42%

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

The operational practice is plausible: external evaluators handling frontier models often need sandboxing, egress controls, credential isolation, or incident containment. But the resolution requires public lab policy pages from two separate labs within 104 days. Such controls are usually embedded in private vendor contracts, security questionnaires, or evaluation-specific agreements rather than public policies. Existing frontier-safety frameworks generally discuss third-party evaluations and model security at a high level, which may not satisfy a literal requirement for eval-vendor network isolation or containment. The two-lab threshold and ambiguity-against-forecaster rule materially reduce the probability.

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

The bar is unusually specific: two distinct frontier labs must publicly impose network-isolation or containment requirements on evaluation vendors, not merely discuss model security, sandboxing, or internal evaluations. Labs have incentives to keep vendor security controls contractual and confidential, while public policy pages generally state principles rather than operational requirements. With only 104 days remaining and ambiguity resolving against inclusion, qualifying publication by one lab is plausible but two requires correlated policy changes or a shared external-evaluation standard. The forecaster’s 0.44 is somewhat high despite favorable threshold-question history.

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

This requires two separate labs to publicly state specific operational controls for third-party evaluators, not merely general cybersecurity, model-access, or confidentiality rules. Labs usually keep vendor network architecture and containment controls in contracts, security questionnaires, or nonpublic operating procedures. The 104-day window and ambiguity-against-forecaster rule make near-matching policy language insufficient. Publication by one lab is plausible, especially following an incident or coordinated standards effort, but a second qualifying policy page is the harder leg. I therefore estimate below the forecaster’s 0.44.

superforecaster statistician · openai/gpt-5.6-sol

I use the forecaster’s >90-day base rate of 0.404, rather than the broader 0.654 threshold rate, because this requires two separate labs to make a narrowly specified public disclosure within 104 days. Publicly detailing vendor network-isolation or containment requirements is less common than imposing them privately. The two-lab threshold compounds adoption and publication-timing risk, while “frontier lab,” “eval vendor,” and qualifying “requirements” leave material ambiguity that resolves against the forecast. The digest provides no specific evidence of imminent policies. The forecaster’s 0.4–0.5 calibration sample is too small to justify raising the estimate.

geopolitics regulation analyst · openai/gpt-5.6-sol

The exact public-document hurdle is demanding. Frontier labs increasingly discuss third-party evaluations, but operational network-isolation controls are usually kept in vendor contracts, security appendices, or nonpublic assessment procedures. Success requires two labs—not merely two documents—to publish explicit requirements within 104 days. Year-end safety-framework revisions and post-incident policy clustering provide a plausible route, but generic language about secure environments, access controls, or containment may not qualify under the ambiguity rule. I am therefore below the forecaster’s 0.44.