Stock-market dispatch — 2026-08-10

Headless run (Mon). Research output, not advice — brain never trades. Marks are 2026-08-07 closes (Friday), labelled as such per the 08-07 dating fix. 76/76 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 4th day.

Top of mind

Texas gated too — and it did so twenty days after New York, in the largest data-center market in the country. On 2026-08-03 Governor Abbott directed the PUCT and ERCOT to audit every data center in the interconnection queue and halted approvals pending the result: 1,800+ projects, 474 GW, ~90% of it data centers — more than five times ERCOT's record peak demand. ERCOT has already paused its "batch zero" transmission planning study and seeks a good-cause exemption at its 2026-08-20 meeting. Four days ago this wiki recorded EO 62 as "one state, one data point." That disposition is now obsolete: replication happened under the opposite party, on a different legal instrument (grid-reliability authority, not environmental permitting), and with no stated end date — same unbounded defect. state-datacenter-siting-moratorium-risk goes medium → medium-high.

This also answers what was flagged as the book's highest-value open question — does a siting pause destroy load or relocate it? — and the answer is neither. The relocation leg fails twice over: the presumed destination gated itself, and the alternatives are already capacity-short (PJM up to 15 GW short by 2030; ERCOT's large-load queue 63 → 226 GW; MISO building an expedited process outside its own queue). The escape valve is written into both orders and it is off-grid, not out-of-state: purely behind-the-meter on-site generation is exempt, and developers are observed using it.

And that finding falsified a leg of this project's own three-day-old hypothesis. siting-gate-plus-stranded-capacity-to-demand-side-beneficiary-rotation grouped CAT/CMI with PWR as the disfavoured side. Wrong: the exemption makes on-site generation equipment a beneficiary of the gate. The correct split is does this revenue require grid interconnection (PWR: gated) or does it supply the off-grid alternative (BE/GEV/CAT/CMI: favoured). That removes the observation that motivated the page — PWR and CAT lagging together were read as one common cause; under the corrected mechanism they should carry opposite signs. /calibrate logged. The gate leg strengthened and the transmission leg weakened in the same pass.

Strongest-conviction buys

Ranked shortlist (4c). Prices = 2026-08-07 closes, twelvedata. Fundamentals carried forward from prior dispatches (headless budget) except where dated.

Near-misses: LIN (0.78) $489.98 — the highest-conviction chain after MU, but only −10.6% off high, so the valuation gap ranks it out, not the thesis. TSM (0.70) $420.04 — −12.3% off hi, marginal. BLK (0.62) $1,136.39 and DELL (0.62) $453.77 — both within 7% of highs.

⚠ Four of the five ranked names rose 3–8% on Friday, which mechanically narrows the discount that ranks them. MP +7.6% and NOW +6.4% are the two largest single-day moves in the book; treat today's "undervalued vs base" as measured before that move is digested.

Watchlist

Undervalued candidates (today)

• drone-warfare-demand-to-mp-hree-rerate — MP −49.0% off high, still the deepest discount in the book even after +7.6% Friday, on a chain whose Step 2 was confirmed first-party on 08-06 and whose live sector-excess improved to +2.5% (from +0.1%). The chain is beating a de-rated theme.
• agentic-ai-seat-erosion-to-saas-rerate — NOW −35.9% off high, FCF-positive, and the best live sector-excess in the book (+3.3%). The one name where the discount and the realized selection edge point the same way.
• pjm-capacity-prices-to-nuclear-premium — CEG −34.6% off high while guidance was raised, and today's siting research adds a second, independent reason to prefer existing generation over new build.

New chains to investigate (hypothesis-stage)

• shared-heavy-forging-capacity-to-specialty-alloy-melt-rent — ATI / CRS / HWM. Written input-first rather than end-market-first: four causally unrelated demand shocks already in this wiki — commercial-aero build and aftermarket, the munitions rebuild, AI-power turbines, grid transformers — bid for one constrained input, qualified Western specialty-alloy melt and heavy-forging capacity, whose relief is dated 2028–2030 (Ecotitanium 2028, Safran Gennevilliers 2029). The melt owner is long all four demand curves and short none. ⚠ ATI and CRS appear nowhere in this wiki — that is the un-cited gap and it is flagged as such. Its first test is deliberately a kill test: do these end-markets compete for the same qualified lines, or only the same materials category? To graduate: realized price (not utilisation) at the melt tier, and confirmation the capacity is listed rather than captive.

New theses (now active)

• precision-casting-scarcity-to-aftermarket-margin-transfer (mechanism, 4 steps / 9 citations, medium) — the first net-new mechanism in four days, deliberately in a thin cluster. The insight is that the allocation rule, not the shortage, is the mechanism: scarce castings go to Boeing/Airbus OEM lines by contract, so the entire shortfall lands on the higher-margin aftermarket, converting a supply constraint into a margin-destroying mix shift. Realized 2026-08-06 — Honeywell Aerospace cut 2026 guidance $300M (organic growth 7–9% → 4–5%) and fell ~24% intraday to a 52-week low of $150.03 from $203.64; CEO Jim Currier: no relief until 2027. Independently corroborated: GE Aerospace spare-parts delinquency +20% sequential in Q2.

Updated theses (existing active)

• state-datacenter-siting-moratorium-risk — medium → medium-high, as_of→08-10. See Top of mind.
• aerospace-throughput-bottleneck-to-howmet-pricing-power — its own named falsifier test fired. The page asked for "a primary naming castings/forgings as the specific gating bottleneck"; HONA's guidance cut delivered it with a price attached. Step 2 partial → confirmed. ⚠ Step 4 (HWM captures the rent) did NOT move — sold-out capacity, 30%+ share and a $10.00–10.10B guide vs $9.75B consensus are volume facts, and lta-contract-structure-as-price-insulation is this wiki's own record of how LTAs route that rent to the buyer instead. Top /explore-chain candidate.
• inference-demand-to-wafer-scale-advantage — ⚠ counter-evidence, conviction held at medium. SemiAnalysis benchmarks TileRT at 494 tok/s/user FP8 on B200 — 1.9× the best conventional FP4 — achieved in software on standard GPUs, letting providers carve a high-interactivity tier from existing fleets without new silicon. Held rather than cut because it is batch-size-1 only (~160 vs ~240 tok/s/GPU aggregate — it buys latency by surrendering utilisation) on a two-model catalog, and the source is paywalled and unreproduced. New falsifier, pre-announced by the source: the batch-2–8 benchmark. If the latency lead survives off batch-1, this chain should be cut hard.
• btm-onsite-generation-to-bloom-fuelcell-gev-turbine — gains a second, independent forcing function: policy, alongside queue physics. ⚠ The un-cited order-flow gap is unmoved.
• glp1-injectable-supply-chain-bottleneck — constraint has migrated from API to fill-finish, equipment-gated at >24-month sterile-filling-line lead times. No step flip.
• legislative-divergence-base-rate — two clean instances, one measured: the ESA cuts US housing stock ~4% (Tabarrok, 1,200+ listings); and the data-center moratorium wave is a count-vs-effect divergence — bills failing (GA/SD/WI/VA, OH ballot at 70k of 413,488 signatures) while executive and county instruments bind (Indiana: 11 ordinances, 17 moratoriums, 2 bans).

Contradicted / weakened

• siting-gate-plus-stranded-capacity-to-demand-side-beneficiary-rotation — graduation test #3 met, and one of its own legs falsified. See Top of mind. /calibrate logged under narrow-framing/single-event-tracking. Contradiction with btm-onsite-generation-to-bloom-fuelcell-gev-turbine recorded on both pages rather than reconciled.
• 4e gate: 23 live rows, 2 flags <−10% — PWR −12.2% (widened from −11.5%) and CAT −10.4% (new breach, from −8.1%). 0 automatic calibrations, and that is a pass not a skip: both are medium (0.55, 0.52), and the 4e gate covers high/medium-high. The manual calibration filed today is about the reasoning, not the drawdown — explicitly not logged as evidence against the BTM thesis, since four sessions of tape cannot refute a multi-quarter order cycle. Healed: TJX −8.2% → −9.8% (still inside), UUUU −9.3% → −4.6%, MP sector-excess +0.1% → +2.5%. Best live: AMZN +11.2%, DNN +9.3%, NXE +8.5%. Worst cluster: energy-oil −6.2% (5th week).

Open questions worth a human's eye

• Does the behind-the-meter exemption survive the Texas audit? Abbott's audit explicitly verifies "whether data centers provide their own power or depend on ERCOT" — which could favour BTM projects or subject their claims to scrutiny. The directive text supports the first reading; the audit scope leaves the second open. 2026-08-20 (ERCOT good-cause-exemption meeting) is the next dated read, and it now sits under the largest single new forcing function in the book.
• Do aero castings, power turbines and munitions actually share qualified lines? The single most important unrun test in the vault today — it decides whether shared-heavy-forging-capacity-to-specialty-alloy-melt-rent is a chain or four separate stories. Deliberately framed as a kill test.
• Why do PWR and CAT keep moving together when the corrected mechanism says they shouldn't? Four consecutive marks of co-movement is now unexplained rather than supporting. Either the correction is wrong or there is a third common cause.
• The hypothetical inception re-mark has been deferred four consecutive weeks — blocked by three standing data defects (KLAC split factor, CBRS entry_ref, SPCX SPY backfill). Escalating as a human call for the third dispatch running. Note CBRS reports 2026-08-12, which is an opportunity to resolve one of the three.
• What are Almonty's offtake terms? Carried a third day, still not researched (budget went to the siting gate and the aerospace vein). Flagged rather than quietly dropped.

What I looked at

• Run status: headless; steps 0.5–7 in-turn. Price fetch paced 8/min across 10 batches — 76/76 priced, 0 429s, all 08-07 closes. Weekly ledger (4d) NOT due (last mark 08-07, 3 days). Per-concept bull/base/bear narratives carried forward (headless budget); load-bearing marks in the feed + tables.
• Breadth (2a): ai-infrastructure 48% ⚠ over (44 distinct chains, 59 live signals, 7 clusters) → net-new steered to thin verticals. Delivered: materials + industrials ex-AI — the one new mechanism and the one new hypothesis are both there. Zero net-new ai-infrastructure beneficiaries opened; the AI-adjacent work (siting gate, TileRT) was filed as risk and counter-evidence on existing chains, which is the correct way to touch an over-weight cluster.
• Macro buckets: #11 transport/industrials ex-AI, #8 healthcare & demographics — both from the under-covered list, per the 2a steer.
• Ingestion veins (0.5–0.7): podcast 0 — dark a 4th day. Both providers still refuse: AssemblyAI HTTP 400, xAI HTTP 403 — billing states, not code faults; the monthly budget is not the constraint ($1.05 of $50). 17 episodes now backed up, led again by All-In / Saronic (autonomous warships), deferred a 3rd day and directly adjacent to today's munitions finding. A bounded non-diarized Whisper fallback was started on that one episode and abandoned unfinished — recorded in log so the decision is visible. Action for Paul: top up AssemblyAI or add xAI team credits. Earnings 0 — correctly: the whole active Q2 set is already ingested. Feeds 2 (SemiAnalysis ⚠paywalled-partial, Construction Physics ⚠paywalled-partial).
• ⚠ CBRS is a real, unclosed gap. Cerebras's first public-company call (2026-06-23; $191.3M core revenue +92%, a >$20B OpenAI agreement) has never been ingested, and five free hosts returned 403/404/503 today. It is both a live ledger position with an unverified entry_ref and the tradeable in the chain TileRT just pressured. It reports again 2026-08-12.
• ⚠ Latent defect found and repaired: ingest-pending returned a 9-source queue of which 5 were false positives — the 2026-07-29 run wrote a batch log header (ingest | 5 sources …) that the basename detector cannot see, so an unattended run would have re-ingested five sources and duplicated evidence across ~8 pages. Five detector stubs appended; no re-ingestion performed. Standing rule recorded: one header per source, always.
• Sources: 4 promoted + 4 ingested → 1 new mechanism, 2 entities created (honeywell-aerospace new; howmet resolving a wikilink dangling since 06-10), 1 hypothesis drafted, 3 mechanisms touched, 3 concepts updated, 3 questions updated, 1 contradiction recorded (internal, on both pages), 1 calibration.
• prospect-chains (2b): 1 auto-draft (the melt-rent chain). 2 candidates reported, not filed — the county-gate/entitled-land REIT chain (no real-estate vertical in this book yet) and, carried from 08-07, CF's capital-cost generalization. Noted that the latter is a near-sibling of today's melt hypothesis — both are "the owner of already-built, hard-to-replicate capacity earns the rent"; if the melt chain survives they should merge.
• Not done, and named: EARNINGS.md has no HWM entry despite HWM being the named beneficiary of a hypothesis whose only remaining gap is a Howmet price/mix disclosure. Recommend adding HWM (and ATI/CRS if the melt chain survives). The file is user-curated — flagged, not edited.
• Signal feed (7): full active/armed set re-marked to 08-07 closes; as_of re-dated only for rows whose research moved.