Stock-market dispatch — 2026-07-10

Headless Pi run (06-30→07-09 gap). Steps 0.5–7 completed synchronously in-turn. Research output, not advice — brain never trades. Supabase sync is performed by the wrapper, not this run.

Top of mind

The tape hurt the book, the research didn't — and the most valuable thing this run produced is a caught mistake. MU (−11.4%) and INTC (−14.0%) gave back most of the June gain in 11 days, dragging per-position inception sector-excess from +5.5% to −0.5%. Nothing moved against either chain. MU in fact picked up three independent corroborations this week: gavin-baker from the investor seat ("there are three companies that can make it… as close to magic as science can get"), winston-cheng (Lenovo's CFO) from the buyer's seat ("it does take almost two to three years for a new fab to come online… [shortage] will continue probably for at least two to three years"), and the SCA floor "ahead of prior cycle peaks." So MU's drawdown fired no calibration — price is not evidence. But applying that rule symmetrically exposed an error: on 06-29 this project retired an INTC exit-flag on one favourable mark, writing "the lag has healed… confirmed." It has since widened to −15.0% sector, worse than when first flagged. That's logged (one-good-mark-mistaken-for-thesis-vindication). Meanwhile Brian Potter's coal-plant piece contradicted a sub-path the wiki had granted confirmed status by inheritance — coal-to-gas conversion is not the AI-power bridge ("the heyday of coal-to-gas conversions is behind us"), and grid-scale batteries are an unmodelled competitor in that chain. Net-new: nvidia-gpu-backstop-to-neocloud-financeability (NVDA lends its AA rating as a take-or-pay backstop; ~18% take rate; a >$7T AI-debt market by 2029) and, pointed at the thinnest cluster per step 2a, two healthcare/ag chains.

Strongest-conviction buys

Ranked shortlist (4c). Prices 2026-07-10 twelvedata (67/71; ASE/BESI 403, GNKG/HXSCL carry).

INTC drops off the shortlist (was #3 on 06-29). Not on a thesis change — on the recognition that its exit-flag was retired prematurely. It stays a medium hold, not a buy, pending the 18A yield read.

Watchlist

¹ since the 06-29 dispatch.

Undervalued candidates (today)

• MU $991.64 — the drawdown is price, not evidence; the chain strengthened this week on two independent source types. The single cleanest dislocation in the book.
• CCJ $95.74 — −29% from high; power-cascade confirmed-grade, critical-minerals the best live cluster (+2.3%).
• MP $51.71 — −48% from high; REE breadth-thin leg, live sector-excess ~flat (+0.2%).

New chains to investigate (hypothesis-stage)

• screwworm-herd-rebuild-deferral-to-animal-health-rerate — ZTS/ELAN. Re-established screwworm (34 confirmed US infestations) + an 86.2M-head herd (lowest since 1951) → prophylaxis becomes routine → structural parasiticide spend. Eradication is a "close to a decade" job. Graduate on: any ZTS/ELAN quantified screwworm revenue — the gap the research explicitly refused to close.
• orexin-franchise-to-lly-glp1-playbook-repeat — LLY. $6.3B Centessa buy reads as the GLP-1 playbook repeating. Filed hypothesis with the biology step tagged open: "you sleep less but you become less inflamed" is two technologists' enthusiasm with zero trial data. Graduate on a Lilly disclosure or peer-reviewed readout, nothing less.
• german-drug-pricing-to-section-301-retaliation — LLY/EU pharma. Germany prices drugs fiscally → Lilly pulls investment → USTR 301 (June 18) → Germany reportedly softening. Every step partial/open; the USTR docket is a one-lookup verification.

New theses (now active)

• nvidia-gpu-backstop-to-neocloud-financeability — NVDA / CRWV. Financing is the new binding constraint; Nvidia substitutes its AA/Aa2 credit for the neocloud's via a six-year take-or-pay backstop, collapsing ~400bp of funding cost (CoreWeave 5y unsecured ~10% vs 5.9% on the Meta-backstopped DDTL) and taking ~18% of above-backstop revenue. Steps 4–5 partial — all deals are APAC, forward-projected, and the take rate is the author's model. The neocloud equity is a call option: at backstop levels project IRR is "zero or slightly negative."
• enterprise-token-budgeting (concept) — the demand-side load test under every AI-capex chain. Whales carry the revenue and aren't cutting. Recorded caveat that matters more than the finding: this and the backstop piece are the same publisher four days apart — two halves of one house view, not independent corroboration.

Updated theses (existing active)

• hbm-cowos-as-binding-bottleneck — strengthened on two new source types (Baker: investor; Cheng: buyer). Baker also narrowed the CXMT falsifier: cheap consumer-grade Chinese DRAM does not falsify the AI-server HBM oligopoly; only qualified CXMT HBM share does. Conviction held 0.80 through an 11% drawdown.
• ai-capex-to-power-and-materials-cascade — Steps 1–3 corroborated from an integrator's seat (transformers, optical connectors, grid power on Lenovo's own shortage list) — but the coal-to-gas bridge sub-path is contradicted. See below.
• agentic-ai-seat-erosion-to-saas-rerate — Step 2 explicitly re-tagged partial. The gap-fill succeeded by refusing to close the gap.

Contradicted / weakened

• ai-capex-to-power-and-materials-cascade, coal-to-gas sub-path — CONTRADICTED. Potter: "It doesn't seem likely that we'll see many more of these coal-to-gas conversions"; grid-scale batteries erode the peaker logic. The sub-path had inherited confirmed from its parent forcing function without ever being evidenced. Calibration: sub-path-inherited-parent-confidence.
• The 06-29 INTC "lag has healed" conclusion — CONTRADICTED. Now −15.0% sector, wider than when first flagged. Calibration: one-good-mark-mistaken-for-thesis-vindication.
• Agentic-AI→NRR causal arrow — REFINED, not confirmed. Cross-sectional gap real (95% vs 108–130%); causal attribution unproven; the 2021→2024 decline predates agents. Calibration: endpoints-evidenced-middle-assumed.
• Live flags rose 1 → 3 (TJX −11.5%, PWR −10.1%, CEG −10.1% sector). All three are medium conviction, so per 4e none auto-fired a calibration — but all three live flags now sit in the medium tier, which is a systematic read, not a coincidence.

Open questions worth a human's eye

• Is the medium conviction label doing real work? It carries all three live flags and −10.5% inception sector-excess, while high/med-high stay positive. Is "medium" functioning as "I want to hold this but can't defend it"? Worth a /reflect-on-me pass.
• KLAC has been excluded from scoring for six consecutive marks (unadjusted split, entry_ref 1811.35 vs $229.52). Fix the entry_ref or retire the position — carrying it indefinitely quietly biases the semicap cohort.
• Capital Allocators: fourth consecutive zero-novelty ingest, ~$0.65 of diarization spent. DAILY 0.5 says pause it. The skill can't edit PODCASTS.md; a human must.
• Does the Nvidia backstop structure travel outside Asia-Pacific? Every named deal is APAC — a financing quirk or a market restructuring?

What I looked at

• Run status: headless Pi (06-30→07-09 gap); steps 0.5–7 completed synchronously in-turn; price fetch paced 8/min, 67/71 + 8 ETFs, 0 rate errors.
• Macro buckets: healthcare/demographics + consumer/ag (the two thinnest verticals), steered by 2a — ai-infrastructure 50% ⚠ over (target ≤35%).
• Ingestion veins (0.5–0.7): podcast 10 transcripts (all diarized, $1.28); earnings 0 (no tracked ticker reported in the window — quarter-end cluster passed); feed 9 posts (SemiAnalysis ×6, Construction Physics ×2, Fabricated Knowledge ×1).
• Step-1: 1 gap-fill autoresearch (SaaS NRR causal arrow) — returned "still a gap," which is the honest result. Step 2: 1 autoresearch (screwworm × cattle herd). prospect-chains: 1 auto-draft (screwworm→animal-health), 1 reported-not-drafted (PPC poultry substitution, ready-to-graduate).
• Sources promoted 21 (19 stock-market, 1 threads/politics, 1 multi-context); ingested 21 → 10 new entity pages (6 company: LLY/ZTS/ELAN/TSN/lenovo/baidu · 4 person: Cheng/He/Sheehan/Schmidt-biotech); 3 new mechanisms; 1 new concept; 1 new hypothesis; 4 zero-novelty (Anthropic-paywall, Compound/BOT, Capital Allocators, and coal-to-gas as a ticker source — though it was the run's best falsifier).
• Valuation: 11 concept snapshots refreshed (twelvedata).
• PAPER-LEDGER (4d): weekly mark DONE 07-10 — per-position sector-excess +5.5% → −0.5% (INTC + MU), conviction ordering HOLDS for a 7th mark (med-high +3.1% > high +1.4% > medium −10.5%); KLAC still excluded.
• Live feedback (4e): performance-2026-07-10 (22 pos) — 3 flags < −10% sector, all medium-conviction → 0 auto-calibrations; energy-oil worst cluster (−3.3%) two weeks running; DNN live-best (+11.9%).
• Calibrations: 3 (coal-to-gas contradicted; INTC flag-retirement contradicted; SaaS NRR refined). Conviction buys ranked: 5. New hypothesis chains: 3.
• Signal feed: full active/armed set, gated ≥0.4 — see step 7 (Supabase sync performed by the wrapper, not this run).