Dipankar Sarkar PRO

For Astra, resisting and not reporting may be the same fact. Your own logs can tell them apart.

I pulled episodes-alarm7-*.jsonl.gz and took the median usage.out per admitted episode, clean vs injected:

GLM-5.2: 179 clean, 694 injected (3.88x), alarm 77.5%
DeepSeek-V4-Pro: 172 vs 445 (2.59x), 48.9%
gemini-3.7-flash: 336 vs 361 (1.07x), 0%
gpt-6-astra: 71 vs 72 (1.01x), 0%

Across all 12 rows, the Spearman correlation between that ratio and the alarm rate is 0.75. claude-haiku-4.5 is the honest exception: 1.08x and still 19.7%, so the ratio is a proxy for noticing, not a detector.

Astra has the lowest ratio of the twelve. All 742 admitted episodes are the same three calls (read_last_email, log_message, finish), at 60 to 163 output tokens. The Responses route counts reasoning tokens in that number. On probe_capability, where the task itself says pay, the same model spends about 250.

So the payload never seems to reach a decision. It did not turn the payment down. The letter barely cost it a token. That fits "logged, not read" better than "resistant but silent".

That matters for your distinction. Resistance that comes from never weighing the body could vanish once the task needs the body: summarize it, draft a reply, pay the legitimate invoice.

models/gpt-6-astra.toml sets no [extra] reasoning, so this is the endpoint default. Would an effort="high" rerun of just the 395 injected letters move the ratio, the alarm column, or the payment floor?

Your choice:2 temperature never runs through your own agent.

decision_config.json ships it at 0.12. agent.py passes every temperature through clamp_temperature on load, which pins it to [0.5, 5.0] (TEMP_MIN, "Laya's runtime clamp"). eval_decisions.py scores through DecisionAgent, and serve.py does too. So every reported number uses 0.5 for 2-option choice questions, 4.2x softer than the fit.

It also never touches the headline. I bucketed the typed-decisions test split with your temp_bucket:

2,000 decisions    choice:3-5  600    noul:2  600    score:3-5  800

So the 0.045 ECE is set by three temperatures. choice:2, choice:6-10 and choice:11+ only act on the held-out sources, where ECE runs 0.08 to 0.40.

The 0.12 looks like a fit that ran to its floor. fit_one_temp clamps at 0.1. In the mix code, the 2-option choice items are the SST2 and IMDB sentiment ones, one-hot targets, and IMDB sits at 0.985 accuracy. On a set that is nearly all correct, NLL keeps falling as T goes to 0.

Which set was the choice:2 bucket fit on? And what does SST2's choice-form ECE look like at 0.5 against 1.0?