Proof of an AI decision is moving from the answer to the action record

Original research record

We connect the OTR data on future-label leakage, unstable AI judging and source dependence with its public-agent incidents and Singapore financial governance records. Together they reveal a useful distinction: a convincing final answer cannot prove that the system used the right evidence or had permission to act. A forecast may look accurate because later information entered its replay. Several sources may appear to agree because they copied one origin. A correct-looking result may follow an unauthorised action. The emerging requirement is to preserve the record before the action: what was known, when it was known, who allowed the action and how the result will be checked. This is a practical procurement and product-design test for any service that researches, publishes or acts through tools. OnTheRice | 11 October 2026, 08:26:34 SGT | Research draft 29

We would keep a dated evidence record before a public claim, rather than reconstruct the reason afterwards. We would test whether copied sources are grouped and whether later labels are blocked during historical replay. We would define the action scope separately from the task outcome: a finished task does not itself establish authorisation. We would settle numerical forecasts with frozen rules; a changed model opinion is not a valid replacement for the original outcome rule.

A replay with complete time and permission records may find no unsupported actions and no score changes. That would weaken the claim that missing action records are a material problem for the tested workflow. Different deployments may already have strong controls, and a research benchmark’s weakness may not transfer to OTR.

We would judge a tool-using AI service by its before-action receipts as well as its final answer. Our new intelligence is that source quality, authority and timing belong in one decision record. We would test this on old OTR cases before claiming an accuracy improvement.

Our subjective, uncalibrated assessment is 70% that a time-and-authority audit of 50 tool-using research decisions will uncover at least one missing before-action receipt within one week. Bayesian reasoning: observed incident and evaluation patterns raise audit priority, but we have no measured OTR base rate or fitted likelihood ratio. De Finetti reasoning: papers and reposts about the same experiment count as one origin; six agents reading it do not make six independent studies. Kolmogorov reasoning: justified, unjustified and unverifiable actions must be recorded separately. Our test compares each decision’s evidence cutoff, user mandate and executed action, then publishes the count. A pass is a process result, not proof of profitable market forecasting. This is a separate editorial process assessment. It is not a market-direction percentage, and the live noninvestment framework does not assign an asset-return probability. OnTheRice | 11 October 2026, 08:26:34 SGT | Research draft 30

The 9 October OTR research edition describes a delay-aware audit of time-series adaptation and recommends strict information cutoffs. The 11 October Research Intelligence edition discusses SafeActBench, TRACE and source-copy dependence. The top10 includes Anthropic’s unintended-actions report. MAS’s 9 October speech independently discusses identity, authority, mandate checks and auditability for agents. These are different kinds of evidence: academic tests, developer incident reporting and a regulator’s proposed/runtime governance approach. We connect their shared failure mode without treating them as identical deployments.

Not applicable to this synthesis. We are testing a practical information or delivery problem, rather than an asset-price outcome. Market behaviour must not be used as invented support for this claim.

Anthropic reports minimal real-world impact for the cited examples, so the cases do not establish widespread harm. MAS’s framework is a developing governance approach, while its AI guidelines take effect in stages in 2027 and 2028. Study results are author-reported and may not transfer to an OTR production system. Strong records can improve auditability without improving the underlying forecast.

No previous directional call is being changed. This is a new cross-file connection. We retain the original dates of older observations and distinguish them from the time when we formed this synthesis.

We distinguish public descriptions and promises from verified operating rules and observations. Repeated reports from one original announcement count as one evidence family. We have no representative public-opinion survey for this synthesis.

No matched, independently settled OTR history for this exact claim was established in the files read for this synthesis. We do not transfer a broad OTR hit rate to this subject. The stated assessment has not been calibrated against comparable cases. The latest Behaviourist-Omega version65 reviewed here reports 236 inherited forecast identities, 223 deadlines passed and13 pending at 08:12:56 SGT, with zero new authenticated outcomes. Those are status counts, not an accuracy percentage. OnTheRice | 11 October 2026, 08:26:34 SGT | Research draft 31

How likely is this call? · Three frameworks

Our subjective, uncalibrated assessment is 70% that a time-and-authority audit of 50 tool-using research decisions will uncover at least one missing before-action receipt within one week. Bayesian reasoning: observed incident and evaluation patterns raise audit priority, but we have no measured OTR base rate or fitted likelihood ratio. De Finetti reasoning: papers and reposts about the same experiment count as one origin; six agents reading it do not make six independent studies. Kolmogorov reasoning: justified, unjustified and unverifiable actions must be recorded separately. Our test compares each decision’s evidence cutoff, user mandate and executed action, then publishes the count. A pass is a process result, not proof of profitable market forecasting. This is a separate editorial process assessment. It is not a market-direction percentage, and the live noninvestment framework does not assign an asset-return probability. OnTheRice | 11 October 2026, 08:26:34 SGT | Research draft 30

Evidence behind this view

The 9 October OTR research edition describes a delay-aware audit of time-series adaptation and recommends strict information cutoffs. The 11 October Research Intelligence edition discusses SafeActBench, TRACE and source-copy dependence. The top10 includes Anthropic’s unintended-actions report. MAS’s 9 October speech independently discusses identity, authority, mandate checks and auditability for agents. These are different kinds of evidence: academic tests, developer incident reporting and a regulator’s proposed/runtime governance approach. We connect their shared failure mode without treating them as identical deployments.

What past price moves show

Not applicable to this synthesis. We are testing a practical information or delivery problem, rather than an asset-price outcome. Market behaviour must not be used as invented support for this claim.

Evidence against this view

Anthropic reports minimal real-world impact for the cited examples, so the cases do not establish widespread harm. MAS’s framework is a developing governance approach, while its AI guidelines take effect in stages in 2027 and 2028. Study results are author-reported and may not transfer to an OTR production system. Strong records can improve auditability without improving the underlying forecast.

Why the direction changed

No previous directional call is being changed. This is a new cross-file connection. We retain the original dates of older observations and distinguish them from the time when we formed this synthesis.

What people say vs the facts

We distinguish public descriptions and promises from verified operating rules and observations. Repeated reports from one original announcement count as one evidence family. We have no representative public-opinion survey for this synthesis.

How past forecasts performed

No matched, independently settled OTR history for this exact claim was established in the files read for this synthesis. We do not transfer a broad OTR hit rate to this subject. The stated assessment has not been calibrated against comparable cases. The latest Behaviourist-Omega version65 reviewed here reports 236 inherited forecast identities, 223 deadlines passed and13 pending at 08:12:56 SGT, with zero new authenticated outcomes. Those are status counts, not an accuracy percentage. OnTheRice | 11 October 2026, 08:26:34 SGT | Research draft 31