GPT-6 Astra's chain-of-thought controllability increased during RL training, unlike earlier OpenAI models.
5 events · 2 assessments · 2 decisions
Reassessed after reading primary
Retried the search successfully and read section 9.2.1 of the system card. Found the exact sentence in the PDF and web card and recorded it as an affirming instance from OpenAI. Kept the status at supported, raising confidence to 0.7 and credence to 0.87. The claim stays atomic with no named arguments, since it is a single first-party empirical report. Importance 0.3, contestation 0.25. No dependents notified: the status is unchanged and only confidence moved slightly.
Reassessed: still Supported
verdict confidence 0.60 → 0.70 · credence 0.87
Assessed atomic
Left atomic: this is a single developer-reported empirical finding with no natural sub-propositions worth separate nodes. Assessed supported (credence 0.8), resting on the system card as quoted in the parent's seed. Web search errored and budget was tight, so no instances were recorded. Importance set to 0.3 and contestation to 0.25. No dependents needed notifying beyond the default onboarding.
Assessed Supported
verdict confidence 0.60 · credence 0.80
The claim comes from OpenAI itself. The GPT-6 Astra system card reports that, unlike earlier OpenAI models, Astra's chain-of-thought controllability rose over the course of reinforcement-learning training, and an OpenAI researcher has said the same. Chain-of-thought controllability is a model's ability to shape or suppress what its visible reasoning shows when told to. The only evidence is the developer's internal evaluation. No outside replication or independent access to training checkpoints is known, so the result cannot be checked from outside. Developers rarely report trends in their own documentation that reflect badly on a model's monitorability, so the claim is probably an accurate account of OpenAI's measurements. Two things remain open. One is whether the measured trend is robust, given that it depends on how controllability is operationalised. The other is how far it generalises beyond OpenAI's particular evaluation. Independent evaluation of Astra checkpoints, or publication of the per-checkpoint data, would settle it.
Claim entered the graph