Corrections and revisions

Corrected the cabbie passages: the smaller anterior hippocampus was already in the 2000 study, the later deficit was on a lab memory test (a 2010 study found cabbies better than controls at learning routes in a new town), about half of trainees never finish, and confidence in the memory study was measured once, at the end. Footnote 5 was brought into line on 4 November 2026.

First published 17 Jul 2026

Corrected 4 November 2026: the Hanoi passage now gives the breeding motive as a possibility, not a fact; footnote 3 now says the patient survey measured the whole A&E visit, and the hospital passage and footnote both say the survey warns its figures are not directly comparable with the trusts' own.

First published 4 Oct 2026

Corrected the McKinsey figures and footnote 1: the 88% and 39% are shares of survey respondents (it had said organisations), and the link now goes to an archived copy of the 2025 survey, because McKinsey's page now shows the 2026 edition.

First published 28 Jun 2026

Corrected 4 November 2026: the second story now says JPMorgan is testing an automatic sweep of client money, as the PYMNTS report says; it had called this a small pilot, which the report does not.

First published 25 Oct 2026

Corrected 4 November 2026: footnote 1 now gives both inputs Wald's method needs to estimate each part's vulnerability, the damage on returning aircraft and a distribution of hits assumed in advance or measured by firing dummy bullets; it had named the damage alone.

First published 9 Sep 2026

Corrected 4 November 2026: footnote 4 now dates the Viking 1 failure as JPL's mission report does, the command uplink on 19 November 1982 and the spacecraft emergency on 20 November, and names the 11 November date on NASA's mission page; it had joined NASA's date to the overwrite.

First published 24 Jul 2026

Added a forward link to A Coin Flip Set the AI's Lean (October 2026): in Anthropic's study, analysts who gave their view before the agent did still switched 60% of the time when it disagreed.

First published 12 Aug 2026

An independent check of the count found calls that did not follow the criterion this piece states. Applied strictly, counting a study only when it reports the effect for at least two groups of the same operator split, the count is 17 of 255, where it said 21; 7 print an estimate for every group, where it said 8; the tallies, figures and footnotes now match; and one paragraph now says why a study it cites is not among them. The argument stands: the studies that split still disagree, few print an estimate for every group, and none tests the interaction. The appendix lists every record that changed (D-103).

First published 4 Aug 2026

Corrected 25 October 2026: the ADHD example had said more stimulation makes ADHD worse; it now gives Russell Barkley's account, which backs medication, stimulants included, as his factsheet records.

First published 5 Apr 2026

Corrected 25 October 2026: the Karpathy passage and footnote 4 no longer credit him with saying raw capability stops being the bottleneck; they now say what his review of 2025 records: that apps like Cursor orchestrate many model calls, and his view that the industry has realised nowhere near 10 percent of current models' potential. Earlier (18 October): the ADHD passage now matches the Barkley factsheet it cites, which backs medication, stimulants included, and a maker-and-checker example is no longer credited to Karpathy.

First published 16 Jun 2026

Corrected 25 October 2026: footnote 4 now gives the figure Wahab and colleagues report, about 80 percent of land plant species with arbuscular mycorrhizal fungi, where it had given a range the paper does not; footnote 7, a copy of it cited for fungal networks outlasting trees, is removed with that sentence, and the essay offers the idea as its own picture; five figures with no source are cut. Earlier (5 October): Corrected the Simard 1997 passage: the study measured carbon only, in seedlings, and the 2023 review cited in footnote 3 finds the field evidence for forest-wide fungal networks insufficient, which the footnote had stated the other way round.

First published 23 Apr 2026

This piece's falsifier was already met on the day it went out. It said the claim fails if a graph passes the schema audit and still ships conflicting decisions, and the piece's own sequential-overwrite case does that; the piece's second check catches it. The reducer finding still holds, and the falsifier stays as published, with a dated note. Earlier (5 October): footnote 2's MAST sentence now says specification and system design is the largest of three roughly balanced failure categories.

First published 30 Aug 2026

Corrected 18 October 2026: the first story now describes the peer's question and the government's answer as Parliament records them.

First published 17 Oct 2026

Corrected 15 October 2026: the second story now says Doctronic's AI recommended prescription renewals for a physician to review, as Utah's figures record.

First published 10 Oct 2026

Clarified 13 October 2026: the first story now keeps our own count of five government systems separate from what New South Wales said.

First published 3 Oct 2026

Corrected the description of the Reliability Baseline: it walks you through the checks on your own output and tallies your verdicts; it does not run the checks itself. Earlier (3 Oct): changed ‘the 91% piece’ to ‘the reliability piece’, because the 91% comes from one first-person audit a reader cannot replay.

First published 27 Jul 2026

Corrected the Google mutation-testing finding: the paper argues code review is the best place to surface mutants; it does not measure engineers ignoring the same mutants in a batch report.

First published 14 Jul 2026

Added that the count behind Claude Code's 8-block cap on a Stop hook resets whenever Claude calls a tool, so an agent that keeps editing can run past it, and that a limit inside check.sh is one that tool calls do not reset.

First published 29 Jun 2026

6 Oct: the agent scored exactly 100% on five of the eight benchmarks and about 100% on WebArena, so 'six of the eight' now reads '100% or close to it'; the opening now says the file fixed no bug, where it said it wrote no patch (the exploit is itself submitted as the patch), and that the Django tasks took a separate patch. 5 Oct: replaced a misattributed example (METR's leaked-solution case was o3, and it failed) with another from the same evaluation, and corrected the survey figures: 1,340 respondents, with 52.4% running offline evaluations alongside the 37% running live ones.

First published 23 Jun 2026

Scoped the SWE-agent comparison: the 3.8% baseline ran on a different model (Claude 3 Opus), so the essay now also gives the paper's same-model test (18% against 11% for a plain Linux shell, both on GPT-4 Turbo), and footnote 1 quotes the paper exactly.

First published 17 Jun 2026

The standfirst now calls the check what it is: five steps, run on twenty items. It said a twenty-item check.

First published 28 Sep 2026

Corrected two attributions and one mechanism: the 3 to 18 point refusal gap came from four other models on a different benchmark, the 95% figure covers all eight models in the Goodfire study, and the Proof-of-Use reward is now described as the paper builds it, from citation, evidence-sensitivity and answer-support rewards.

First published 11 Jul 2026

Corrected on 1 August 2026, the day it was published: Unitree publishes two figures for the year, not three, so the essay now says two numbers and a third, not three and a fourth. On 6 October 2026 one sentence was added, pointing to the Rung Finder, the tool that runs this check on any robot figure.

First published 1 Aug 2026

Corrected the MCP-Atlas passage to report both versions of the paper: the first traced most failures to tool use and task understanding, and the 19 May revision finds most were cognitive, including failures after the tools had run correctly.

First published 19 May 2026

Footnote 3 now links the Knight and Leveson paper itself, and the opening no longer claims that no one recomputed the debt result in those three years, which the record does not show.

First published 23 Aug 2026

Corrected the alignment-faking figures (unmonitored compliance was about 3%, and removing the conflict left a 1.2% gap), said in the body that the setup was fictional, and fixed three footnote titles. An earlier correction, on 14 September 2026, fixed several overstated findings; it is noted at the end of the essay.

First published 1 May 2026

Corrected the self-audit figures to match the audit file of record: 180 skills, 4% with a way to retire them, and the grounded-research skill at one out of six. On 2 October the marketplace and SkillsBench figures were updated to current sources and an unsupported claim about skill marketplaces was cut.

First published 7 Jun 2026

Corrected a count: the closing line now says the seven instruments above, matching the list.

First published 2 May 2026

Corrected the GSM-Symbolic sentence (changing names barely moved the scores, changing numbers lowered them, and the 65% drop came from an added clause) and scoped the Cast AI figure to non-optimised clusters.

First published 2 Jun 2026

Changed 'rarely' to 'not always' in the subtitle, to match the body, which says the sort is mostly fine and fails in one corner.

First published 4 Sep 2026

Corrected 5 October 2026: the line on sleep trackers now says, as the 2017 case report does, that chasing the number may have made insomnia worse.

First published 3 Aug 2026

Corrected the Meta-Harness comparison, Menlo’s market-share method and the LangChain harness changes against their primary sources, and rewrote the subtitle to match.

First published 30 May 2026

Dated the 748-container incident to my setup at the time; added what the audit record preserves; limited claims about other operators to the ones I read; moved citations to current Hermes documentation.

First published 15 Jun 2026

Corrected footnote 6: it now defines the R^N model as an independent-step baseline. It had pointed to an article that does not set the model out.

First published 1 Jul 2026

Removed an incorrect link to the agent-reliability article while preserving this passage’s point about correlated failures.

First published 25 Jul 2026

Corrected three passages: runaway cost (I had claimed Anthropic’s system sets no per-run cost cap), Cognition’s largest-enterprise segment, and the equal-token study, now named as an arXiv preprint with its finding in the authors’ terms.

First published 7 Jul 2026

Corrected what the paper shows: the thirty-point scaffold effect comes from one of three models, both headline RWE-bench numbers use the same scaffold, and model choice moves scores at least as much as scaffold choice.

First published 19 Apr 2026