What happened: For the automated-R&D threat model, Anthropic ran semi-structured interviews with 31 academics, scientists, industry experts, officials and practitioners across seven fields, most external to Anthropic, and flags the method plainly as "a rough litmus test, rather than a rigorous or comprehensive assessment". The recurring finding (§3.6.1): frontier models handle established knowledge well but "still lack research taste", and so "cannot yet substitute for the judgment, intuition, and creativity of top-tier researchers". An energy interviewee expected the gap might close "in perhaps a year". The internal picture matches: §3.4.2's researcher survey (self-reported uplift with a geometric mean "on the order of 4×") lists the weaknesses as self-managing ambiguous week-long tasks, organisational priorities, "taste, verification, instruction-following, and epistemics"; §3.4.1 distils the failure modes to "calibration, self-monitoring, and judgment, which are precisely the properties" distinguishing autonomous research work from work needing a human in the loop — and adds "We note that this judgment gap is narrowing on at least some measures." The report also states its own incentive direction: it is "strongly incentivized to have our models substitute" and reports the negative finding anyway.
Diagnostic reading: Capability rises across the board and the named residual migrates to taste, novel idea generation, hypothesis quality, judgment, intuition and epistemics — the properties conventionally held to require human experiential contribution, and the ones no source in the document supplies criteria for. Three features make this a stronger data point than the usual commentary instance. Thirty-one experts across seven fields, most of them external to the lab, converge on the same remainder, so this is a distributed pattern in the vocabulary practitioners reach for rather than one commentator's reclassification. The lab records that its own incentive runs the other way and reports the negative finding regardless. And the residual is dated by the people drawing it — "perhaps a year", a judgment gap "narrowing on at least some measures" — which is the AI-effect pattern captured in advance rather than reconstructed afterwards, the more valuable and the rarer form. Nothing here says the experts are wrong or the residual empty: the diagnostic point holds either way, and it is that the terms doing the work are exactly the terms ARG-03 v2.6 asks after, with ARG-01 v1.5's cognition-pressure-point extension as the canonical treatment they feed. Thread 2 says nothing about experiencing, and this entry does not imply otherwise.
Fold, 2026-10-15 — Spangher, Huang, Haupt, Goodman, Yang, Ho & Koyejo, "Verifiable, Articulable, and Tacit Components of Preference" (arXiv:2610.03025v2 [cs.AI], 2 October 2026, revised 6 October, 136pp; abstract and results headings read from the arXiv HTML). The residual this entry records as a vocabulary — taste, judgement, intuition — is here measured as a gap. On a dataset of some 2.8 million texts and 317 million judgements across seven domains, the authors find "articulability gaps occur across nearly all tasks and datasets", that the gaps survive the articulation of more metrics, better judges and better aggregation ("No" to each, by the paper's own section headings), and that they grow with the number of judges, which the authors read through Collins's collective tacit knowledge; articulating a preference moves it away from its tacit dimension. The datum for this entry is that the experts' remainder is not only named but has a size, measured in domains that include mathematics and software, and that the measurement is made by people with every incentive to close it. It does not say what the remainder is, and neither does this entry; ARG-03 v2.6's question about which terms carry the claim stands, now with a quantity attached.
Source held at inbox/anthropic_risk_report_2026-08/; quotations verified against the PDF text layer 2026-08-27. The report itself calls the interviews 'a rough litmus test, rather than a rigorous or comprehensive assessment' — that scoping travels with every use. The entry must not claim the experts are wrong or the residual empty; the diagnostic point holds either way.