Stock-market dispatch — 2026-08-17

Headless run (Mon, pre-open). Research output, not advice — brain never trades. Marks are 2026-08-14 closes (Friday, the latest completed session). 119/123 symbols priced, 1 recovered 429, 4 persistent failures. Weekly ledger (4d) not due — ran 08-14, next 08-21. Supabase upsert + push are performed by the wrapper, not this run.

Top of mind

A hypothesis this project filed on Friday named a public dataset as its own falsification test. The dataset had already printed. It says no.

driver-not-equipment-scarcity-to-class-8-oem-derate argued that a driver-constrained truckload market inverts the truck cycle — carriers shrink fleets, Class-8 OEMs derate into a rate upcycle. Its stated bar: "ACT/FTR Class-8 net orders falling while TL contract rates rise, with the 2027 emissions pre-buy ruled out." July 2026 net orders were 22,100 (ACT) / 22,000 (FTR) — up 68–75% year over year, with calendar-2026 production "essentially sold out" and the −30% sequential print caused by "a lack of 2026 build slots", i.e. an OEM supply constraint, not weak demand. And the pre-buy is not ruled out — both forecasters name it, and Werner's CEO calls the EPA change "more relief than a delay."

Filed as /calibrate #42, reusing process/unverified-denominator rather than minting a variant, because it is the same defect as #41 one step earlier in the pipeline: the check that would have caught it had already been identified — it just wasn't run. #41 was validating a claim over a filtered population without checking the filter; this is naming the decisive dataset in a falsification bar and not querying it before filing. Rule recorded: if a hypothesis's graduation bar names a specific, already-published dataset, query it before filing — the bar is not a to-do list for future-me.

The premise survived and got stronger; only the conclusion died. RXO — a broker, whose costs rise with truckload rates — states "the primary constraint is driver availability rather than equipment." What was wrong is the inference that a driver constraint shrinks fleets. The observed behaviour is replace, don't expand: Werner cut fleet-growth guidance to 16–18% from 23–28% while raising net capex to $215–250M, targeting fleet age "closer to mid-2s"; Schneider cut capex on "a lower need for trailing equipment." Replacement-heavy, trailer-light capex is consistent with a driver shortage and +68% tractor orders simultaneously. Deliberately not closed — one month's print, in a month when build slots set the number, is thin ground to kill a chain, and closing on one datapoint would mirror the opposite error corrected on 08-13/14. Priority medium → low; the surviving idea relocated one layer down the channel (below).

Strongest-conviction buys

Ranked shortlist (4c). Prices = 2026-08-14 closes, twelvedata. Fundamentals carried forward except where dated today.

⚠ WERN's promotion to a cited source cuts both ways and the tension is worth stating. Its call corroborates the chain — but it also discloses "low to mid-single-digit" contract renewals against Schneider's "double digits" on the same quarter. Two carriers, one market, a two-to-threefold gap in realized price, and no source explains it. Step 4's magnitude currently rests on Schneider's number alone. That is a new open question, and it sits underneath the #5 rank.

Deliberately not ranked, for the fourth consecutive day: VLO / PSX / MPC. All three still within ~1.3% of 52-week highs (VLO $341.67, PSX $233.61, MPC $355.42). Zero valuation gap. The ranking refusing to chase remains the ranking working.

⚠ BWXT (0.55) at $173.22 (+1.7%, −28.4% off hi) remains the deepest discount in the tier after MP and NOW.

Watchlist

Undervalued candidates (today)

• drone-warfare-demand-to-mp-hree-rerate — MP −41.4% off high and live sector-excess +12.4%, both the best in the book. ⚠ But it rose 5.5% on Friday: the discount is closing on the position with the strongest realized edge, exactly the setup the freight complex presented last week.
• agentic-ai-seat-erosion-to-saas-rerate — NOW −36.3% off high, its deepest discount to date, and it fell 2.6% while the chain's evidence was unchanged. FCF-positive, live sector-excess +1.3%.
• legacy-priced-backlog-rolloff-to-bwxt-margin-inflection — BWXT −28.4% off high on a dated end-2026 margin inflection.
• driver-supply-removal-to-truckload-contract-rate-inflection — WERN −18.1%; KNX has cheapened back to −12.6% after three sessions of rallying. The window the 08-14 dispatch called "closing" reopened slightly.

New chains to investigate (hypothesis-stage)

• driver-constrained-replacement-cycle-to-truck-dealer-aftermarket-capture ⭐ new — RUSHA (Rush Enterprises). The surviving half of the falsified OEM short, relocated one layer down the channel. If a driver constraint makes carriers replace rather than expand, the value may shift from OEM unit volume to the dealer / parts / service layer — new-truck retail, used-truck remarketing off the trade-in flow, and aftermarket service on an installed base management has just said is older than it wants. Anchored on a confirmed forcing function; step 5 flagged ⚠ unverified — RUSHA appears nowhere in this vault before today. Most likely failure: if OEM build slots are the scarce asset, the dealer is volume-capped and the rent stays upstream. What it would take to graduate: RUSHA segment margin history showing parts-and-service counter-cyclicality, plus evidence dealers aren't slot-capped. → /explore-chain
• ⚠ Still no hypothesis page for the healthcare find — second consecutive day, and now for a stronger reason. See below.

New theses (now active)

• None. One hypothesis drafted (above), zero graduations. Today's research produced two negative results of high value and one refinement — that is the honest output.
• 4 new entities: werner-enterprises (WERN), derek-leathers, chris-wikoff, and rush-enterprises — the last filed with zero supporting evidence in sources/, explicitly, to hold the new hypothesis honestly rather than to assert exposure.

Updated theses (existing active)

• driver-supply-removal-to-truckload-contract-rate-inflection — second-carrier corroboration, an independent non-carrier confirmation, and a provenance defect that matters more than either. Werner's Q2 confirms the attrition ("playing out as predicted"; "long-haul truckload employment has dropped to below pre-COVID levels"); RXO — a net loser from truckload inflation, and now the second such party after CHRW — confirms the constraint is drivers, not equipment, with tender rejections at their highest since 2022 in the slowest quarter. But RXO also states "The FMCSA estimates the non-domiciled CDL rule will push 194,000 drivers out of the market." This project's own govinfo primary check (2026-07-16) established that the rule contains no driver-supply economic analysis and that 194,000 is J.B. Hunt's model. An interested carrier's estimate has been laundered into a regulator's, and is now circulating in a third venue. Step 2 stays partial; the laundering is an argument for more caution, not less.
• pjm-capacity-prices-to-nuclear-premium — a serious modeling challenge lands, and the chain is REFINED, not falsified. SemiAnalysis argues PJM's 2024 Reserve Requirement Study understates winter gas capacity (cold-air uplift 8.4–25%; ~400 of ~700 units winterized post-Elliott), inflating the demand curve and overcharging ratepayers ~$11.6B across two auctions; E3 is reviewing the methodology. It does not reach the revenue line inside the collar window: the 2028/29 auction cleared at the FERC $325/MW-day cap while PJM's own uncapped estimate was $554.72 ($16.4B vs ~$30B) — a ~$230/MW-day buffer a 3.8 GW correction must eat through before CEG's realized price moves. CEG cleared 18,875 MW (15,700 MW nuclear) at the capped price. ⚠ But this page had been conflating the capped realized price with the uncapped shadow price, and step 1's magnitude is now formally marked ambiguous. New dated risk: collar expiry after 2029/30, when a corrected model and a free price would compound. Corroborating the chain from the other side: only 525 MW of new generation cleared, and 220 GW applied when the interconnection window reopened in April 2026 while "neither has energized a single megawatt." Conviction held medium.
• sole-source-concentration-in-generic-sterile-injectables — the decisive question is ANSWERED, and the answer is "not tradeable." Brookings: GPO contracts "commit manufacturers to a price but rarely carry a minimum quantity agreement", and hospitals "can and do buy off contract." Price capped on the upside, volume unguaranteed on the downside — so the shortage clears as allocation, not price, and nobody has an incentive to build the redundancy that would end it. Sterile injectables in shortage average ~8.5× cheaper than those not in shortage: the drugs that go short are the cheap ones. The proposed remedy — $2B in subsidised loans plus a public buffer inventory — is itself the diagnosis; if price could clear it, policy would let it. The tradeable version, if any, is whoever receives the $2B, and that has no bill and no date.

Contradicted / weakened

• /calibrate #42 — process/unverified-denominator (reused, not re-minted). See Top of mind. driver-not-equipment-scarcity-to-class-8-oem-derate → priority: low, held at hypothesis, deliberately not closed.
• ⚠ UNRESOLVED CONTRADICTION, surfaced not reconciled — biosecure-cdmo-scarcity-to-western-cdmo-pricing-power. That active chain asserts Western CDMO pricing power. Today's healthcare pass measures the only available price series going the other way: "Contract fill-finish pricing for monoclonal antibodies decreased by 15% from 2023 to 2025", attributed to overcapacity — while Catalent, Delpharm and PCI add $150M/$140M/$365M of Western capacity into it. The two may be compatible (BIOSECURE is prospective; −15% is realized history) but the page currently reads as though the pricing power is already in evidence. No conviction or status changed unilaterally — this is a human call, and the run was headless. It also explains why the sterile-injectables concept got no ticker for a second day: the CDMO relocation was attempted and failed on its own numbers.
• 4e gate: 23 live rows, 2 flags < −10% — TJX −13.1% (⚠ worsened a fourth consecutive mark, from −12.8%) and PWR −11.5% (worsened from −10.6%). CAT −9.5% still cleared. 0 automatic calibrations fired. The 08-14 instrument repair is confirmed working: iran-fuel-shock-consumer-bifurcation now carries status: active / conviction: low, so it is visible to the gate and correctly below threshold — the failure mode has moved from "structurally invisible" to "seen and ranked," which is the outcome the fix was for. Best live: MP(drone) +12.4%, DNN +9.2%, NXE +7.3%, AMZN +5.6%, STVN/CF +3.5%. Worst cluster: energy-oil −6.1%, 10th week — ⚠ and for the fourth consecutive week the label mismatch is noted without being fixed: the live energy-oil rows are CCJ/CF/NTR, not the refiners, whose chain is the cluster's best and sits at 52-week highs. Best cluster: critical-minerals +5.7%.

Open questions worth a human's eye

• Why do Werner and Schneider report contract renewals two-to-three-fold apart on the same quarter in the same market? WERN "low to mid-single-digit" vs SNDR "double digits." Step 4 of the book's best-evidenced new chain currently rests on Schneider's number alone. New, and the highest-value item here.
• Should biosecure-cdmo-scarcity-to-western-cdmo-pricing-power be downgraded on a −15% realized price series, or is the segment mismatch (mAb fill-finish vs generic small-molecule) doing the work? A genuine adjudication, flagged not decided.
• The energy-oil cluster label is measuring the wrong names — fourth week flagged, still unfixed. It corrupts the trader's cluster-level risk budgeting and now also its P&L attribution.
• Does the PJM collar get made permanent? A published $11.6B modeling-artifact claim, four days before the 2026-08-21 FERC comment deadline, is exactly the input that entrenches a cap on merchant generator revenue. Watch that docket.
• Why do PWR and CAT still move together when the corrected mechanism says they shouldn't? PWR +1.9%, CAT +0.2% Friday — divergent for once, but ninth consecutive mark under observation.
• The hypothetical inception re-mark is deferred a ninth week — blocked by three standing data defects (KLAC split factor, CBRS entry_ref, SPCX SPY backfill). Escalating as a human call for the eighth dispatch running.

What I looked at

• Run status: headless; steps 0.5–7 in-turn. Price fetch paced 8/min across 16 batches — 119/123 priced, 1 × 429 (recovered on retry after a 65s back-off; the batch boundary between two shell calls dropped the pacing gap), all 08-14 closes. ⚠ 4 persistent failures: ASE, BESI (not on the Twelve Data free plan) and GNKG, HXSCL (symbol not found). ASE/BESI still matter: cowos-packaging-capacity-crunch names both and neither can be priced, so that snapshot stays structurally incomplete — fifth day flagged.
• ⚠ A defect I introduced and fixed inside this run, recorded rather than hidden. The first pass of the 4b snapshot refresh rewrote every valuation table with a 5-column layout, which mis-filled the 24 pages whose tables carry 7 columns (day-% landed in the "52w range" column, and so on). Caught on spot-check, repaired from git HEAD, and those 24 pages now carry a dated mark blockquote above an untouched table instead — the convention several of them already used. No page was left with transposed data, but the near-miss is the reason the spot-check step exists.
• Breadth (2a): ai-infrastructure 44% ⚠ over (48 distinct chains, 66 live signals, 8 clusters) → net-new steered to thin verticals. Delivered: the one new hypothesis is transport/industrials-ex-AI; the two macro buckets were healthcare and transport — both target verticals. Zero net-new ai-infrastructure chains opened, fourth consecutive day.
• Macro buckets: #11 transport, industrials & logistics ex-AI and #8 healthcare & demographics. Both produced filable material; both produced negative results, which is the honest characterization.
• Research passes: 3 (2 bucket scans + 1 gap-fill on the PJM chain). Below the 5-call step-1 cap, deliberately — the day's inbound earnings and feed material was dense enough that a fourth query would have padded rather than added.
• Ingestion veins (0.5–0.7): Earnings 2 — SNDR Q2 and WERN Q2, both pulled ad-hoc; the watchlist proper is dry (every active ticker's Q2 is ingested; DE and WMT report 08-20, NVDA late Aug, MU/COST late Sep). Feeds 2 — SemiAnalysis (08-16, the PJM piece) and Construction Physics (08-14, ROAD Act Part II). Podcast 1 — see below.
• ⚠ The podcast vein is working again. AssemblyAI is funded (budget $7.76 of $50 for August after this run) and the 08-15 backlog cleared — the eight-day funding block reported through 08-14 is resolved. Only one new episode qualified today (Capital Allocators, "Best of Pre-Mortem Analysis: Gary Klein", diarized, 7 speakers, $0.15) and it was transcribed but deliberately NOT promoted: it is a decision-process rerun with no causal chain to a tradeable, matching the treatment of the two Annie Duke "Best of Decision-Making" episodes that were transcribed and never promoted. It sits in clippings/.
• Feeds needing a human call (user-curated — flagged, not edited): Fabricated Knowledge is fully paywalled for a third consecutive post (08-13, a 157-char stub) — candidate for paused. Apricitas Economics has published nothing since 2026-05-03, i.e. 3.5 months — the feed is not broken, the writer has stopped. Both are Active. Meanwhile EARNINGS.md still has no HWM, ATI, BWXT (sixth consecutive day) and no SNDR/KNX/WERN/CHRW despite the freight names supplying today's two best sources, both pulled ad-hoc.
• Sources: 7 promoted + 7 ingested → 4 new entities, 1 new hypothesis, 0 new mechanisms (3 extended), 1 hypothesis contradicted, 1 contradiction surfaced unresolved, 1 calibration (#42).
• prospect-chains (2b): 1 auto-drafted (driver-constrained-replacement-cycle-to-truck-dealer-aftermarket-capture, move: second-order beneficiary). One candidate routed to an existing chain as a falsifier rather than drafted (the PJM capped-vs-uncapped ambiguity); two reported without drafting; one dropped as a duplicate.
• Valuation snapshots refreshed: 40 pages (16 table rewrites + 24 dated mark blockquotes), all to 08-14 closes.
• Weekly ledger (4d): not due (ran 08-14; next 08-21). Hypothetical inception re-mark deferred a 6th consecutive week, not fabricated.
• Signal feed (7): full active/armed set re-marked to 08-14 closes; as_of re-dated only for rows whose research moved today.