# Recovered-nine independent visual judging

This directory packages fresh ordered viewport evidence for AG1, Ridge, HexClad, Gymshark, Jones Road, Grüns, Dr. Squatch, True Classic, and Allbirds. It is a read-only static evaluation of traditional PDP presentation. Source pages, routes, offers, checkout, and analytics are unchanged.

## Inputs and outputs

- `evidence-input.json`: exact selected sets, URLs, capture dates, device statuses, limitations, absolute frame paths, and SHA256 values.
- `SCORING-BRIEF.md`: byte-for-byte copy of the original fixed six-dimension rubric.
- `scoring-schema.json`: compact independent-judge output contract.
- `scores/craft/` and `scores/journey/`: isolated judge lanes; each judge records the images actually inspected.
- `validate_package.py`: validates the 222 source hashes and, when scores exist, exact inspection coverage, citations, score precision, weights, totals, and full-scope null rules.
- `build_report.py` and `report-shell.html`: build the review artifact with visibly separate original-eleven and recovered-nine cohorts and native-size evidence viewing.
- `publish_report.py`: defaults to a non-mutating plan and requires an explicit confirmation token for an immutable public revision.

Run `python3 validate_package.py` while scores are pending, then `python3 validate_package.py --require-scores`. Build a score-withheld viewer for shell QA with `python3 build_report.py --shell-only`, or a clearly labeled provisional viewer in the separate `report-site-draft/` directory with `python3 build_report.py --draft`. A final `python3 build_report.py` run validates both lanes and requires `synthesis.json` with `review_status: root_accepted`. Root accepted the synthesis and authorized publication of the immutable `recovered-nine-benchmark-20260913-r1` revision after final desktop and mobile browser QA.

The exact evidence states are `verified pass`, `verified failure`, and `unknown`. Offer and transaction facts not present in static evidence remain unknown. The zero-valued commerce-run offer tier is a schema-only placeholder marked `fact_state: unknown`; it is not an observed price or offer claim. `inspected_images` is a judge self-attestation and coverage record; mechanical validation cannot prove that a human or model actually viewed the pixels, so root spot checks and substantive review remain required. Heuristic presentation scores do not prove conversion lift or commercial truth.