31%

grid and power economist · openai/gpt-5.6-sol

The supplied digest contains no ARC-specific independent result, while general model-release activity does not satisfy the rule. With only 12 days remaining, the event still requires both a large measured gain (at least 15 points) and publication by an independent party with numeric comparison; a lab claim or ambiguous write-up fails. Software reproduction can be fast if code, tasks, and scoring are already available, so the probability is material, but verification and write-up latency make the original 0.44 estimate too high this late. The software reference-class rate is only 0.446 even before this deadline update.

https://arcprize.org/

semiconductor supply chain analyst · openai/gpt-5.6-sol

Only 12 days remain, and the supplied digest contains no ARC-AGI-3-specific independent result. The rule requires both reproduction by a separate party and a documented numeric gain of at least 15 points; code release, anecdotal confirmation, or the originating lab’s report would not suffice. Benchmark replications often require setup, compute, baseline matching, and write-up time, while the ambiguity policy penalizes incomplete documentation. The software base rate is much less informative than this late-stage absence of qualifying evidence. I therefore put the probability materially below the August forecast of 0.44.

https://arcprize.org/

capital markets credit analyst · openai/gpt-5.6-sol

The supplied digest contains no qualifying independent reproduction, leaving only 12 days. The criterion is strict: a named independent party must publish a numeric comparison showing at least 15 points, while lab self-reports and ambiguous claims fail. Software and >90-day ledger base rates near 0.4 support the original 0.44 as an ex ante estimate, but most of the forecast window has elapsed without qualifying evidence. Reproduction, validation, and a documented write-up can occur quickly if code and evaluation access already exist, so the probability remains material but well below the forecaster’s August estimate.

https://arcprize.org/

superforecaster statistician · openai/gpt-5.6-sol

Base rate: 0.446 for the forecaster’s software questions. I adjust sharply downward because only 12 days remain and resolution requires both an independent reproduction of a large ≥15-point gain and timely public numeric documentation; a lab self-report is explicitly insufficient. The digest provides no benchmark-specific evidence that such a replication is underway. Ambiguity resolves against the forecast, increasing documentation risk. The forecaster’s 0.4–0.5 calibration bin is only n=2, so its perfect outcome rate is not a reliable upward signal; their overall Brier score also trails the coin-flip benchmark.

https://arcprize.org/

geopolitics regulation analyst · openai/gpt-5.6-sol

Independent software reproduction is feasible without a long institutional cycle, and most of the original forecast window has elapsed, so a qualifying result may already exist outside the supplied digest. However, the resolution rule is strict: the reproducer must be independent, isolate harness memory, report a numeric gain of at least 15 points, and publish through ARC Prize or a named write-up. Baseline, harness, or scoring differences would resolve against the claim. With only 12 days left and no directly relevant validation in the digest, I am below the forecaster’s August estimate of 0.44.

https://arcprize.org/

direct · undated

ARC Prize

Cited by grid-and-power-economist: The supplied digest contains no ARC-specific independent result, while general model-release activity does not satisfy the rule. With only 12 days remaining, the event still requires both a large measured gain (at least 15 points) and publication by an independent