Fix live-web false positives: view agreement, live hard negatives - #33
Merged
Merged
Conversation
Three coordinated changes derived from the published 2026-08-16 live smoke failure (docs/live-smoke-2026-08-16.md): 1. CF view agreement: a lone TTA view in [0.65, 0.85) no longer carries an AI verdict; it falls back to the runner-up view. Views at or above 0.85 keep single-view authority and early-exit. The adaptive band extends through the mid band so corroborating views always exist. Applied in the shared scoring fold and in the live predictAdaptiveViews path. 2. Probe retrained with 118 verified-real live CDN images (camera EXIF present in Unsplash API metadata; rows disjoint held-out guard) on top of the existing hard negatives. 3. Strong rescue tier tightened to [0.02, 0.20) with DINO >= 0.96. Also: harness and extractor ingest .avif (the format blind spot that hid live CDN bytes from evaluation), and the update check is skipped while offline. Verified with the implemented adaptive pipeline: bench 87.9 BA / 76.0 TPR / 99.8 TNR, stress 3/240, live guard 4/132 (3.0 percent, was 9.2 percent for v1.2.0 on identical bytes). Full suite 216 green.
benchmark-results.json now carries the verified live-guarded numbers (bench 87.9 BA / 76.0 TPR / 99.8 TNR, stress 3/240, live guard 4/132 at 3.0 percent) plus a liveGuard section, README and AGENTS.md are rewritten to the view-agreement policy, and the published findings doc gains a same-day fix section with the isolation results. Version references move to 1.3.0 across manifest, package, and store copy.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes the gap published in docs/live-smoke-2026-08-16.md and issue #32.
The fix (three coordinated changes)
Also: the harness and feature extractor now ingest .avif (the format blind spot that kept live CDN bytes invisible to evaluation), and the update check is skipped while offline.
Verified results (implemented adaptive pipeline, not simulation)
The 2.6-point bench trade for a threefold live FP reduction is deliberate: false positives on real photos are the visible kill criterion. Full suite 228 green, including new tests encoding the agreement contract; benchmark-results.json, README, AGENTS.md, and the site all update from the single source of truth.
After merge
Tag v1.3.0 (version references already bumped), then a full live smoke rerun of the built zip. No bounty claim until it passes.