Calibration
Beliefs updated by evidence — what was thought before, what changed it, and what is thought now. Over time the pattern tags reveal where intuition runs over- or under-confident.
CEG/TMI restart target 2027 — PJM interconnection delay risk to 2031
Crane Clean Energy Center (Three Mile Island Unit 1 restart, 835 MW) has a near-term 2027 online target secured by Microsoft's 20-year PPA. The wiki treated the 2027 date as the primary concrete near-term nuclear-for-AI catalyst — the thesis's most credible near-term revenue event.
PJM interconnection is a load-bearing gating constraint, not a formality. Constellation Energy Q1 2026 commentary flagged that the PJM interconnection process has potential to delay the online date from 2027 to 2031 — a 4-year slip. If 2031 materializes, the Microsoft PPA (20-year agreement, delivery starting from restart date) is disrupted and the $1.6B restart investment does not produce cash flows for 5+ years. The "near-term catalyst" framing collapses; CEG re-rates away from nuclear-pure-play premium toward integration-risk utility. The thesis itself (nuclear as only viable AI baseload) remains structurally correct, but this specific near-term catalyst may not be as near-term as believed.
2026-05-29-autoresearch-nuclear-ai-datacenter-ppa-smr-may-29 — Constellation Energy Q1 2026 commentary cited PJM interconnection process delay risk, with potential to slip the Crane/TMI online date from 2027 to 2031.
correctly identified nuclear restarts as the near-term path but didn't stress-test the interconnection queue as a binding constraint. Same class of error as the helium-repair timeline entry (`underestimating-physical-supply-chain-lead-times`): assumed regulatory/process timelines would be proportionate to urgency and capital commitment, not dominated by queue position in an overwhelmed process.
Section 232 Phase 2 "July 1 gate": implementation trigger → Commerce report submission date
July 1, 2026 was the trigger date for Section 232 Phase 2 semiconductor tariffs — the date on which broader tariffs would be implemented or activated. Multiple research entries carried the "July 1 Phase 2 gate" framing as though it were an implementation event.
Per the White House Presidential Proclamation (verbatim): "By July 1, 2026, the Secretary shall provide the President with an update on the market for semiconductors used in United States data centers, so that the President may determine whether it is appropriate to modify the tariff." July 1 = Commerce report submission date only. Phase 2 tariffs activate only after a separate Presidential determination following that report — no automatic Phase 2 on July 1. Financial media characterizations of "Phase 2 on July 1" are a common mischaracterization. The actual gate structure is: July 1 report → Presidential determination → tariff modification (if any).
2026-05-28-autoresearch-us-industrial-policy-tariffs-may-28 — verbatim Presidential Proclamation language cited; notes financial media commonly collapses reporting milestones and implementation triggers.
carried forward simplified financial media framing ("July 1 is the Phase 2 gate") without checking the actual proclamation. Regulatory deadlines in financial media often collapse multi-stage processes (report → determination → implementation) into a single "trigger date."
Ras Laffan helium infrastructure recovery: political gate → 3-5 year physical-supply-chain gate
Ras Laffan helium infrastructure would recover within 6-12 months post-ceasefire, restoring global helium supply relatively quickly once the geopolitical situation resolved. The primary gating factor was viewed as political will / ceasefire agreement.
Recovery is NOT primarily gated by geopolitics. QatarEnergy confirmed a 3-5 year repair timeline due to the global gas turbine shortage: replacement turbines have 2-4 year delivery queues. Physical damage repair extends to 2029+ regardless of when a ceasefire occurs. "Ceasefire is NOT the end" — the supply restriction persists through turbine supply chains, not political resolution. The helium thesis is now a multi-year structural supply deficit, not a geopolitical-resolution binary. Conviction upgraded from medium-high to HIGH.
2026-05-28-autoresearch-helium-supply-crisis-ras-laffan-may-2026 — QatarEnergy confirmed global gas turbine shortage → 2-4 year delivery queues → 3-5 year total repair timeline. JP Morgan upgrade of LIN with explicit helium framing.
assumed repair was primarily gated by human decisions (ceasefire, political will) rather than physical equipment supply chains (gas turbine manufacturing lead times). Physical capital equipment lead times are systematically longer than intuition suggests and are not shortened by urgency or capital spending alone — same pattern as transformer lead times in the datacenter thesis.
CUDA moat erosion at inference — Nvidia actively defending the moat by neutralizing Groq
The CUDA moat at inference is eroding: two of three frontier models (Gemini/TPU, Claude/Trainium) have exited the CUDA ecosystem; moving a model from GPU to Cerebras takes "10 keystrokes"; the inference layer is CUDA-irrelevant. The wiki concept cuda-moat-erosion-at-inference was filed at medium conviction on Andrew Feldman's articulation. The investment implication was that Nvidia's software-lock-in premium was at risk from inference alternatives.
Nvidia's Groq acquisition ($20B acqui-hire, December 2025) is direct counterevidence. Groq's LPU was the most credible CUDA-alternative for inference — confirmed "genuinely faster and more energy-efficient for certain workloads" by Senators Warren and Blumenthal (March 20, 2026). By acquiring and neutralizing Groq, Nvidia has actively defended the moat that Feldman claimed was eroding. This doesn't mean the erosion thesis is false for the long run (TPU, Trainium, wafer-scale are all real), but it does mean Nvidia is not passively watching its moat erode — it is spending $20B to shore it up. The specific Groq threat is now eliminated. The AMD-Meta $60B deal is the remaining primary non-CUDA inference pathway, but AMD ROCm is not yet at Groq's LPU level for low-latency inference. Conviction revised medium → low-medium: still a real tension, but Nvidia's active defense reduces the near-term re-rate risk.
2026-05-27-autoresearch-regulatory-antitrust-semis-ai-may-2026 — Nvidia-Groq deal structure (December 2025); senator confirmation that Groq's LPU was genuinely superior for inference workloads.
once a thesis is filed in the wiki (CUDA erosion), it accumulates evidence on the thesis side but is slower to absorb counterevidence that doesn't fit the narrative. The Groq acquisition was a December 2025 event; it took until the May 2026 antitrust scan to surface it as direct counterevidence. Counterevidence that requires searching outside the thesis's usual sources takes longer to integrate.
Goldman Sachs copper surplus forecast is now an isolated outlier — ICSG flipped to deficit
The copper structural thesis faced a genuine near-term contradiction: Goldman Sachs forecasting a 490K tonne surplus while the wiki's structural thesis rested on a deficit narrative. I logged this as a calibration candidate on May 18, downgrading copper conviction from medium-high to medium on the basis that two independent near-term sources (Goldman, S&P Global January "elevated prices overextended") cut against the structural thesis for 2026 specifically.
Goldman Sachs is now the isolated outlier. The ICSG (International Copper Study Group, the authoritative international body) revised to a 150K tonne deficit for 2026 — the first ICSG deficit since 2009. LME spot hit $6.65/lb on May 13, a record. Multiple analysts (JPM, Citi, BMO) cluster around deficit or near-balance. FCX Grasberg ramp confirmed started end-March. Goldman's revised estimate (still 160K tonne surplus) now represents a ~310K tonne gap with ICSG. ICSG's methodology is considered more comprehensive than Goldman's near-term model. Conviction upgraded medium → medium-high. The structural thesis is substantially confirmed for 2026.
2026-05-27-autoresearch-copper-supercycle-may-27-2026 — ICSG deficit projection, LME record $6.65/lb (May 13), Goldman now isolated.
gave Goldman's near-term surplus call roughly equal weight to ICSG's structural methodology because Goldman is more prominent. The right prior for "Goldman disagrees with ICSG on copper balances" is probably "ICSG is right more often," given the respective institutional remits.
INTC "undervalued" framing is no longer accurate after +478% YTD run
INTC represented the clearest undervalued position in the stock-market thesis universe — the anchor-customer thesis (AWS, Microsoft confirmed; Apple preliminary; Terafab $25B) was visible but not yet priced in. The re-rate event lay ahead. "Higher conviction than TSMC on re-rate potential."
At ~$119–120 (May 24), +478% YTD and ATH $132 (May 11), INTC is now trading 14–35% above BofA's post-Apple-deal PT of $96, which itself was raised from $56 after explicitly modeling the Apple deal as adding $10B/year IFS revenue by 2030. The market has already priced a substantial IFS re-rate. The thesis chain is intact, but the "undervalued" framing is not — INTC is now at-or-above base-case fair value by the most comprehensive sell-side model available, and trading significantly above it. The correct framing is: "thesis intact, but the easy money is made; remaining upside requires bull-case execution (Apple deal finalizing with volume, 14A securing 2+ major customers by year-end, yield curve reaching industry-standard by 2027)."
2026-05-25-autoresearch-intel-18a-yield-ifs-status-may-2026 surface the valuation context: BofA PT $96 (post-deal), stock consistently trading 14–35% above throughout May; +478% YTD; ATH $132 May 11 followed by pullback to $115 May 18, then recovery to $119–120 May 24. The Apple deal remains preliminary (no orders, no specified volumes yet). IFS external revenue is $174M/quarter — still tiny. The gap between "deal announced" and "revenue recognized" is multi-year; the stock has moved as if the revenue is already in the bag.
thesis calls that work well can generate a follow-on error: the re-rate *event* occurs, and the bull framing continues as if the thesis hasn't been priced. Need to re-baseline "undervalued vs base case" after a stock's +4x move. The corrective is to ask "if BofA models the thesis fully and still has a PT 20% below current, am I seeing something the sell-side isn't, or is the market over-pricing?"
FCX Q1 2026 unit cost: management's $2.60/lb guidance vs $1.91/lb actual
FCX's Q1 2026 unit cost would be approximately $2.60/lb — the company's own guidance, which I treated as the expected-value estimate.
FCX's gold by-product credits act as a meaningful downside buffer in quarters with favorable ore grades. The $2.60/lb figure was a conservative ceiling, not an expected value. Actual Q1 2026 unit cost came in at $1.91/lb — a ~$0.69/lb beat driven by gold credits and favorable Morenci ore sequencing. The correct mental model: FCX guidance is issued before by-product credit realization and therefore systematically skews conservative when gold is elevated.
2026-05-20-autoresearch-fcx-grasberg-q2-2026-restart-copper-lme — Q1 2026 results: $1.91/lb actual vs $2.60/lb guidance, with explicit attribution to gold by-product credits and favorable ore sequencing at Morenci.
accepted management guidance as a point estimate without adjusting for a systematic by-product-credit effect that FCX has historically delivered. The anchoring was to the guidance number specifically, not to an independent cost model.
Friedberg's "99% confidence" El Niño framing was accurate
gave outsized weight to a credentialed domain expert's expressed confidence without triangulating against official forecasting agencies. Expert confidence is a prior, not a fact.
Copper structural deficit thesis is a multi-year story, not a 2026 price catalyst
A 304K tonne structural copper deficit in 2026, combined with AI data centers adding 500K tonnes/yr by 2030, means copper prices are in a supercycle. FCX and SCCO are primary beneficiaries with medium-high conviction.
The long-run structural thesis remains intact (mine supply genuinely cannot respond to 2030+ demand in time), but the near-term is not necessarily a 2026 price catalyst. Goldman Sachs (May 2026) forecasts continued global surplus and prices capped below $11,000/ton; S&P Global (January 2026): "elevated prices overextended." FCX's Grasberg Q1 2026 incident (-27% YoY volume, cost spiking from $1.40/lb to $2.60/lb) adds idiosyncratic operational risk. Conviction downgraded to medium; FCX specifically weakened by Grasberg to medium (from medium-high); SCCO holds medium-high on cleaner cost profile.
2026-05-18-autoresearch-copper-supercycle-may-2026-update surfaced Goldman Sachs (May 2026) forecast of continued surplus + price cap below $11,000/ton, and S&P Global January 2026 "elevated prices overextended" framing — direct sell-side contradictions of the structural deficit narrative for the near-term horizon.
drawn to the structural-deficit framing because it is intellectually elegant (mine timelines genuinely long, AI demand signal real). Underweighted near-term supply-demand balance data from sell-side analysts who have more granular copper balance visibility than bottom-up macro research.
Mainstream news gives a roughly accurate picture of state-program fraud scale
Mainstream news coverage plus state-government statements give a roughly accurate picture of the scale of state-program fraud and of what is being covered up. If a major fraud story were as large as occasional viral citizen-journalist claims suggest, mainstream investigative reporting would have already led on it.
Mainstream coverage materially underweights the scale of state-program fraud, often pivots to process-criticism of citizen journalists (e.g., Nick Shirley) rather than engaging with the substance they surface, and is unreliable on politically-charged factual questions where coverage incentives are weakest. Primary documents (court filings, indictments, OIG reports, state audits) plus on-the-ground citizen-journalism (video of the thing happening, walked-into-the-empty-building, visible-luxury-cars-in-parking-lot) are stronger evidence than either mainstream summaries or alt-source commentary on contested claims. The right posture for politically-charged factual questions is triangulation across primary documents + mainstream + alt sources, with bias calibration applied to all three.
the 2026-05-13 autoresearch pass on MN/CA state-program fraud (filed as 2026-05-13-autoresearch-recent-fraud-minnesota-california-hospice-daycare), which surfaced a gap between concrete federal prosecutions ($300–400M across Feeding Our Future, autism EIDBI, Housing Stabilization) and the US Attorney's aggregate estimate ($9B); and the synthesis's own evidentiary tilt — it leaned on mainstream sources (CBS, KFF, CalMatters, LAist) and dismissed Nick Shirley primarily through the CBS-debunking framing of his MN daycare claims rather than engaging with the broader on-the-ground evidence pattern he represents. Reading the synthesis after the fact made the institutional-trust bias visible.
(none assigned — see whether future entries cluster with this one before tagging)
The Intel alternative-foundry thesis was bilateral; Samsung is a credible third
The leading-edge foundry world is effectively bilateral for AI customers: TSMC (capped) or Intel (re-rating). Apple's options were: wait for TSMC capacity, or go to Intel. Samsung was a background actor that had lost significant yield credibility and lagged both TSMC and Intel on US capacity and customer wins.
Samsung is a credible third alternative, not a background actor. SF2P at 70% yield (January 2026) is within striking distance of TSMC N2's 65–75%. Taylor, Texas fab is 90% production-ready. Apple executives physically visited the Taylor facility the same week the Apple-Intel preliminary deal was reported. Qualcomm and AMD are in "final negotiations" for Samsung 2nm. The thesis frame should be "trilateral with Samsung as a live option," not "bilateral Intel vs. TSMC." Intel retains structural advantages (first-mover, CHIPS Act $8.9B grant), but the Intel-exclusivity narrative is materially weaker.
2026-05-11-autoresearch-samsung-foundry-third-alternative-2026 provided concrete data: 70% SF2P yield, Taylor 90% ready, Apple physical visit, Qualcomm/AMD negotiations.
I undercounted Samsung's recovery because the seed research was thin on Samsung data. When a source doesn't give me data on X, I tend to underweight X's importance rather than flagging "unknown." Samsung was an unknown; I should have flagged it as a gap requiring its own research round rather than discounting it.
Nvidia's $5B Intel commitment was equity, not a wafer production order
Nvidia's $5B commitment to Intel validated 18A for production-scale AI silicon, and Nvidia would be a significant IFS customer for inference chip wafers. This was treated as evidence that Intel's 18A yields were production-ready enough for Nvidia's standards.
Nvidia's $5B is an equity stake + co-design partnership, not a wafer production order. More significantly, Nvidia reportedly stopped progressing on 18A foundry production. The AWS and MSFT commitments remain intact (those are wafer orders). But Nvidia should be removed from the "production anchor" list and treated as a strategic investor with co-design involvement.
2026-05-11-autoresearch-apple-intel-deal-may-2026-update explicitly flagged the distinction: "Nvidia $5B = equity stake + co-design, NOT wafer production order."
I accepted the framing of a vague commitment as production-level validation without verifying the contract structure. Press releases around equity investments and co-design partnerships often use language ("commitment," "partnership") that reads as production orders without being one.