell a reader to act, then (2) the most recent open call. That gives CRDO (buy, 2026-07-08), STX (buy, 2026-07-05) and LITE (watch, 2026-10-02, the newest open call). The only open avoid (HAL 07-31) ranks below them.

• Math. Code did all of it (scratch scripts: tape.py, washout.py, rescreen.py, monster.py, calls.py in the session scratchpad). The scripts mirror the engine's definitions:

  • Wilder RSI over the full daily series.
  • Calendar-day change7d/30d/3m from the newest bar on or before the target date.
  • change1y/3y/5y from the nearest bar within 10 days.
  • 52-week high = maximum intraday high in the trailing year.
  • relVolume against the prior 20 bars.
  • Desk-call returns use the close on or after the call date and the close on or after +N calendar days, against SPY.

  Before trusting the helper, it was checked against eight dated MPWR readings and reproduced all eight to two decimals.

• Tolerances. A value inside its printed rounding counts as MATCH. So do x.x5 rounding boundaries and integer RSI rounded twice (x.48 → x.5 → x+1).

• Caveat (A746). Summaries are overwritten in place. Every dated tape value here is therefore consistent with on-disk bars, not proven equal to the summary row the author read. Massive became the primary OHLC source on 10-06, and tonight's batch rewrote bars at 13:36 and 20:34 PT, so provider and basis revisions since publication count as UNVERIFIABLE, not MISMATCH.

What we found

Per-artifact table

What each artifact's counts cover:

• MPWR. Tape at 09-28 and 09-23, plus the 09-18 history table. Every financial cell was checked against EDGAR (company-facts.json, including the 2025 10-K restatement) and Massive. 15 refresh-history closes and their readings.
  • Unverifiable: segment revenue and growth (Q2 release, not on disk), the $194.6M restatement, 1,500 sockets, the $6B goal, the insider feed (90 sales / $848M), the social count, the Q3 guide, quote-service multiples (19.1x–19.4x / 78.1x–79.2x, which are internally consistent with price), and the volume ratios. The ratios are 2.3x and 2.1x as printed; they recompute to 2.36x and 2.26x on the rewritten Massive volume.
  • SBC: printed $53.5M; XBRL AllocatedShareBasedCompensationExpense for Q2 is $53.18M. The release basis is not on disk, so this stays unverifiable rather than a mismatch.
• POWI. The same set of checks.
  • Every EDGAR and Massive figure matches. Q3 2025 revenue really was $118.919M, the same as Q2 2026's $118.939M to the hundred-thousand. The +2.7% YoY is right on unrounded inputs (118.939/115.852).
  • "116x" is the rounded 114x repriced to $51.84. The exact value is 115.3x. The text says "roughly", so this counts as a match.
• Washout. All 217 table cells match.
  • The screen was re-run independently over the 1,262-symbol summary universe on 10-08 bars. It returns the same 30 names plus none extra. The 31st row, WRT1V.HE, is a stale-bar member (latent finding L1 below).
• AI infrastructure. All 34 typed prose figures match the 10-02 close: TSM RSI 71.5, P RSI 78.2 and 30D +51.62%, NVDA RSI 62.9 and +16.6% versus its 200-day, ARM +5.18% Friday and 1Y +102.1%, AMZN, PLTR, TAC, "10 of 11", and so on. The one mismatch is structural (D-row proposal 4).
• Monster. 2,814 table and list cells plus the header counts.
  • The header's 437 rows equal its tier counts (32/63/49/180/113). The overbought list matches tiers 1–3 exactly. No row violates its tier bounds.
  • 47 cells were rounding boundaries.
  • ATRO's six secular cells only match once its 2026-06-15 ×1.253 split is folded (latent finding L2).
  • The 30 unverifiable cells are crypto bars re-stamped since publication (LINK, SOL, AVAX and BTC: 16 cells, all ≤0.8 pt) and history-basis revisions since the 10-06 provider switch (SMH 5Y/3Y/alpha, DHT, NMM, HCC, IMOS, BITQ, TTMI, P and POWL: 13 cells, ≤0.75 pt except SMH's dividend-basis 5Y gap of 14.7 pt). The last one is the 144-watchlist count.
• Desk calls.
  • LITE: the 10-02 note, the 20-day $941.73, the 50-day $879.03 and the to-date +1.65% all match. The $1.225B guide is in the August 11 release, which is not on disk.
  • CRDO: 33.3% operating margin on $1.335B FY26 revenue (10-K 2026-06-15). The ~$255.50 reclaim and "flagging under its 20-day" match 07-07, the data date. The ~$160 200-day, the 09-14 flag figures, the 10-02 note, the 30- and 90-day returns and excess, and SMA200 179.2113 all match.
  • STX: the note, the returns, and the four margin values match. The label does not.

Every MISMATCH, with evidence

1. company-profiles/2026-09-28-MPWR.md:158: Q3 2025 diluted EPS "$3.65". EDGAR EarningsPerShareDiluted for 2025-07-01..2025-09-30 is $3.74 (restated, 10-K filed 2026-02-27; the original 10-Q said $3.71). Massive eps_diluted is 3.74. The gap is −$0.09 (−2.4%). Mechanism: 179.766M ÷ 49.26M (the Q2 2026 diluted share count) = 3.649. The cell is derived, not reported, and the line under the table says "Q3 2025 is the restated value published in the 2025 Form 10-K". SHIPPED. It has been carried since the 2026-08-22 edition, through 6 archived editions.

2. …MPWR.md:159: Q4 2025 diluted EPS "$3.48". The filed FY figure of $12.86 minus the 9-month $9.37 gives $3.49 (Massive also 3.49). Same mechanism: 171.656M ÷ 49.26M = 3.485. Q1 and Q2 2026 happen to coincide with the filed values (3.92 and 5.22). SHIPPED.

3. …MPWR.md:158: Q3 2025 gross margin "55.0%". Filed GrossProfit / revenue = 406.228 / 737.176 = 55.1%. The gap is 0.1 pt, which is immaterial; it belongs to the same table as the EPS cells. (Q4's "55.1%" against 55.1506% is a rounding boundary and counts as a match.) SHIPPED (immaterial).

4. …MPWR.md:34: Target "The June 22 close area ($1,563.70)". The June 22 bar closed $1,537.88. $1,563.70 is the June 18 close (June 19 was the Juneteenth holiday). The 2026-06-22 refresh-history row records the latest settled close at refresh time (see latent finding L3), and the matrix writer read that refresh date as the session date. The line was introduced in the 09-23 edition. SHIPPED (low stakes: the $1,560 target is near either value).

5. …MPWR.md:32: Action Matrix "reviewed 2026-09-28": "Just off an 18% week, RSI 60". RSI on 09-28 was 58.3; 60.1 was the 09-23 reading, and the 09-28 7D was +5.8%. The 09-23 matrix was copied forward under a new review date. SHIPPED (minor; stale value presented as current).

6. company-profiles/2026-09-28-POWI.md:35: Target "Near the August 8 close ($64.59)". August 8 2026 is a Saturday. $64.59 is the August 7 close. The value is right and the date label is wrong; same mechanism as item 4. SHIPPED (label only).

7. …POWI.md:212: the 09-18 row says POWI is "the one name in the tracked AI-power group with no twelve-month excess return". On 2026-09-18 bars, 20 of the 54 ai-power watchlist names had negative 1Y alpha against SPY: CRWV −48.0, PSIX −71.8, LEU −59.7, AOSL −28.8 (a power semi), HUBB −13.5, ETN −0.6 and others. ETN is also a key_ticker of the profile's own ai-power-delivery campaign. What was true is the narrower statement on line 169: VICR (+298.7), MPWR (+17.7) and NVTS (+54.8) had positive alpha and POWI (−4.44) did not. The row extends a three-peer comparison into a claim about the whole group (subset-as-universe), and the profile builds its reading on it ("a broken long trend rather than a deep pullback inside a good one"). SHIPPED. It appeared in the 09-18 edition, the 09-23 row refers back to it ("retiring last week's 'only name with no excess return' line"), and it is still printed in the 09-28 edition.

8. scans/2026-10-02-ai-infrastructure.md:30,72,80–90,109–113. The page's own prose dates the data to 10-02, but its tables pull current numbers:

   • Line 72 says "Percentages in the tables are … changes ending at the 2026-10-02 close", and line 30 says "Settled close 2026-10-02 for all 11 names".
   • The Price Table and Entry Zones are 104 live sum:… binding bindings with no sum:SYM.latestDate binding label.
   • The shipped projection research/api-data/events/2026-10-02-ai-infrastructure.json (generated 2026-10-09T20:44Z, also in feed.json and the ticker feeds) renders the 10-09 values.

   As a result, the same page reads:

   • TSM $453.31 | 52.4 | … | 🔴 Extended, while its prose says "TSMC joined the extended band at RSI 71.5" and its Entry Zones say "Just crossed 70".
   • P RSI 81.2 against 78.2 in the prose.
   • NVDA $229.28, RSI 53, against the 10-02 values of $233.95 and 62.9.

   This is the D176/D177 invariant ("no column labelled with a past date renders from a live binding") recurring in the scan writer. Five sibling 10-02 scans have the same shape: bindings, a dated as-of sentence, and no latestDate. They are airlines, ai-scan, chemicals, market-pulse and travel-leisure. Tonight's 10-09 ai-scan, ai-infrastructure and market-pulse repeat it; they were not audited because they are still being written. SHIPPED.

9. research/api-data/desk-calls.json, STX 2026-07-05 (open buy). The call says "nearline-HDD margins compounding 23%→32% over four consecutive quarters". From EDGAR, Seagate's total-company GAAP operating margin was:

   Quarter	Operating income	Revenue	Margin
   FQ4'25	568	2,444	23.2%
   FQ1'26	694	2,629	26.4%
   FQ2'26	843	2,825	29.8%
   FQ3'26	998	3,112	32.1%

   The values match; the basis does not. Seagate reports no nearline margin, and the call's invalidation ("nearline margin progression stalls … two consecutive quarters") is keyed to a metric that cannot be computed. The source issue's table at research/issues/2026-07-05-memory-hardware-supercycle-industry-report.md:65 labels the column correctly ("Operating margin (last 4Q)"). The basis was lost when the row was condensed into the call text at :21 and :106. SHIPPED (low; unlabeled-basis).

Proposed DEFECTS.md rows (shipped; file BEFORE any fix)

• A — MPWR profile prints derived EPS as filed diluted EPS (items 1–3).
  • Inject: quarterly NI ÷ the latest quarter's diluted shares, printed under "Diluted EPS" beside a "restated value published in the 2025 Form 10-K" claim.
  • Scream: each "Diluted EPS" cell equals the period's filed EarningsPerShareDiluted (EDGAR or Massive), or carries a "derived" label. $3.65 against $3.74 trips it.
  • Caught by: blind re-derivation spot check, loop-34, 2026-10-09. No gate flagged it across 7 editions.
  • Class: formula-substitution.
• B — profile Action Matrix carries prior-refresh readings under a new review date and reads refresh dates as session dates (items 4–6).
  • Inject: a refresh-history row (date = refresh day, price = latest settled close at refresh), and the 09-23 matrix copied under "reviewed 2026-09-28".
  • Scream: "the close ($X)" equals the bar close on that date; a cell under "reviewed D" uses D's reading or carries its own date.
  • Caught by: spot check.
  • Class: stale-reference / window-mislabel (date form).
• C — POWI "the one name in the tracked AI-power group with no twelve-month excess return" (item 7).
  • Inject: a three-peer comparison generalized to the group.
  • Scream: an "only / one name in " claim is computed over that group's membership (the watchlist or the campaign's key_tickers) on the stated date. Here 20 of 54 had negative alpha.
  • Caught by: spot check.
  • Class: subset-as-universe.
• D — 2026-10-02 AI Infrastructure scan renders live bindings under a 10-02 label (item 8; D176/D177 class recurring in scans).
  • Inject: newer summaries into a scan whose bound tables are dated by prose.
  • Scream: a bound table either carries sum:SYM.latestDate binding or is frozen. Prose interpreting it is true against what it renders.
  • Caught by: spot check.
  • Class: stale-reference. The D176/D177 re-fix was per-page; the scan writer's path never got the invariant.
• E — STX desk call labels total-company operating margin as "nearline-HDD margins" (item 9).
  • Inject: a column basis dropped when a table row is condensed into a call.
  • Scream: a call's figure names its reported concept, and its invalidation names a computable metric.
  • Caught by: spot check.
  • Class: unlabeled-basis. If label-only basis errors are not counted as escapes, this one belongs in TASKS-AUDIT instead.

LATENT findings (not reader-wrong values; belong in TASKS-AUDIT)

• L1 — washout 10-08 includes WRT1V.HE on a stale bar. Its row matches its 10-07 bar on all seven cells, but only price and vs-SMA200 say "(as of 10-07)". The 10-08 Helsinki bar, now on disk (yfinance, fetched 2026-10-09T20:41Z), has 30D −0.75%, so the name fails the both-windows-positive screen on the scan's own data date. On settled 10-08 data the scan would show 30 turns, not 31. Class: ragged-universe-max-date (disclosed only partly).
• L2 — python/src/bars.py (and lib/bars.ts) return ATRO's unrescaled 2026-06-15 ×1.253 split step. ATRO's max.json (Schwab history, rewritten 2026-10-09T20:34Z) shows a −21.2% day, 96.00 → 75.69. Only precompute-summaries rescales seams in memory (the summary rows are adjusted). "The ONE reader" hands every study a false crash: on raw bars the 09-30 monster row recomputes to 5Y +380% and vs-SMA200 −8%, where the correct values are +501% and +5%. Class: corporate-action-seam.
• L3 — refresh-history "Price" means the latest settled close at refresh time, but the column does not say so. MPWR's 2026-07-31 row reads "$1,316.18 — Small bounce". That is the 07-30 close. 07-31 itself closed $1,426.03, up 8.35% after earnings. This convention is what produced items 4 and 6.
• L4 — CRDO's buy call is still status: open in /calls. Its 09-14 flag reads "Met as written … Void", and a 10-02 status note still says "OPEN". This is a lifecycle problem, not a wrong number. It is noted because it is the highest-stakes call in the sample.

Verdict + reasoning

Where the errors are. Machine-written scans re-derive clean. In the monster and washout scans, about 3,000 cells are mismatch-free once rounding, crypto re-stamps and the 10-06 basis switch are separated out. Every mismatch is in agent-written prose, an agent-built table, or a label:

• a derived EPS presented as filed (A);
• refresh dates read as session dates, and stale readings carried under a new review date (B);
• a peer comparison turned into a whole-group claim (C);
• live bindings under a dated label, which the D176 invariant never reached on the scan path (D);
• a dropped basis in a desk call (E).

What is not wrong. Two inspected anomalies did not hold up: ATRO's monster row is correct once the split is folded, and POWI's identical $118.9M revenues are real.

Next step. Recommend filing rows A–E in DEFECTS.md before any fix, and the latent findings L1–L4 in TASKS-AUDIT. This note edits neither file.

Sources

• L1 bars: research/market-engine/data/stocks/{MPWR,POWI,CRDO,STX,LITE,ATRO,WRT1V.HE,SPY,…}/ohlc/max.json via research/market-engine/python/src/bars.py.
• Filings: research/market-engine/data/stocks/{MPWR,POWI,CRDO,STX}/edgar/company-facts.json; …/{MPWR,POWI}/massive/{income-statements,balance-sheets,cash-flow-statements}.json.
• Engine definitions mirrored: research/market-engine/cli/chart/precompute-summaries.ts (calculateChangeDetail, computeMultiYearReturns, computePctFrom52wkHigh), research/market-engine/lib/indicators.ts (computeVolumeMetrics), research/market-engine/lib/washout-scan.ts (isWashoutTurn), research/market-engine/lib/forward-return.ts.
• Shipped projections: research/api-data/events/2026-10-02-ai-infrastructure.json, research/api-data/desk-calls.json (generated 2026-10-09T20:44Z).
• Membership: research/market-engine/watchlists/ai-power.json, research/perspectives/ACTIVE.json (ai-power-delivery), research/api-data/calendar.json.