Interpretation with receipts attached
AI evaluation scores are easy to publish and hard to read honestly. This desk starts with the controls, carries uncertainty into the headline, and binds every analytical sentence to a sealed run.
Eval integrity / control failure
Meta Llama 3.3 70B Instruct passed all 144 prompt arms in the latest full-sweep run. The prior failed controls remain part of the record; the change is descriptive rather than causal.
Eval method / uncertainty
4 control-clean model runs observed 0 refused families out of 34, yet each zero result still carries a Wilson 95% upper bound of 10.2%. That interval is part of the finding, not fine print.
Model behavior / adjacent-run drift
Across 4 sealed endpoint runs, 3 of 163 paired family comparisons changed state: 0 toward refusal and 3 toward answer. This is a dated transition, not a trend claim.
Eval integrity / registry status
The current served chain contains 7 preregistrations and 620 runs across 7 model endpoints. Its hashes make revision detectable; they do not make the eval construct valid by themselves.
Publication standard
The prompt commitment and result are verified against the append-only registry before a sentence is written.
A failed ordinary control turns the article into an instrument warning. It cannot become a suppression claim.
Rates keep their denominator, interval, model label, and timestamp. Nothing becomes a standing leaderboard.
Every piece carries the strongest alternative reading, limitations, and the command needed to verify the chain.
Follow the record The prose is a route into the evidence, never a replacement for it.