Stock-market dispatch — 2026-08-11

Headless run (Tue, pre-open). Research output, not advice — brain never trades. Marks are 2026-08-10 closes (Monday, the latest completed session). 110/114 symbols priced, 0 429s. Weekly ledger (4d) not due (last 08-07). Supabase upsert + push are performed by the wrapper, not this run. Podcast vein dark a 5th day — now with the cause decoded.

Top of mind

Two hypotheses set their own kill tests; both fired within 24 hours of being written; one graduated and the other was corrected. This is the most useful day the wiki has had, and not because the news was big.

aerospace-throughput-bottleneck-to-howmet-pricing-power had demanded, since 2026-06-10, one specific thing: a Howmet disclosure that splits price from volume. HWM reported 2026-08-06 — a call absent from EARNINGS.md and pulled ad-hoc today. CFO patrick-winterlich: "Higher metal pass-through diluted margins by approximately 360 basis points year-over-year, but had no material impact on EBITDA dollars." So the +340bps of reported EBITDA-margin expansion was achieved despite a 360bps pass-through headwind — roughly 700bps underlying, on a 46% incremental margin. Step 3 → confirmed, step 4 ⚠ unverified → partial, and the hypothesis graduated into aerospace-casting-scarcity-to-howmet-margin-capture (5 steps, 12 citations).

The other test failed, and that is the more valuable result. shared-heavy-forging-capacity-to-specialty-alloy-melt-rent — opened yesterday — set its bar as "realized price (not utilisation) at the melt tier." Carpenter's CFO: "The decline in average selling price per pound is due to higher proportion of lower-priced products in the mix." ASP fell, while CRS's SAO hit a record 37.8% operating margin. The rent is real; the mechanism was wrong. ATI's CEO states the actual one: capacity allocated "more heavily weighted towards the margin and EBITDA line more so than the revenue line." It is an allocation rent, not a price rent, and it sits one tier downstream of where the page put it. CRS dropped from the ticker list; ranking inverts to HWM first, ATI second. /calibrate #38 logged under overconfidence/favored-narrative — I let an elegant, portable story pick the tier before any disclosure said where the money was.

The shared-facility test that page called "the single most important thing to check first" also fired, from john-plant: missile programs are "actually competing for the space in our Virginia facility."

Strongest-conviction buys

Ranked shortlist (4c). Prices = 2026-08-10 closes, twelvedata. Fundamentals carried forward except where dated today.

Deliberately not ranked, and the reason is the interesting part: HWM. Today's strongest research result is the Howmet graduation — but HWM trades at $283.71, only −8.5% off its 52-week high, the narrowest discount of any name in the book's top tier. The chain got materially stronger and the valuation gap ranks it out, which is the ranking working as designed rather than a contradiction. Same for ATI ($228.19, −4.6% off hi) and CRS ($559.98, −10.5%) — the melt/forge names are the least discounted part of a thesis that just improved. LIN (0.78) $492.46 at −10.2% off hi is the same story for the third day.

⚠ MP has now risen +6.9% on top of +7.6% — roughly +15% in two sessions. Its discount is closing fast; treat "undervalued vs base" as measured before that move is digested.

Watchlist

Undervalued candidates (today)

• drone-warfare-demand-to-mp-hree-rerate — MP −45.5% off high, still the deepest discount in the book, but +15% across two sessions has narrowed it from −49.0%. The chain is beating a de-rated theme (live sector-excess +5.1%).
• agentic-ai-seat-erosion-to-saas-rerate — NOW −34.6% off high, FCF-positive, +2.7% live sector-excess. The one name where the discount and the realized selection edge still point the same way.
• pjm-capacity-prices-to-nuclear-premium — CEG −34.5% off high while guidance was raised; +2.0% live sector-excess.

New chains to investigate (hypothesis-stage)

None drafted today, and that is deliberate. prospect-chains surfaced 4 candidates and none cleared the auto-draft bar; the 3–5/day hypothesis budget was already absorbed by one new mechanism, one new concept and two major hypothesis revisions. Padding the cap would have been the wrong call. Reported instead:

• Venezuelan heavy-sour return → US Gulf Coast complex refiners (VLO/PSX/MPC). Exports >1M bpd for a third straight month, up from ~0.5M in December. Spine-distinct from refining-bottleneck-to-refiner-crack-capture (feedstock differential, not crack capacity). Blocked only on a single-sourced, partial forcing function — the strongest candidate of the day and tomorrow's first autoresearch target.
• ATI's $1B five-year naval-nuclear contract (~double the prior) as an independent read on bwxt — filed as a falsifier-check, not a thesis: BWXT is the book's second-worst live position (−16.6%), and a supplier independently confirming naval-nuclear demand doubling says the lag is not demand-driven.
• Turbine secondary market as an asset class; Chinese power-module bypass beneficiaries — both dropped for want of a liquid tradeable.

New theses (now active)

• aerospace-casting-scarcity-to-howmet-margin-capture (mechanism, 5 steps / 12 citations, medium) — aerospace-throughput-bottleneck-to-howmet-pricing-power graduating after 62 days. Filed new rather than merged into precision-casting-scarcity-to-aftermarket-margin-transfer: same forcing function, opposite sign — that chain routes to the aftermarket victim, this one to the capacity-owning beneficiary. The wiki now holds one allocation rule observed from both ends in the same week.
• constraint-bypass-by-capitalized-buyers (concept, medium) — the day's genuinely new idea, and it is a risk lens on this book's own premise: nearly every chain here assumes the buyer cannot route around the bottleneck. SemiAnalysis documents SpaceX routing around three at once — LPTs ("buy power modules from China," medium-voltage architecture), GEV turbines (30+ alternates, a surging secondary market), and permitting ("1km across the border"). Its useful output is a discriminator: qualification-gated chains (casting, HBM/CoWoS) are durable; capacity-gated ones are capped at the bypass cost.

Updated theses (existing active)

• shared-heavy-forging-capacity-to-specialty-alloy-melt-rent — corrected, not confirmed. See Top of mind. Test table added; CRS dropped; "four demand shocks" downgraded to an honest two-and-a-half (grid transformers uncorroborated).
• howmet · ati-inc ⭐ new · carpenter-technology ⭐ new — the two entities the 08-10 page flagged as absent from this wiki now exist and carry Q2 evidence.
• pwr-transformer-moat-to-eps-doubling — new second falsification path on step 1: demand routing around LPTs rather than lead times compressing. ⚠ Conviction held — the bypass is a behind-the-meter topology, not PWR's grid-interconnected backlog.
• ai-power-gap-to-genset-bridge-power — the turbine backlog is supplier-specific, not sector-wide. ⚠ Conviction held: a secondary market clearing at "very high" prices confirms scarcity; what changes is who captures the rent.
• btm-onsite-generation-to-bloom-fuelcell-gev-turbine — first measured on-site build rate in the wiki (Southaven 27 turbines/~495MW Feb → 69/1.7GW Jul). ⚠ Steps 1–3 strengthen; 4–5 weaken — no supplier named, GEV mentioned only as the thing being routed around, BE not mentioned at all.
• state-datacenter-siting-moratorium-risk — ⚠ tension with yesterday's own conclusion, refined rather than reversed: interconnection gates don't relocate load; permitting gates do. Both gates currently in force are interconnection gates, which is why the off-grid response dominates in those two cases.
• spr-releases-as-loans-not-sales — independently corroborated from a different source type; premium quantified at "up to 24%", 35–40M barrels of premium oil due back, and the political mechanism the wiki lacked: no cash changes hands, so no congressional approval is needed.
• naic-rating-scrutiny-to-pe-insurer-flywheel-derate — the catalyst window is today (08-11→08-14), and the new fact cuts against the chain: the Discretion Amendment took effect 2026-01-01 but is still not operational, RBC changes unfinalized. Recorded as an information event, not an action event.
• legislative-divergence-base-rate — a new and unusually clean instance; proposes a third category beyond introduced ≠ passed: commenced ≠ operable.
• glp1-injectable-supply-chain-bottleneck — fill-finish corroborated from a second source type; first named supply response (Sharp doubling Lee MA; CordenPharma two isolator lines). ⚠ No step flip — trade press, one item sponsored.

Contradicted / weakened

• shared-heavy-forging-capacity-to-specialty-alloy-melt-rent's price leg — falsified on its own stated bar. /calibrate #38, overconfidence/favored-narrative. Recorded on the page: the kill test was well-designed and worked within one day; because the un-cited leg was flagged in bold, nothing downstream had been built on ATI/CRS, no signal was emitted, and no position taken. The process caught this, not hindsight.
• 4e gate: 23 live rows, 3 flags < −10% — PWR −12.7% (from −12.2%), TJX −11.1% (⚠ new breach, from −9.8%), CAT −10.6% (from −10.4%). 0 automatic calibrations, and that is a pass not a skip — the gate covers high/medium-high and all three are medium or below. ⚠ Worth stating: for the third consecutive dispatch 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-trading evidence the tiering is currently sorting correctly. Three dispatches is not proof — re-check, don't assume. Best live: AMZN +13.3%, DNN +10.5%, NXE +7.3%. Worst cluster: energy-oil −6.0% (6th week).

Open questions worth a human's eye

• Does Howmet's margin gain survive decomposition? Spares grew +37% vs +24% and are now ~22% of revenue; aftermarket is higher-margin, and 21% organic growth expands margin through leverage alone. Price, mix and leverage remain unseparated — this is now the single number that both aerospace pages turn on, and the partial source means the analyst Q&A that would force it is missing.
• Does the allocation rent survive the 2028 capacity? Every relief date found today is 2028–2030 (HWM "earliest August of 2028"; CRS brownfield early FY2028; IGT adds 2028–2030). An allocation rent should compress faster than a price rent once capacity lands, because allocation power evaporates the moment buyers stop being rationed. Sharpest falsifier on the new mechanism.
• Why do PWR and CAT still move together when the corrected mechanism says they shouldn't? Five consecutive marks of co-movement, both widening again today. Still unexplained.
• The hypothetical inception re-mark is deferred a fifth week — blocked by three standing data defects (KLAC split factor, CBRS entry_ref, SPCX SPY backfill). Escalating as a human call for the fourth dispatch running. CBRS reports tomorrow (2026-08-12), confirmed — an opportunity to resolve one of the three.
• What are Almonty's offtake terms? Carried a fourth 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 — 110/114 priced, 0 429s, all 08-10 closes. ⚠ Batch 9 hit a 429 and was silently skipped on the first pass — and it contained MU and NOW, the two top-ranked names. Caught by reconciling priced-vs-requested and re-fetched 8/8. The failure mode is silent; recorded so the next run checks the count, not the log. 4 persistent failures: ASE/BESI (not in free plan), GNKG/HXSCL (malformed legacy symbols in the feed — worth cleaning).
• Breadth (2a): ai-infrastructure 47% ⚠ over (45 distinct chains, 60 live signals, 7 clusters) → net-new steered to thin verticals. Delivered: materials + industrials ex-AI (the new mechanism, both new entities) and financials + healthcare (the bucket scan). Zero net-new ai-infrastructure beneficiaries opened — the AI-adjacent work landed as a risk concept and as counter-evidence on three existing chains, which is the correct way to touch an over-weight cluster.
• Macro buckets: #9 financials & rate regime, #8 healthcare & demographics — both from the under-covered list, per the 2a steer. Compact one-round scans, recorded as such rather than presented as deep passes.
• Ingestion veins (0.5–0.7): podcast 0 — dark a 5th day, 19 episodes backed up. ⚠ The cause is now decoded rather than inferred: the AssemblyAI key authenticates fine (200), and reproducing the submit call returned the real body — "Your current account balance is negative." xAI: "your newly created team doesn't have any credits." Defect found and fixed: _lib/diarize.py printed only the HTTP status line and discarded the response body, which is why five days of logs said "HTTP Error 400: Bad Request" instead of the actionable reason. Now surfaces the body; verified. Budget is not the constraint ($1.05/$50). Action for Paul: top up AssemblyAI, or add xAI team credits and switch diarize_provider: to grok. Earnings 1 (HWM, ad-hoc — the day's best source; the watchlist itself is correctly complete for Q2). Feeds 3 (SemiAnalysis SpaceX ⚠paywalled-partial, Doomberg ⚠paywalled-partial, + 1 discarded — see below); Mule's Musings feed failed DNS.
• ⚠ Dedup miss, caught by the basename check. A SemiAnalysis post was filed, then found to have been already ingested on 08-07 — my feed scan deduped on publication date rather than on basename against sources/. clipping-promote's vault-wide basename check caught it before promotion; it was discarded. Without that check the source would have been re-ingested and evidence double-counted across ~4 pages. Second dedup-class defect in two days — standing rule recorded in log: dedup on canonical URL/basename, never on date.
• Sources: 5 promoted + 5 ingested → 1 new mechanism, 1 new concept, 4 entities created (ati-inc, carpenter-technology, john-plant, patrick-winterlich), 2 hypotheses substantially rewritten, 6 mechanisms/concepts updated, 1 calibration, 0 contradictions needing adjudication.
• prospect-chains (2b): 0 auto-drafted (bar not met), 4 reported. Recorded as a legitimate outcome, not a shortfall — the day's generative budget went into a graduation and a correction.
• Not done, and named: EARNINGS.md still has no HWM entry (today's best source was pulled ad-hoc) and no ATI entry now that it is a cited name. The file is user-curated — flagged, not edited, for the second day running.
• Signal feed (7): full active/armed set re-marked to 08-10 closes; as_of re-dated only for rows whose research moved.