Keep TTA extras on busy pages; raise CF early-exit to 0.95

[felirami]

Why Unsplash still failed on v1.3.2

A live pass of the public v1.3.2 zip on the Unsplash featured feed scored 15 AI of 26 badges (ordinary editorial photos at 81-94 percent, raw p(AI) equal to neural). Wikipedia photos stayed clean. The published 3.0 percent held-out live-CDN guard did not match that feed.

Two production behaviors caused it. Neither remaps 0.65.

1. Center-only load-shed. ttaModeForLoad returned center when more than 12 images were pending. A 26-image masonry scored the first ~14 images on the official 440 center crop alone, so agreedMax never ran. A lone 0.81-0.94 center crop became the AI verdict. The 132-image guard was scored with harness --tta=adaptive, so it never saw this path.
2. Early-exit at 0.85. Even when adaptive TTA ran, a center crop of 0.85-0.94 stopped extras and kept single-view authority. The two FPs on the new stand-in were exactly this: livecdn-121 center 0.902 (runner-up 0.399) and livecdn-142 center 0.926 (runner-up 0.588).

Adaptive already skips extras on confident reals (center below 0.15). Shedding to center-only only changed behavior when the center crop was suspicious: exactly the band that needs extras for agreement.

Change

• ttaModeForLoad always returns adaptive. Busy pages keep view agreement.
• TTA_EARLY_EXIT moves from 0.85 to 0.95. A lone 0.81-0.94 crop must gather extras and survive agreement. Two agreeing high views, or a single view at or above 0.95, still carry the verdict.
• After an agreement fallback, DINO must be >= 0.96 to lift. A 0.76 probe on a 0.40 runner-up (livecdn-121) no longer re-flags the photo.
• DEFAULT_THRESHOLD stays raw 0.65. No Platt, no logit shift, no score remap.

Evidence (run in this checkout)

The published 893 / 240 / 132 / 100 fixtures are not in git. eval/fetch-live-guard.mjs built a stand-in: 39 held-out unsplash-lite rows (80-149) re-fetched as live imgix grid JPEGs (w=700 q=60 fit=crop) plus 16 DALL-E 3 images. Product and Flux sources returned zero files.

Offline comparison at raw 0.65, with harness DINO scores (eval/compare-tta-policy.mjs):

Policy	Live-CDN FP	TPR (n=16 DALL-E 3)	BA (n=55)
Load-shed (center only)	2/39 (5.1%)	25.0%	59.9%
v1.3.2 (early-exit 0.85)	3/39 (7.7%)	43.8%	68.0%
Production	1/39 (2.6%)	43.8%	70.6%

TPR did not move on this 16-image DALL-E 3 set. The remaining production FP is livecdn-145 (CF 0.020, DINO 0.994), the existing strong-tier rescue, not introduced here.

Removing load-shed does not change the harness adaptive path that produced the last published 893-image row (87.9 BA / 76.0 TPR / 99.8 TNR). Raising early-exit may still cost TPR on images with a lone 0.85-0.94 view; that fixture was not re-scored. A full live Unsplash re-smoke of this build has not been published. README and the site say so.

npm test passes.

Docs

STORE.md, eval/README.md, docs/POIDH-323.md, AGENTS.md, and README now describe the 0.95 agreement band, the no-load-shed rule, and the fallback DINO gate. README does not claim a live pass we did not run.