Stock-market dispatch — 2026-08-12

Headless run (Wed, pre-open). Research output, not advice — brain never trades. Marks are 2026-08-11 closes (Tuesday, the latest completed session). 113/115 symbols priced, 0 429s. Weekly ledger (4d) not due (last 08-07; next 08-14). Supabase upsert + push are performed by the wrapper, not this run.

Top of mind

The day produced two good chains and the market had already found one of them. Both were filed from first-party evidence; the difference between them is the whole lesson.

The refiner feedstock chain — venezuelan-heavy-sour-return-to-usgc-coker-differential-capture, yesterday's designated first research target — validated cleanly. Venezuelan exports are corroborated above 1M bpd from a second source type (1.07M bpd production in June; US imports +300% since January), differentials have widened (WCS ~$11–11.50 under Brent, ~$4 cheaper than the Q4 average), and it lands on the tape: Valero's Q2 refining operating income was $4.5B against $1.3B a year earlier, with the COO naming the cause — "an improvement in delivered crude costs relative to the benchmarks." Phillips 66 dollarizes it at ~$140M of annual EBITDA per $1 of WTI–WCS widening. Then the mark: VLO +2.85%, PSX +4.10%, MPC +5.03% — and all three closed at or within 0.1% of their 52-week highs. MPC closed exactly at its high. The chain is right and it is fully priced. Filed as a mechanism because it is true and because the next differential turn will need it; not ranked, because there is no discount left.

The BWXT chain is the opposite, and it is today's actual find. BWXT has been a live paper-ledger position since 2026-07-16, sat at −16.6% sector-excess (the book's second-worst), and the wiki held no first-party evidence on the name at all. Its Q2 call, pulled ad-hoc, explains the lag and dates its end: revenue +18% against adjusted EBITDA +7%, and CFO mike-fitzgerald names the wedge — "backlog associated with older pricing arrangements with the customer…fully expect that to be done by the end of 2026." Government-segment margin guidance was raised 19% → 20.5% on the way. BWXT trades −29.8% off its 52-week high, ~8% above its 52-week low. Filed as legacy-priced-backlog-rolloff-to-bwxt-margin-inflection.

Alongside it, a pairing worth more than either half: russ-jevnin says BWXT can get "zirconium tubes, large forgings, whatever we need" — while three days later kim-fields says ATI cut specialty-energy shipments 6% "reflecting our decision to prioritize available production capacity toward higher-value naval nuclear demand." The buyer says it is not short; the supplier says it made that buyer whole by shorting someone else. That is the allocation rent observed from both ends of the same transaction, and it is a materially better-specified version of the mechanism than "heavy forgings are scarce."

Strongest-conviction buys

Ranked shortlist (4c). Prices = 2026-08-11 closes, twelvedata. Fundamentals carried forward except where dated today.

Deliberately not ranked, and it is the point of the day: VLO / PSX / MPC. Best new chain of the run; at 52-week highs after +2.8/+4.1/+5.0%. The ranking is doing its job by refusing to call a name undervalued the day it prints a high.

⚠ CEG (0.58) ranks just outside at $278.36 (+2.93%, −32.6% off hi, live +3.8%) — the largest single-day move in the top tier and the reason it slipped behind BWXT on valuation gap rather than on chain quality.

Watchlist

Undervalued candidates (today)

• drone-warfare-demand-to-mp-hree-rerate — MP −44.9% off high, still the deepest discount in the book, and now the best live sector-excess at +8.5%. Discount and realized selection edge point the same way.
• legacy-priced-backlog-rolloff-to-bwxt-margin-inflection — BWXT −29.8% off high, ~8% above its 52-week low, with a dated margin inflection and a raised guide. New today; the asymmetry is that the market has been pricing a demand problem that the supplier's own contract disclosed as a vintage problem.
• agentic-ai-seat-erosion-to-saas-rerate — NOW −34.5% off high, FCF-positive, live sector-excess +4.1%.

New chains to investigate (hypothesis-stage)

No new hypothesis pages drafted, and — as on 08-11 — that is deliberate. Both candidates the day generated turned out to already exist in the wiki, which is itself the finding: prospect-chains' dedup surfaced that the freight chain (trucking-regulatory-capacity-removal-to-tl-carrier-rate-recovery) and the private-label chain (consumer-trade-down-to-private-label-manufacturer-treehouse) were both already open hypotheses. The budget went into two new mechanisms and two hypothesis strengthenings instead of a third restatement.

• trucking-regulatory-capacity-removal-to-tl-carrier-rate-recovery — one primary confirm from graduating, and the blocker finally moved. Since 2026-07-16 this page has been held not for lack of rate data but for lack of attribution (the LMI's record price print was confounded by the Hormuz fuel shock). Today supplies a discriminator: truck costs +16.0% versus rail +0.7%. A fuel shock raises both modes; it cannot produce a 15.3-point wedge with rail at +0.7%. Plus an independent capacity measure that doesn't rest on Werner's CEO (−5.5% effective capacity in 2025) and a quantified conversion trigger (intermodal 30% below truckload against the 10–15% J.B. Hunt says is needed). Next action: /explore-chain trucking-regulatory-capacity-removal-to-tl-carrier-rate-recovery, targeting a primary for the mode wedge and the KNX/WERN/SNDR Q2 prints, which reported in early August and are still not ingested here.

New theses (now active)

• legacy-priced-backlog-rolloff-to-bwxt-margin-inflection (mechanism, 4 steps / 11 citations, medium) — see Top of mind. Also yields a reusable lens: when a supplier's revenue growth outruns its EBITDA growth with no cost shock disclosed, check contract vintage before concluding pricing power is weak. That is the mirror of lta-contract-structure-as-price-insulation — the structure that insulates a buyer from a spike strands the supplier at stale prices.
• venezuelan-heavy-sour-return-to-usgc-coker-differential-capture (mechanism, 5 steps / 14 citations, medium) — filed spine-distinct from refining-bottleneck-to-refiner-crack-capture: that chain is product-side (crack), this one is input-side (feedstock). ⚠ Flagged on both pages: VLO's Q2 cannot be attributed to either alone, so the ledger must not double-count one quarter across two mechanisms.
• 4 new entities: bwx-technologies, valero, phillips-66, marathon-petroleum — all four were named tickers in the signal feed or the ledger with no entity page. Plus 4 person entities (russ-jevnin, mike-fitzgerald, kim-fields, rob-foster).

Contradicted / weakened

• ⚠ aerospace-casting-scarcity-to-howmet-margin-capture — self-correction, /calibrate #39. Yesterday's headline claim ("+340bps despite a 360bps headwind → ~700bps underlying") is withdrawn: the 360bps metal pass-through dilution is a Forged Wheels segment figure, not consolidated. Forged Wheels is ~12% of revenue, so the consolidated equivalent is 40–45bps, not 360. The EBITDA-dollar-neutrality point survives. The mechanism survives too, on better evidence — the segment distribution of the gain: Engine Products (the constrained casting/forging segment) +470bps on +32% revenue, versus Forged Wheels +30bps. A rising tide expands margin broadly; an allocation rent concentrates it in the constrained segment. The data matches the second by ~15x. Step 4 re-tagged partial.
  • The pattern tag is overconfidence/favored-narrative — the same tag as #38, one day apart, and the repetition is the finding. The new sub-shape: I compounded my own prior-day summary instead of the primary document, and a summary is exactly where a scope qualifier ("in the Forged Wheels segment") gets dropped. Operational rule recorded: when a figure is load-bearing enough to headline a dispatch, re-derive it from the primary before building on it.
• The 2026-08-11 "VLO/PSX/MPC" basket is corrected from MPC's own mouth. Chief Commercial Officer Rick Hessling says there are "better options" than Venezuelan barrels. MPC carries differential beta; only VLO and PSX carry Venezuelan-sourcing alpha.
• 4e gate: 23 live rows, 3 flags < −10% — PWR −12.5% (improved from −12.7%), TJX −11.8% (⚠ worsened from −11.1%), CAT −10.6% (flat). 0 automatic calibrations, a pass not a skip — the gate covers high/medium-high and all three are medium or below. ⚠ Fourth consecutive dispatch in which every breach sits in the medium-or-below tier and the high/med-high tier has zero. CALIBRATION.md's standing worry is that conviction labels have been a poor guide; on live evidence the tiering keeps sorting correctly. Four is still not proof. Best live: AMZN +11.3%, DNN +10.4%, MP(drone) +8.5%, NXE +8.0%. Worst cluster: energy-oil −5.9%, 7th week — pointedly, on the same day the refiners ripped, because the live energy-oil positions are CCJ/CF/NTR, not refiners.

Open questions worth a human's eye

• ⚠ BWXT's ledger position sits under an AI-power chain but its 2026 result is a defense story. The position is booked to ai-capex-to-power-and-materials-cascade (sector XLU); the evidence says naval nuclear drives the year. That is an attribution problem and a cluster-independence problem — it inflates apparent ai-infrastructure exposure with something that isn't. Flagged, not silently re-clustered; the new mechanism is emitted under defense.
• Does Howmet's margin gain survive full decomposition? Now bounded by segment but still not solved — price is not separated from mix or leverage anywhere, and Engine Products' +470bps is itself undecomposed with gas turbines (+38%) the fastest-growing and plausibly richest-mix end market.
• Net heavy-sour availability is not the Venezuelan number. Mexican Maya imports have collapsed to 22,000 b/d, a six-year low — one heavy supplier arrives as another withdraws, and no source nets them out. Sharpest open weakness in the new refining chain.
• Why do PWR and CAT still move together when the corrected mechanism says they shouldn't? PWR +1.5%, CAT +0.7% today; sixth consecutive mark of co-movement. Still unexplained.
• The hypothetical inception re-mark is deferred a sixth week — blocked by three standing data defects (KLAC split factor, CBRS entry_ref, SPCX SPY backfill). Escalating as a human call for the fifth dispatch running. CBRS was expected to report today; no transcript existed pre-open.
• What are Almonty's offtake terms? Carried a fifth day, still not researched. Flagged rather than quietly dropped.

What I looked at

• Run status: headless; steps 0.5–7 in-turn. Price fetch paced 8/min across 15 batches — 113/115 priced, 0 429s, all 08-11 closes. Reconciled priced-vs-requested (the 08-11 defect check) — clean. 2 persistent failures: ASE/BESI (not in the free plan). ⚠ Correction to yesterday's characterisation of GNKG/HXSCL. They were called "malformed legacy symbols in the feed" on 08-11; they are neither malformed nor in the signal feed. HXSCL is SK Hynix's OTC line and GNKG is Ginkgo Bioworks — both real tickers, both suggested-tickers on concept pages (hbm-supply-bottleneck, us-biotech-strategic-technology-offshoring), and both simply unavailable on the Twelve Data free plan (recurring since 2026-07-20, already recorded as such on those pages). HXSCL additionally has a working substitute already in use — the SKHY ADR line, which priced fine today at $154-ish. They were excluded from this run's request set, which is correct behaviour, but the right fix is a substitute (SKHY) or an explicit unpriced marker on the concept page, not deletion.
• Breadth (2a): ai-infrastructure 47% ⚠ over (45 distinct chains, 60 live signals, 7 clusters) → net-new steered to thin verticals. Delivered: both new mechanisms are outside the dominant cluster (defense and energy-oil), and both bucket scans were from the under-covered list. Zero net-new ai-infrastructure chains opened; the one AI-adjacent source landed as evidence on two existing pages, which is the correct way to touch an over-weight cluster.
• Macro buckets: #11 transport & industrials ex-AI, #10 consumer shift — rotated forward from yesterday's #9/#8. ⚠ Compact one-round scans with zero direct fetches (search-result extraction only) — recorded as such, and no step on any page was re-tagged on them.
• Ingestion veins (0.5–0.7): Earnings 2 (BWXT + ATI, both ad-hoc — the Q2 watchlist cycle is genuinely complete, and both of today's best sources were off-watchlist). Feeds 1 (SemiAnalysis Kimi K3, ⚠ paywalled-partial); the other three active feeds had nothing new and undeduped — dedup by canonical URL slug per the standing rule, which correctly caught two Construction Physics posts already in sources/. Podcast 1 — and the vein is no longer fully dark.
• ⚠ Podcast providers still unfunded, 6th day — but the free path is now proven. Both providers re-probed directly: AssemblyAI "Your current account balance is negative", xAI "your newly created team doesn't have any credits." Budget is not the constraint ($1.05 / $50). Rather than log a sixth no-op, this run passed diarize: false in the episode descriptor to route podcast_ingest.py to local Whisper, which works fine here (~1.2× realtime) — one episode transcribed and ingested. A 48-minute All-In episode (Saronic founders; defense/industrials, a thin vertical) was also started and will not finish inside this run; the helper writes atomically, so it either lands in vault/clippings/ for tomorrow's step 3 or leaves nothing. 23 episodes still backed up. Action for Paul: top up AssemblyAI, or fund xAI and switch diarize_provider: grok.
• Sources: 7 promoted + 7 ingested → 2 new mechanisms, 8 entities created (4 companies, 4 people), 2 hypotheses substantially strengthened, 9 mechanisms/concepts updated, 1 calibration (#39), 1 zero-novelty ingest surfaced (the 6-minute Seides essay — a lens, not data; attached to hidden-leverage-beyond-margin-debt).
• prospect-chains (2b): 0 auto-drafted. Both of the run's candidate chains deduped against existing open hypotheses — a healthy signal about wiki coverage rather than a shortfall, and it redirected the budget into strengthening the two pages that already held them.
• Valuation snapshots refreshed: 16 mechanism pages, 08-11 closes. Conviction buys ranked: 5.
• Not done, and named: EARNINGS.md still has no HWM, ATI or BWXT entry — third consecutive day, and today it cost two ad-hoc pulls for the run's two best sources. User-curated, so flagged, not edited. The KNX/WERN/SNDR Q2 prints (early Aug) are also un-ingested and are the freight chain's stated graduation gate.
• Signal feed (7): full active/armed set re-marked to 08-11 closes; as_of re-dated only for rows whose research moved.