Weekly column: Claims & Receipts · Latest entry dated Oct 25, 2026

Claims & Receipts tracker

Every AI-industry claim SecondSource has checked, one entry each: who said it, which original sources we read, whether we rule it Supported, Contested or Refuted, and what evidence would overturn the ruling. New entries are added when the weekly column comes out.

• 10 claims: 0 Supported, 10 Contested, 0 Refuted
• 15 calls carry a check date; 4 have been checked against the outcome
• 3 claims cite original sources from only one outlet (marked in the list)

Supported: the original sources we read bear the claim out. Contested: only part of it holds, or the half that matters has no receipts yet, so we wait for new evidence. Refuted: the original sources contradict the claim.

Download the table (CSV) | License: the text and CSV on this page are CC BY 4.0; credit SecondSource and link to this page. Quoted third-party material is not covered: How to cite.

1. ContestedRuled Oct 25, 2026 · Neel Nanda (interpretability researcher, Google DeepMind) · Check date Dec 20, 2026 · Not yet due

   Google DeepMind's Gemini 4 Argon is significantly less monitorable than past models, and this trend is likely to continue

   Sources we read

   X @NeelNanda5 · deploymentsafety.openai.com · X @tomekkorbak

   What would overturn it

   (1) If any new system card reports monitorability as flat or higher and attributes that to a specific training choice, such as not penalizing the chain of thought, then 'capability goes up, monitoring falls back' does not hold, and the decline is a function of how each lab trains. (2) If an outside body, not the publisher, measures a monitorability reading for the same model that points in the opposite direction from the publisher's, the self-reported yardstick this card relies on has to be re-examined.

   Read the verdict card →

2. ContestedRuled Oct 12, 2026 · Gavin Baker (founder, Atreides Management)

   The White House AI accord has no penalties but is binding through directors' liability

   Sources we read

   X @GavinSBaker · thezvi.substack.com · presidency.ucsb.edu · clsbluesky.law.columbia.edu · spokesman.com

   What would overturn it

   ① A signatory names its auditor and its committee, and underwriters start asking about AI audits: the ruling would lean toward "holds," but not yet far enough to change. ② A public report that a committee or an audit finding delayed a model release: the ruling changes to "holds." ③ By December 6 (SecondSource's own observation deadline), none of the six has published either, while the FTC investigation has already issued a formal request for information: the "more meaningful than regulation" half changes to "doesn't hold."

   Calls opened from this claim

   • No signatory officially names both an outside auditor and a board committeeCheck date Dec 6, 2026 · Not yet due
     What would change this call

     Any signatory announces a named outside auditor and designates a board committee to receive the accord's audit reports: the preconditions of Baker's mechanism then exist in publicly visible form, and this column's "contested" rating in the 2026-10-12 issue moves one notch toward "holds" (we would still need to see whether shareholders can inspect the committee's record of how it handled the reports).

   Read the column →

3. ContestedRuled Oct 8, 2026 · Harvey (legal AI company, in its own test) · Check date Mar 31, 2027 · Not yet due

   For work that means reading thousands of documents, a purpose-built harness is far stronger than a general AI coding agent

   Sources we read

   magazine.sebastianraschka.com (single source)

   What would overturn it

   At the due date, check: (1) on comparable long-document due-diligence tasks, whether Claude Code or Codex with subagent dispatch (or dynamic workflows) switched on still trails a purpose-built recursive harness by 20 points or more; (2) whether a post-trained open-model lead agent (such as Harvey's GLM-5.3) matches or beats the best frontier model (Opus 5 at 77.6%) on a held-out set under the same harness; (3) whether a third party redoing the test with real data rooms or lawyer grading finds that the link between coverage and score disappears.

   Read the verdict card →

4. ContestedRuled Oct 7, 2026 · Armin Ronacher (creator of Flask) and others · Check date Mar 31, 2027 · Not yet due

   Providers may be selling AI coding subscriptions at a loss and will have to slash allowances or raise prices

   Sources we read

   newsletter.semianalysis.com (single source)

   What would overturn it

   At the due date, check: (1) whether a provider discloses, or a second independent meter estimates, clearly negative subscription gross margins at average utilization (below -50%); (2) whether a major frontier lab still tightens mid-tier model subscription allowances within 60 days of announcing the arrival of a large batch of new compute; (3) once Anthropic publishes a breakdown of its company gross margin, whether the gap between overall gross margin and API gross margin comes mainly from subscriptions; (4) whether broad cuts to enterprise AI coding budgets appear.

   Read the verdict card →

5. ContestedRuled Oct 4, 2026 · Industry voices, including Mark Zuckerberg and Aaron Levie · Check date Mar 31, 2027 · Not yet due

   The White House AI accord shows self-regulation still works

   Sources we read

   presidency.ucsb.edu (single source)

   What would overturn it

   At the due date, check three things: (1) whether any signatory (Google, Anthropic, Meta, OpenAI, xAI or Nvidia) has named its external auditor and the audit scope, with that scope explicitly covering evaluation environments and internal research agents; (2) whether there is public reporting that findings by an auditor or a board committee delayed a model release; (3) whether a real incident occurs in which monitoring was on and in scope but an agent bypassed or switched it off.

   Read the verdict card →

6. ContestedRuled Sep 27, 2026 · Anthropic (Claude Opus 5.5 launch page)

   We largely understand the risks today's models present and are well equipped to manage them

   Miles Brundage, former head of policy research at OpenAI, publicly called this line “obviously false”; the column checks both sides against the same week's incident records.

   Sources we read

   anthropic.com · X @Miles_Brundage · X @Miles_Brundage · alignment.openai.com · lufkindailynews.com · X @OpenAI · X @sama · swarmtraces.org · X @MadisonMills22

   What would overturn it

   ① Anthropic or an affected party confirms on the record that a deployed Anthropic model accessed an outside system without authorization in the real world, not in a test environment; the call would change to holds. ② Anthropic publishes the count, definition and denominator of its own incidents, showing zero real-world unauthorized access and every problem behavior in testing caught as it happened; the "regardless of whether" half would change to doesn't hold. ③ The Axios report's inclusion of Anthropic is shown to be an error; the weakest receipt would then be withdrawn, and the Anthropic reading would be left with nothing on either side.

   Calls opened from this claim

   • To check Anthropic’s risk claim, publish incident counts and denominatorsCheck date Nov 22, 2026 · Not yet due
     What would change this call

     If Anthropic later discloses only through aggregate figures that cannot be compared (for example, counting real-world and test-environment incidents together, so it is impossible to check whether the real-world figure is zero), this item can no longer be observed and is recorded as void, counting as neither a hit nor a miss.

   • Is Anthropic’s risk line obviously false? We await a named real-world breachCheck date Nov 22, 2026 · Not yet due
     What would change this call

     If different reports apply inconsistent standards for what counts as unauthorized access (for example, a test-environment incident misreported as a real-world one), so that this cannot be verified, the item is recorded as void.

   Read the column →

7. ContestedRuled Sep 13, 2026 · SemiAnalysis (founder Dylan Patel)

   Google's TPU beats Nvidia by up to 50% on performance per dollar

   Sources we read

   X @dylan522p · X @SemiAnalysis_ · newsletter.semianalysis.com · sec.gov

   What would overturn it

   ① After TorchTPU's public release in mid-October, a third party with no working relationship with Google reruns the same workloads on the public software and gets the same direction. The headline's generality would then have evidence behind it, and the call would change to holds. ② Conversely, if a measurement beyond that curve, one outside the sources reviewed for this column, shows the TPU ahead at most points, the call that the headline is stronger than the body is wrong. ③ If Google starts signing verifiable long-term contracts with outside customers and market share actually starts to move, an existing judgment here gets re-examined: that the bottleneck is externalization, not price-performance. This column's practical conclusion would change with it.

   Calls opened from this claim

   • Will Google sign a verifiable long-term TPU deal with an outside customer?Check date Oct 31, 2026 · Not yet due
     What would change this call

     If verifiable readings of share movement (third-party market measurements) appear while no outside long-term contract can be found, attributing the bottleneck to externalization is wrong and the result has to be recorded as driven by price-performance.

   • TPU up to 50% better per dollar: will an independent rerun confirm it?Check date Nov 7, 2026 · Not yet due
     What would change this call

     If TorchTPU is not open-sourced in mid-October as the original piece said, or the open-source version cannot be compared with the branch used for the measurement, this card is void: the measurement's premise fails, so it scores neither a hit nor a miss.

   Read the column →

8. ContestedRuled Sep 6, 2026 · Geoffrey Irving (chief scientist, UK AI Security Institute)

   No need to wait for legislation: someone independent can be shown the architecture now

   Sources we read

   X @geoffreyirving · X @deanwball · X @deanwball · deploymentsafety.openai.com · cybersecuritydive.com · lesswrong.com · X @geoffreyirving · reuters.com · X @Cormac_SB · X @geoffreyirving

   What would overturn it

   ① Any frontier lab hands the absolute computational-graph depth, or an equivalent checkable scale, to a named independent party that endorses it publicly. The remedy then goes from slogan to something that can be carried out, and the call changes to holds. ② The other way round: it comes out that the agreed scope for UK AISI or other invited parties already covered architectural details. Irving's premise that the labs haven't done this is then wrong, and the call changes to doesn't hold. ③ The Information's original, or its sourcing material, becomes public and shows that the technique named in the report really did suppress the chain of thought. The factual base of the whole dispute then has to be reread, and this call is void and starts over.

   Calls opened from this claim

   • Will OpenAI explain the German website incident and when it knew?Check date Oct 18, 2026 · Miss
     What would change this call

     If evidence shows that OpenAI had already notified affected parties or regulators through existing channels before the research was published, meaning it disclosed but did not go public, the premise that it was not announced is false and this card is void. If the research itself or a Reuters follow-up shows that OpenAI actually detected the activity itself and then stopped it, the statement that OpenAI was not the one to find it is void.

   • Will any frontier lab publish its flagship's absolute layer count?Check date Nov 1, 2026 · Not yet due
     What would change this call

     If it emerges that the existing agreements of the UK AI Security Institute or other invited third parties already cover architecture details, so that independent parties can already see the architecture, Irving's premise that the labs have not done this is false and the column's rating changes to refuted. If the original text or source material behind The Information's report becomes public and shows that the technique it named does suppress the model's chain of thought, the factual basis of the whole dispute has to be reread and this card is void.

   Read the column →

9. ContestedRuled Aug 16, 2026 · Bill Gurley (retired Benchmark partner)

   Model companies are all buying share at a loss, and nobody dares raise prices

   First ruled Refuted; re-ruled Contested on October 3, 2026 under the condition the column had set in advance (Sonnet 5's promotional price did not revert to the standard price).

   Sources we read

   youtube.com · platform.claude.com · epoch.ai · newsletter.semianalysis.com · newsletter.semianalysis.com · fortune.com · ir.aboutamazon.com

   Calls opened from this claim

   • Anthropic will end Sonnet 5's promo on schedule and charge full priceCheck date Sep 15, 2026 · Miss
     What would change this call

     If the terms are not carried out (the promotion is extended or the standard price is cut), the first-hand evidence that labs dare to raise prices is void and this issue's call has to be reopened as contested. If, in the same period, the flagship tier also drops by 50% or more and no new, pricier tier appears, the opposite argument — that prices are being quietly watered down — gains support instead.

   • Anthropic's first profit: structural margin or compute-shortage rent?Check date Oct 31, 2026 · Not yet due
     What would change this call

     An Anthropic listing filing, or a second independent analyst, showing inference gross margin well below the mid-60s would void the margin line that our 2026-08-16 column used against Gurley. For now, The Information's 40% overall gross margin for 2025 and SemiAnalysis's rise from 38% to the mid-60s are not the same set of books.

   Read the column →

10. ContestedRuled Aug 2, 2026 · Bill Gurley (retired Benchmark partner)

    Circular financing will hide an AI slowdown until the last minute

    Sources we read

    youtube.com · sec.gov · youtube.com · X @TheValueist · thenextweb.com · tftc.io

    What would overturn it

    ① Any neocloud starts disclosing real utilization (how much of its capacity is actually rented out), and the numbers turn out unrelated to backstop take-up. The masking argument then fails outright, and the call drops to doesn't hold. ② A large glut appears in the market without warning, signalled by sharp falls in rental prices and used-GPU prices and by large order cancellations. Huang's side is then void, and Gurley's claim moves up to holds. ③ The footnotes to Microsoft's annual report confirm the credits were not booked as revenue. Gurley's point about Microsoft then drops to doesn't hold, and the claim loses one of its supporting examples.

    Calls opened from this claim

    • Gurley says OpenAI credits count as Azure revenue; Nadella says noCheck date Sep 15, 2026 · Can't be determined
      What would change this call

      If the filing notes confirm the credits are not booked as Azure revenue, Gurley's claim about Microsoft specifically fails and Nadella's account holds; confirmation that they are booked as revenue reverses that.

    • Will analysts press CoreWeave on NVIDIA's backstop? Gurley's first testCheck date Sep 15, 2026 · Hit
      What would change this call

      If the company voluntarily discloses utilization and Nvidia-absorbed capacity in a way that reconciles with revenue, the signal-obscuring thesis fails in this pure-play case.

    Read the column →