Step 8: tool results

{
  "status": "supported",
  "claim_id": "83df12e2-68f5-4d25-8327-28d24ab513be",
  "assessment": "The forecast rests mainly on METR's measurements of AI \"time horizons\": the length of software tasks, measured by how long skilled humans take, that frontier agents complete with 50% success. In early 2026 the leading public models measured about 12 hours (Claude Opus 4.6, after a correction from 14.5), with stronger models put at 16 hours or more, and [[claim:d8be49b1-0f3c-43e5-8426-cc8bc25d9165|recent doubling times]] are estimated at roughly 3.5 months. If growth continues even at the older seven-month pace, the 50% horizon passes a 40-hour working week during 2027, and at the recent pace it passes several weeks well before the end of the year.\n\nThe human-hours scale behind those figures is itself uncertain at the long end. [[claim:14228127-3c75-4f02-aff0-9253356e34f6|Whether METR's baseline times reliably measure task length]] is doubtful for the longest tasks: METR measured human times for only 5 of its 31 tasks of eight hours or more and estimated the rest, and those long tasks anchor both the recent horizons and the post-2024 slope. If the estimates overstate length, today's horizons are lower than reported; if the error grows with length, the fitted doubling is faster than the true one. The extrapolation has a wide margin (several hundred hours at the recent pace against a threshold of one to a few working weeks), so a uniform error of a factor of two would not by itself defeat it, but the combination of a slower true doubling and an inflated starting level could leave the horizon near or below two working weeks at end-2027. This weakens the trend argument somewhat without reversing it.\n\nThere is also direct evidence that the threshold is already crossed for some tasks. In April 2026, Epoch AI and METR's MirrorCode benchmark reported Claude Opus 4.6 fully reimplementing a 16,000-line program that engineers estimated would take a human 2 to 17 weeks. Those tasks are precisely specified and checkable, and their lengths are engineers' estimates rather than baselines, so they show capability on a favourable kind of task, not on weeks-long work in general; their estimated lengths are, however, independent of METR's suite and sit far above the week threshold.\n\nRemaining uncertainties concern meaning as much as trend. The measured horizons are for software tasks at 50% reliability; 80% horizons are several times shorter, and messier work in other fields has not been shown to follow the same curve. METR's suite was nearly saturated in 2026, so a weeks-long horizon may not be cleanly measurable by the end of 2027. Read as \"frontier AI will be able to complete, with substantial reliability, some software tasks that take a skilled person weeks\", the claim is well supported. Read as reliable performance across weeks-long tasks in general, it is much less certain. The most prominent statement of the forecast comes from Anthropic, whose own models are the subject of the evidence, though the underlying measurements are METR's and Epoch's.",
  "confidence": 0.62,
  "claim_credence": 0.7,
  "marginal_yield": 0.35,
  "reasoning_trace": "Re-pass prompted by a consistency flag that the earlier assessment did not weigh the doubtful standing of METR's human baseline times. The flag is correct: the trend argument reads horizons in human-hours, and the baseline claim (doubtful, 0.35) reports that only 5 of 31 tasks of 8h+ were timed (METR Time Horizon 1.1, Jan 2026), with the rest estimated, and possible success-conditioning bias. That premise is now linked under the extrapolation argument.\n\nSensitivity check (arithmetic): from 12 h in early 2026 over ~21 months to end-2027: at 3.5-month doubling, 12 x 2^6 = 768 h; at 7-month doubling, 12 x 2^3 = 96 h (about 2.4 working weeks). If long-task lengths are inflated by 2x and the true doubling is 7 months, 6 x 2^3 = 48 h, just over one working week and below \"weeks\" read as two or more. So the conclusion survives most error scenarios but not a combined pessimistic one. Length-dependent overestimation would make the fitted doubling (about 103 days) too fast; the doubling-time quantity's 90% interval (82-132 days) does not include this measurement-scale uncertainty, which is a reason to widen my own effective range toward the 7-month case.\n\nMirrorCode (https://epoch.ai/publications/mirrorcode-preliminary-results) is independent of METR's baselines but its 2-17 week lengths are also estimates; even the low end is above the threshold.\n\nNet: credence lowered from 0.75 to 0.70 on the natural reading (some meaningful class of tasks taking skilled humans weeks, completed at 50%+ success by end-2027); status unchanged. Neighbour verdicts cohere: the baseline doubt is real but bounded, and does not contradict a supported forecast with this margin."
}