GPT-6 Astra's reduced chain-of-thought monitorability is primarily due to capability gains rather than architectural or training changes.
9 events · 4 assessments · 4 decisions
Reassessed no change
The controllability-capability correlation subclaim was first assessed as supported (0.62/0.75). That matches this claim's prior treatment of it as plausible but uncontrolled developer testimony, so the change is not material. Status (inconclusive 0.65), credence (0.45), importance and contestation are unchanged. The for-argument evaluation was updated to reflect the premise statuses, and the against-argument evaluation was confirmed. Since nothing material changed, no dependents were notified.
Reassessed: still Inconclusive
verdict confidence 0.65 · credence 0.45
Reassessed no material change
The excess-over-trend subclaim's first assessment (inconclusive, 0.5) matches the prior reading of that premise as weak. Status, confidence and credence are kept: inconclusive, 0.65, 0.45. The against-argument's evaluation was updated to reflect its premise's new status, and the for-argument's evaluation was re-confirmed. No notification was sent, because the change is not material.
Reassessed: still Inconclusive
verdict confidence 0.65 · credence 0.45
Reassessed no status change
The recurrent-depth subclaim's first assessment (supported, 0.55) confirms how the prior pass already treated that premise. The change is not material: status stays inconclusive, confidence 0.65, credence 0.45. I updated the for-argument evaluation to reflect the premise's new standing, reconfirmed the against-argument, and kept importance at 0.55 and contestation at 0.6. No dependents were notified because nothing changed materially.
Reassessed: still Inconclusive
verdict confidence 0.65 · credence 0.45
Structured and assessed
First pass. Decomposition: assumes the Astra decline (existing claim 33a06089). The for-argument is root-cause elimination, with new subclaims on direct CoT pressure (contradicts, seeded 0.15), recurrent depth having a minor effect (requires, 0.65), and controllability tracking no-CoT capability (supports, 0.75). The against-argument is the unexplained excess decline, with new subclaims on decline exceeding trend (0.55) and controllability rising over RL unlike earlier models (0.85). Related links: the rival explanation that training practices degrade monitorability (74c2c8af), and the claim that CoT length tracks monitorability (4c7a1288). Delegated a reading of the system card ($0.78) and verified the key passage myself. Recorded instances from Korbak (affirms) and Zvi (denies, conf 0.6). Assessed inconclusive 0.65 with credence 0.45. Importance 0.55, contestation 0.7. Notified the parent.
Assessed Inconclusive
verdict confidence 0.65 · credence 0.45
The attribution comes from OpenAI. Its researchers say GPT-6 Astra's lower chain-of-thought monitorability stems from "a jump in intelligence", not from architecture or direct pressure on the trace. The system card itself is more cautious. It says OpenAI is "actively investigating the causes" of Astra's rise in CoT controllability. It says OpenAI is "fairly confident" there was no significant direct CoT grading pressure and "quite confident" the change is not differentially due to architecture, meaning Astra's limited recurrent depth has only a minor effect. And it reports that controllability gains track no-CoT capability gains. The case is elimination of suspects plus a correlation. No ablation, and no comparison that holds capability fixed, has been published. Two things keep the attribution from being established. First, the developer itself reports that Astra's controllability rose during RL training, unlike earlier models'. That is a change in training dynamics. When more or different RL both raises capability and lowers monitorability, "capability gains" and "training changes" are not separable causes, so the claim's either/or partly dissolves. Second, outside analysts, notably Zvi Mowshowitz, argue that the decline exceeds what capability trends predict. They point to recurrent depth, indirect pressure, and pretraining data full of discussion about CoT monitoring as candidates for the residual. That argument rests on extrapolating trend lines across very few data points. Astra's no-CoT horizon estimate may also be inflated by contamination, so it is suggestive but not decisive. The card's own breakdown shows the proximate mechanism: most of the measured decline comes from shorter, often empty chains of thought. Residual gaps persist at matched length in a few environments, such as destructive actions and patient-opinion health queries. That is compatible with both stories. Capability is very likely a large contributor. Whether it is the primary one, as opposed to an inseparable part of a changed RL regime, is not settled by the published evidence. A root-cause publication from OpenAI would resolve it, as would independent replication using controlled comparisons: controllability against RL steps at fixed capability, or a run without recurrent depth.
Claim entered the graph