Methods, sources and limits

AI-generated review input, not an Unjournal evaluation. No person has reviewed its content or checked its numbers.

Generated with Joe Francis’s isitcredible review suite v0.9.4 (MIT licence, The Catalogue of Errors Ltd). The Unjournal ran it with Claude Code (Anthropic’s Claude), 30 October–2 November 2026, blind to earlier Unjournal and AI assessments.

Review stages

  1. Review: AI agents examined the paper and available materials, ran mathematical and consistency checks, challenged findings in fresh contexts, and assembled the claim-by-claim synthesis.
  2. Audit: A mechanical audit checked completeness, scripts and source locators, correcting or routing issues back through the review.
  3. Digest: The issue ledger passed fresh fact, coverage and rendered-digest checks. Fairness and prose passes were recorded separately before finalization.

These are AI checks. They do not establish human verification.

Version and source limits

The paper version is arXiv:2609.15802v1, 14 September 2026. The review and audit are complete within their access limits. Replication remains OPEN because two cited sources could not be read: The Economist’s paywalled article and Halperin’s unpublished paper. No ruling rests on either.

No author code or data was deposited. The calibration was rebuilt from printed inputs and Epoch AI’s public data.

Source texts

Open source-text errata (E1–E5)

Five upstream slips remain open; none affects a finding, and the finalized digest already carries the corrected forms. They concern finding order, a label collision, a withdrawn 2× reversal, two page locators, and whether Whitfill and Wu (2025) was read. The synthesis’s “Not read” statement is incorrect: an assumptions agent read the source. The source texts remain unchanged; prefer the finalized digest where they differ.

Attribution and conflicts

Conflict: Claude is made by Anthropic. No author lists an Anthropic affiliation, but the paper leans on Anthropic sources (the Claude Mythos Preview system card and Favaro and Clark’s Anthropic piece) and illustrates its threshold with Anthropic models. METR funds the paper and employs the lead author; the system card names METR and Epoch AI (a co-author’s affiliation) as pre-release testers of Mythos Preview.

This disclosure identifies a potential conflict. It does not show that bias occurred.

The earlier Unjournal and AI assessments were not inputs to this blind run. This workspace presents only the finished isitcredible output.

Comment with Hypothes.is. Select a passage or open the sidebar. Public annotations may be incorporated into The Unjournal’s knowledge base and quoted, summarized or reposted on other Unjournal pages or public channels. Do not include private or confidential information. Help.