Chain-of-thought monitorability stayed roughly steady over the frontier RL training runs of OpenAI o3 and GPT-5.1 Codex Max.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The only evidence is OpenAI's own report on its monitorability evaluation suite (arXiv 2512.18311, section 6.1 and appendix I), which tracked two frontier RL runs, o3 and GPT-5.1 Codex Max. Measured monitorability rose sharply early in training and fell considerably in the middle of the Codex Max run. The authors say these movements largely track the number of thinking tokens, and they conclude that RL at current frontier scale "does not meaningfully degrade monitorability". Read as "no material net degradation across the run", the claim fits that report. Read as stable from start to finish, it fits less well.
The claim stands or falls with the suite. Whether OpenAI's evaluations validly measure chain-of-thought monitorability is itself unsettled. The suite can detect gross obfuscation, but its labels are noisy, two of its three archetypes face construct objections, and its scores follow CoT length. That last point matters here. Steady scores across training show that the runs produced no gross loss of legibility the suite can see. They do not rule out subtler obfuscation, and the score changes may partly reflect reasoning length rather than monitorability. The claim therefore holds for monitorability as this suite measures it, and the evidence for monitorability in a broader sense is indirect. The sample is two runs, and outsiders cannot replicate them. Two things would change the verdict: per-step data showing a lasting decline late in either run, or an independent check showing that the suite misses obfuscation that training actually produces.
Full reasoning: the evidence and decisions behind this verdict
The trigger was the first assessment of the framework premise, that the intervention, process, and outcome-property evaluations validly measure monitorability. It came back inconclusive, at confidence 0.6. The earlier assessment of this claim (supported, 0.65) had already named the suite's validity as a caveat. The new verdict makes that caveat concrete, and one point bears directly on this claim. The premise's assessment finds that the suite's scores track CoT length. The paper itself attributes the early rise, the Codex Max dip and the o3 declines on GPQA, DAFT and sandbagging to changes in thinking-token count. So the training curves partly measure reasoning length. That confound cuts both ways, and the main conclusion (no net degradation) survives it. The cost falls on what the scores can show: they speak to gross legibility, not to subtle obfuscation.
The status stays supported because the evidence still favors the claim as measured and is indirect for the broader construct. The doubt sits with the framework, not with the observed trend. Credence goes from 0.65 to 0.6 and confidence from 0.65 to 0.6 to reflect the inconclusive framework premise. Doubtful was considered and rejected: nothing points to degradation, only to limits on what the measure can see. Source: arxiv.org/html/2512.18311, read on the earlier pass.
Decomposition
The claims this one rests on directly. ↗︎ opens a subclaim; the map shows how they fit together.
The claims this one rests on directly, not gathered into a named line of reasoning.
- assumesbackground the parent's framing takes as givensteward instructions →OpenAI's intervention, process, and outcome-property evaluations validly measure chain-of-thought monitorability. ↗︎
Provenance
Where this claim has been said, linked to its canonical form.
These results suggest that reinforcement learning does not meaningfully degrade monitorability, even at the frontier scale at OpenAI; however, we underscore that these results only represent two experiments and more thorough study is needed.
Section 6.1, Effect of high scale RL: average monitorability across the suite tracked over the o3 and GPT-5.1 Codex Max RL runs; the same passage reports a significant early rise and a considerable mid-training dip for Codex Max.
The source's own evidence bears what it asserts. The paper's own data show an early rise in average monitorability and a considerable mid-training dip for GPT-5.1 Codex Max, with no net decline; it states only that RL does not meaningfully degrade monitorability, and flags that the result rests on two runs. The trends track the number of thinking tokens, and three toy evaluations declined over the o3 run. Worth reading closely: Section 6.1 and Appendix I hold the only data on the two runs; the per-evaluation curves show how steady "steady" is.
Assessment history
0 status changes over 2 assessments. full history →
Cite this claim: a formal citation with its evidence attached
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Oct 30, 2026. Every judgment on this page is accompanied by a reasoning trace.