LTL-UI combined annotation browser

Read-only public snapshot of the latest30 FP-fix 2,288 candidates and historical cross-replay 484 candidates. Entry point: index.html. No build dependencies.

All 2,772 rows retain their notes and dataset-scoped IDs (R:X… and H:X…). On 2026-09-06 the user confirmed reviewing all 14 unsure cases and classified them as TP. Each changed row records that confirmation in humanReview, retains its prior decision, and sets humanConfirmed: true. Existing Duplicate groups are retained. The 493 representatives now comprise 424 TP and 69 FP (0 unsure). The old 2,358-case development baseline is not included.

All latest30 rows include their existing AI notes verbatim. Historical notes are empty in the canonical source; do not invent historical annotation provenance. The 69-case AI mechanism-review supplement includes 28 historical FP cases and 41 latest30 FP cases. It is shown separately and does not change their labels. The other 456 historical rows have no notes in these sources.

Latest30 non-Duplicate representatives include one exact current-rerun trigger before/after screenshot pair each (250 pairs, 500 images). The existing after images were checked against their exact trace entries before the preceding trace frame was added as the before image. Duplicate rows preserve trigger references and link to their canonical representative; that representative’s screenshot is not claimed to show the duplicate row’s action. Historical before/after evidence uses existing relative assets in ../ltl-ui-cross-replay-error-review/. Screenshots show visual state, not screen-reader speech or reading position.

The replay table and filter cover both sources: 484 historical + 2,288 latest30. Each run counts rows whose replays include that run. Pass@3 counts all 2,772 scoped IDs once. Correctness is Correct / (Correct + FP), excluding Duplicate. summary.json records this combined breakdown as well as the historical-only breakdown for reference. The table is generated from the current case labels.

cases.json is the downloadable snapshot, schema ltl-ui-combined-error-review/1; it is not an import for either older review UI. summary.json contains scope, provenance filenames, and counts without cases. This page does not read, mutate, or overwrite other review pages’ localStorage.

Canonical inputs (annotation snapshot 2026-09-05):

• ltl-ui-cross-replay-review-2026-09-03.json
• ltl-ui-latest30-fpfix-rerun-blv-review-2026-09-05.json
• ltl-current-69-fp-source-review-2026-09-05.json

Checker commit: bfb4996a. Machine-local paths are omitted from metadata; notes are preserved. Only the 14 user-confirmed unsure labels were changed in this update.

Run node verify.mjs cases.json to check the scoped IDs, labels, replay filters, Pass@3 counts, and all latest30 screenshot pairs. Passing the previous published cases.json (722e4eb4) as a second argument also verifies that only the 14 user-confirmed annotations and their summary metadata changed.