Stock-market dispatch — 2026-07-17

Headless run, steps 0.5–7 in-turn. Research output, not advice — brain never trades. Prices are pre-open; marks are the 2026-07-16 close. Supabase sync is performed by the wrapper, not this run.

Top of mind

TSMC's Q2 call handed the book its bottleneck confirmation from the one party with no reason to give it — and the stock fell 2.3% anyway, on a day the whole AI-infra complex broke. C.C. Wei, unprompted, on competitors' packaging: "our packaging capacity is so tight that now it's limiting my customers' growth." Then he went further and welcomed Intel's EMIB-T — "I welcome that the competitor offers some of the flexibility to my customer so that their front-end wafer can be put into the package" — because packaging scarcity is stranding TSMC's larger front-end wafer revenue. That is a monopolist testifying against its own scarcity rent. hbm-cowos-as-binding-bottleneck Step 6 → confirmed. It also inverts the wiki's own 2026-05-19 read that EMIB-T is a CoWoS headwind.

The same call raised capex to $60–64B (from $52–56B in January), added $100B in Arizona — with, in Wei's own words, no schedule — and put FY26 growth "slightly above 40%". The tape's answer: SOXX −4.5%, MU −5.6%, INTC −5.8%, ARM −5.4%, ASTS −17%, while DG +4.8%, ROST +3.1%, UNP +3.8%, IHI +4.7% went green. The two podcasts ingested this morning were titled "The AI Unwind Is Forcing A Historic Market Rotation" and "You're About to See the Real AI Winners Stand Up." They printed it the same session.

And the day's second find defends the book's biggest position. A gap-fill pass found the 07-16 bear leg on MU was mis-scoped: the pull-forward IBM's 8-K documented sits in the non-LTA enterprise-server channel — buyers with no contractual price cap, who therefore have a reason to pre-buy. Micron's HBM book is take-or-pay through CY2027. If your price is contractually capped, pre-buying ahead of a hike is pointless. The bear leg didn't die; it got narrower (MU's non-LTA/consumer bits). Separately, MU's forward P/E was carrying a stale ~11× since June — it is 5.95×.

Strongest-conviction buys

Ranked shortlist (4c). Prices 2026-07-16 close, twelvedata (73/77; ASE/BESI free-plan-gated, HXSCL/GNKG OTC-carry).

⚠ Caveat on #3 — the cap came from inside the call. Wei named his own margin ceiling: "I'm really jealous about memory companies, 86%… About 68%, I would be happy about that" and "I don't want to squeeze them out from the market." Booked as a new machine-checkable falsifier (tsm_gross_margin_pct <= 68), not a conviction cut — it caps upside, not chain-truth.

⚠ Caveat on #1 and #3 — the new bear leg is the relief Wei is recruiting. A bottleneck whose incumbent actively wants competitors to relieve it has a dated erosion path. No prior falsifier was worded to catch this; one was added to both rows today.

Watchlist

Undervalued candidates (today)

• MU $853.20 — second straight session where the chain got stronger as the price fell. −32% from high at 5.95× forward, first-party confirmation from the supplier, and the bear leg the last dispatch feared got scoped out of its book. ⚠ The unmodelled risk is positioning, not thesis: ~89% DRAM earnings-revision breadth — the chain can be entirely right and MU still de-rate.
• AMKR $63.02 — −34.8% from high on the day the monopolist confirmed the bottleneck AMKR exists to absorb. Not in the feed; the honest reason is it has no mechanism page of its own yet.
• MP $45.46 — $1.03 above its 52-week low with the Nov 10 China deadline still ahead and a $110/kg DoD floor underneath. Sector-excess only −2.4%: a theme drawdown, not a selection failure.

New chains to investigate (hypothesis-stage)

• grid-conductor-replacement-to-cable-maker-pricing-power — ATKR / NVT. prospect-chains promoted a lead the wiki left itself in writing: wire-and-cable inflation +150% since 2019 is the steepest line in the T&D bill of materials, and nothing in the vault holds it. Root cluster is ex-AI — Yake (ex-Xcel): "we're looking to replace a significant portion of the infrastructure — even without data centers, even without load growth." To graduate: cite a cable ticker (none is cited anywhere in the wiki today) and separate supplier margin from copper pass-through — link 2 is written as a chain-killer.

New theses (now active)

Four mechanisms filed, none emitted — each fails on the tradeable step, and that is the finding:

• memory-shock-to-oem-value-transfer — IDC: PC units −11.3% but PC revenue +1.6%; smartphone units −12.9%, revenue −0.5%. Units fall, revenue holds: the consumer dollar transfers to the memory supplier. The book already owns the benefit side; the un-owned leg is the hurt side, and the trader is long-only.
• memory-crowd-out-to-trade-in-funnel-scarcity — AIZ. Round 3 of the research reversed the naive read: the secondary market is supply-constrained, not demand-constrained (the same price shock that drives buyers to used devices makes owners hold longer). ⚠ Step 6 stays open — the cited $1.63B is trade-in value returned to consumers, not Assurant revenue. It names a ticker it cannot size.
• packaging-scarcity-to-emib-t-second-source — INTC/TSM. Steps 1–3 confirmed on Wei's own words against interest. Step 5 open: he validates a second source is wanted; nothing shows Intel booking volume.
• agentic-ai-cpu-resurgence-to-tsm-silicon-broadening — steps 1–2 only partial, because Wei's claim here is with interest (he is justifying an $8B capex raise). The same call's "x86, Arm-based, or RISC-V" phrasing weakens vera-cpu-to-arm-datacenter-royalty's ARM-specific step 4 while strengthening its step 1. ARM held at 0.42 — flat is the honest read.

Contradicted / weakened

• drug-pricing-reform-pbm-spread-compression — premise inverted. The spread is no longer where PBM profit is: rebates ~50% → under 15% of revenue mix (2012→2023), specialty 16%→35%, admin fees 5%→22%. Optum's "95% pass-through by end-2026" is a promise about the smallest, fastest-shrinking pool — and Optum is the last mover: Express Scripts went first (Nov 2023), then CVS (2024-25). Calibration filed; pattern: regulation-targets-the-profit-pool-that-already-moved. Sharper forcing function filed instead: state-nadac-floor-reimbursement-mandates (binds the commercial book now, not in 2028-29).
• The 07-16 MU bear leg was mis-scoped (see Top of mind) — narrowed, not deleted.
• ibm-preannounce-to-enterprise-saas-capex-reallocation weakened. The wiki treats the memory line as the whole explanation of IBM's miss; IBM's own letter attributes it partly to Z-systems and cybersecurity, and Power/storage grew 37% — its best in reported history. A pull-forward borrows from the future; a reallocation moves line items inside a period. IBM's quarter shows both.
• CEG cut 0.58 → 0.52 — the one cut, and the only one with a reason the tape can't give. It is flagged below −10% sector on both books (hypothetical −14.9% from inception, live −10.4% from a 06-01 entry). Two different entry dates agreeing excludes entry timing by construction — which is the one thing a single flagged position can never rule out. What's left is thesis doubt.
• INTC ×3 flagged for exit at −21.0% sector — 3rd consecutive widening mark, flagged since 06-18, entire inception gain gone. Not downgraded and not re-dated: the thesis is not falsified, and a flag is not evidence. Same rule that protected MU through a −32% drawdown; it has to cut both ways. The question has moved from "is the chain right" to "how long does a correct chain get to lag its sector by 21% before the position is the problem." That's a human call.

Open questions worth a human's eye

• ⚠ The calibration gate has never fired, and now we know why. Seven consecutive marks: every live sector-excess flag < −10% lands in medium (today PWR −13.6%, TJX −13.4%, CEG −10.4%), and step 4e only reads high/med-high. Meanwhile the by-tier live data went monotone — high +4.0% (n=1, not evidence), med-high −0.7% (n=8, 50% hit), medium −2.9% (n=14, 29% hit). The labels rank correctly; the monitor is aimed where the failures aren't. Calibration filed. The live question: is medium a size-small tier or a don't-size tier? n=14 in one drawdown regime, all long into the same unwind, is suggestive — not settled.
• ⚠ Two tooling defects, one of which silently destroyed data for a week. (1) The 07-10 ledger mark corrupted 11 of its own position blocks — a regex wrote a literal \g<1>, deleting each position's id, mechanism, entry_ref and falsifier. Recovered from commit 96f9a2a; the numbers were fine but for a week the ledger couldn't say what half its positions were long or why. (2) _lib/causal_extract.py::_parse_evidence_sections doesn't match the schema's status-tagged ### Step N · confirmed: form — a test against a copy of hbm-cowos destroyed 44 of 68 evidence bullets. It is create-only-safe today. Both filed in log; neither is fixed.
• Is medium where the wiki hides its inference? Three separate chains today died on the same step: AIZ (can't size it), MCK/COR (terminal link is the researcher's own inference, with live counter-evidence), EMIB-T (nothing shows Intel booking volume). All three had strong, cited forcing functions. The gap is never the cause — it's always the last link to the ticker.
• Odd Lots filter miss: "NY Governor Kathy Hochul on Her One Year Data Center Moratorium" was filtered out — no keyword matched. A state moratorium on datacenter buildout is squarely a forcing function for the book's largest cluster. The filter is user-curated; the skill never edits it. Worth a data center / moratorium keyword.
• vera-cpu-to-arm-datacenter-royalty page/feed disagree — the page says conviction: low-medium + status/hypothesis; the feed emits active at 0.42. Pre-existing.

What I looked at

• Run status: headless; steps 0.5–7 in-turn; price fetch paced 8/min to completion — 73/77, 0 rate errors (api.twelvedata.com reachable; the 4 failures are the known ASE/BESI plan-gating + HXSCL/GNKG OTC).
• Breadth (2a): ai-infrastructure 53% ⚠ over. Net-new steered to thin verticals: healthcare (bucket 8 — PBM), consumer shift (bucket 10 — device price shock). Post-emit breadth holds 53% — 3 of the 4 new mechanisms were honestly ai-infrastructure-clustered, and non-emitting them is what kept the number from going to ~57%.
• Ingestion veins (0.5–0.7): podcast 3 diarized ($0.73; budget $10.65/$50 July); earnings 2 (TSM Q2 full; UNH Q2 flagged partial — the host returned a condensed extract, so only quoted fragments are citable. ISRG also reported 07-16 — deferred to the next run, honestly, for run-time budget); feed 1 of 4 (SemiAnalysis/Fabricated Knowledge/Apricitas fully deduped).
• Steps 1–2: 3 autoresearch (memory pull-forward vs level-shift; PBM pass-through; consumer device price shock). 2b prospect-chains: 1 auto-drafted of 10 candidates — 7 reported-not-drafted, 2 rejected (one for having no citable causal verb; writing it would have been invention).
• Sources: 9 promoted (7 → stock-market, 2 multi-context → vault/sources/), 9 ingested → 4 new mechanisms, 12 concepts, 3 company entities, 10 person entities, 34 pages updated. Zero-novelty: 0.
• Valuation (4b): 21 concept snapshots re-marked to 07-16; 11 had material bull/base/bear changes; 10 got a first snapshot. Cleaned malformed "Last refreshed" stamps on 8 pages, and corrected a systematic "near 52w high" carried on AMAT/LRCX/KLAC/FCX — all are 19–29% off their highs.
• PAPER-LEDGER weekly mark: DONE (due; last 07-10). 16 scored, 7 unpriced (data debt is now the book's biggest weakness — KLAC unscoreable for 7 marks, CBRS for 4; SPCX hid a +19.2% → −18.5% round-trip behind a missing benchmark). Scorecard: per-position exSector −3.0% (from −0.5%), per-instrument −0.4%, hit 50%. ⚠ Not apples-to-apples — n went 11→16 as NVDA/NOW re-priced. One name is the whole negative print: INTC ×3 at −21.0% sector is −3.9pts of the −3.0% mean. Ex-INTC the book is +1.2%, and the ordering is emphatic: high+med-high +7.7% vs medium −6.4%.
• Live feedback (4e): 24 rows, 3 flags < −10% sector (PWR, TJX, CEG), all medium → 0 auto-calibrations (see Open questions). Best: DNN +9.4%, AMZN +7.0%, STVN +5.4%, NXE +5.1%.
• Calibrations: 2 (the PBM premise inversion; the mis-specified 4e gate — pattern: monitoring-gate-aimed-where-the-failures-are-not). 9 entries now — /reflect-on-me is overdue (past the ~5 clustering threshold for two runs running).
• Signal feed (7): full active/armed set, 48 rows, validated 48/48 against the v3 contract. 47 re-marked to 07-16 closes (SSNLF left on its 07-15 OTC carry rather than falsely re-dated). as_of re-dated on only 4 — MU + TSM (step 6 → confirmed), ARM (one source moved two steps in opposite directions), CEG (conviction) — the other 44 carry forward, which is what lets the trader see staleness. +0 / −0 rows. New falsifiers added on MU + TSM for the EMIB-T packaging-relief path. Deliberately not emitted (auditable, not silent): all 4 of today's new mechanisms — AIZ (below gate, tradeable step open), the OEM value-transfer (long-only book can't express the hurt side; the long leg double-counts MU and would contradict the existing long DELL row), EMIB-T (tradeable step open), CPU-resurgence (tradeable step open; TSM already held). Also flagged for a human: 5 pre-existing rows now sit below the 0.4 gate (NVDA 0.32, ASTS/RKLB 0.34, DG/ROST 0.39) — retiring them is a reconcile --prune call this run was told not to make. Supabase sync performed by the wrapper, not this run.