GPT-6 Astra's training applied significant direct optimization pressure to its chain of thought.
Assessment
The evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.
The only direct evidence on this question is the developer's own statement. OpenAI's GPT-6 Astra system card says the company is "fairly confident that there are no significant direct CoT grading pressures" in Astra's training. It attributes Astra's reduced chain-of-thought monitorability to greater single-forward-pass reasoning and greater control over the chain of thought, not to training on the chain of thought itself. That confidence has a concrete basis: by May 2026, before Astra was released, OpenAI had built a system that scans its RL runs for accidental chain-of-thought grading.
The case for the claim is circumstantial. OpenAI has accidentally graded chain-of-thought reasoning before, in earlier deployed models, so such pressure can arise unnoticed. In addition, Astra's ability to control its chain of thought rose during RL. Rising controllability fits capability gained through RL just as well as it fits pressure on the chain of thought, though, so it does not single out direct pressure. The precedent is exactly what led to the scanning system, which makes a repeat less likely to go undetected.
The evidence weighs against significant direct pressure, but it does not settle the matter. No outside party can inspect the training setup. "Fairly confident" is a hedged statement. Indirect pressure, such as spillover from grading final outputs, is not covered by the denial. Independent audit access to Astra's RL reward pipelines would resolve the question, and so would a published account of what the scanner covered.
Full reasoning: the evidence and decisions behind this verdict
Model formed before searching: only the developer can observe whether reward signals touched the chain of thought. The developer's own denial would be the main evidence, and evidence for the claim would most likely come from leaks, a developer admission, or indirect signs like rising CoT controllability. Searching confirmed this picture.
Evidence read: - GPT-6 Astra System Card (deploymentsafety.openai.com/gpt-6-astra). It says: "We are fairly confident that there are no significant direct CoT grading pressures." It attributes the monitorability drop to greater forward-pass reasoning and greater CoT control compared with GPT-5.6 Sol. - Micah Carroll and Tomek Korbak, OpenAI, May 2026 (x.com/tomekkorbak/status/2052456355244707885). OpenAI found instances of CoT grading in previously deployed models after building a system that scans all OpenAI RL runs for accidental CoT grading. They found no clear evidence that this degraded monitorability. The seed note described this as happening "at another lab". It was OpenAI itself. - Zvi Mowshowitz, "Astra is hard to monitor" (thezvi.wordpress.com/2026/09/08/astra-is-hard-to-monitor/). He reads rising CoT control over training as implying that OpenAI is "inadvertently training this". He also notes that OpenAI could have changed training in many ways. This is an inference about an indirect effect, not evidence of direct grading. - Korbak et al., arXiv 2510.19476, on feedback spillover. Pressure on outputs can shape the chain of thought indirectly. This is a reason why "no direct grading" does not exclude all pressure. The claim as worded concerns direct pressure, however.
Weighing: the evidence against the claim is developer testimony, and the scanning infrastructure makes it more credible than a bare assertion would be. The evidence for the claim is a base rate (accidental grading has happened at this developer) plus a non-specific signal (rising controllability). The status is doubtful rather than contradicted because the evidence is indirect and comes from a single interested party. Searches across the system card, press coverage (Transformer, PCWorld, Yahoo Tech) and commentary found no report alleging direct CoT grading of Astra. Such a report could be expected only from the developer itself, so this absence weighs modestly.
What would change the verdict: an OpenAI disclosure that the scanner found CoT grading in Astra runs, independent audit findings, or evidence that the scanner missed classes of direct grading.
Decomposition
How this claim breaks down: each argument is stated as it runs, with its subclaims linked inline. ↗︎ opens a subclaim; the map shows how they fit together.
Because OpenAI has inadvertently graded chain-of-thought reasoning in earlier models' RL training, such pressure can arise without the developer intending or noticing it. And because Astra's CoT controllability rose over the course of RL, training evidently shaped how Astra uses its chain of thought, which is what optimization pressure on the CoT would produce. Together these suggest Astra's training applied significant direct pressure to its chain of thought.
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 1 supported1 not separately assessed
Even granting both premises, the conclusion gains little. Past accidental CoT grading only raises a base rate, and it prompted the scanning system that makes a recurrence easier to catch. Rising controllability during RL is equally expected if RL simply increases capability or applies only indirect pressure, so it does not pick out direct grading. At most, the argument keeps the question from being closed.
OpenAI's GPT-6 Astra system card states that the developer is fairly confident there were no significant direct CoT grading pressures in Astra's training. That statement is grounded by the fact that OpenAI automatically scans its RL runs for accidental CoT grading, a check that would have caught significant direct grading of Astra's chain of thought. So Astra's training applied no significant direct optimization pressure to its chain of thought.
- Does it follow?
- likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
- Do the premises hold?
- 1 not separately assessed
Granting that OpenAI made the statement and that it scans its RL runs for accidental CoT grading, the conclusion is much more likely, though not guaranteed. The scanner's coverage and sensitivity are not public, and the developer is an interested party that itself hedges its statement ("fairly confident"). The argument rests mainly on the scanning premise, which is publicly attested by OpenAI staff but not independently verified. It is not decisive.
Provenance
Where this claim has been said, linked to its canonical form.
We are fairly confident that there are no significant direct CoT grading pressures.
Discussion of causes of Astra's decreased CoT monitorability.
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 · Nov 1, 2026. Every judgment on this page is accompanied by a reasoning trace.