# Proof audit pass 2 — read-only

Issue [#20883](https://git.cafe/pingdotgg/t3code/issues/20883), PR [#24313](https://github.com/pingdotgg/t3code/pull/24313).

Reviewed exact Before [c79ef8c454](https://github.com/pingdotgg/t3code/commit/c79ef8c454e1268ff79da5a9dcb74572e86f7861) and After [23cff4333e](https://github.com/pingdotgg/t3code/commit/23cff4333ed7ea1ff3a64a5b836761d1b06ddab3). Local integration HEAD equals full requested After SHA and is clean. Local git comparison independently confirms two changed files, 136 additions/3 deletions; production coordinator is the same 22-line ownership change prescribed by pass1. Reviewed the actual production diff. No new scope or product defect is asserted.

[Current index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e8fc96a7007dd7ad/index-current.md) and its linked latest README consistently name those full SHAs. The archive README is older; latest report supersedes it. Capture-time clean state and executing binary SHA are reported provenance, not independently attested by a runtime build stamp in the inspected event logs. Environment metadata is consistent with PNG/video dimensions: reported Linux Chrome154, CSS1440x900 or390x844, DPR2, scale1, root16px; desktop PNG1280x800 versus video2880x1800, narrow both780x1688. Browser zoom separately unmeasured. Date frozen, real timers live, differing focus/hover treatment disclosed. No desktop shell/native mobile execution claimed.

## Per-row status

| Cell | R1 | R2 | R3 | C0 |
|---|---|---|---|---|
|1440/light|Partial: full After first pre lost completion; separate pilot supplements first step, later selective steps delivered.|Partial: repeated ownership/action delivered; Before limited timing not certified.|Invalid pair: Before removal lacks required unrelated A completion witness; active-target milestone unverified.|Not established; changed witnesses/focus and residual differences.|
|1440/dark|Delivered selective sequence, timing/focus qualification remains.|Invalid Before witness expiry; After action settled post and limited missing.|Missing both phases.|R1-only comparisons; incomplete variants and unequal focus.|
|390/light|Delivered selective sequence, timing qualification remains.|Invalid After unrelated failure witness loss; action post/limited missing.|Before dismiss/expiry only, Settings remainder missing; After absent.|Partial; underlying R2/R3 coverage and stability fail gate.|
|390/dark|R1 only; independently LOOKED at first-step pre/post and decoded video samples.|Missing both phases.|Missing both phases.|R1 pre exact; full required variant coverage missing.|

## Confirmed evidence issues and visual inspection

- Missing rows are independently visible in archive filenames and event chapter labels: neither dark phase has R3; 390/dark has no R2/R3; 390/light After stops during R2. Eight cell videos are not eight complete matrix recordings.
- Opened original delivered PNGs with view_image: 390/dark R1 first post Before visibly retains A completion, A input and B approval; After removes only A completion. Opened corresponding full diff0 pre (black/exact) and diff8 post (completion silhouette only). This is positive visual evidence for that step, not structural approval.
- Opened Before1440/light R3-removal-pre: B completion is visible, required unrelated A completion is absent; an exiting/clipped failure card is not a substitute. Opened After390/light R2-1-post: no failure toast remains, so selective preservation cannot be established.
- Decoded both390/dark MP4s at24s and viewed frames: Before at B target retains all three; After at B target retains A input only. Confirms accessible recording content, not full playback or certified timing for all chapters.
- Independently recomputed all49 matched checkpoint and109 stability count records using raw RGB deltas: every count matches comparison.json;32 stability pairs exact,60 have no delta>8. These are stability counts, not49 valid acceptance pairs. Looked at actual diffs; counts cannot cure missing witnesses/chapters.
- Timestamp mapping needs correction before precise video timing claims: Before390/dark final event offset41.224s exceeds actual ffprobe duration33.888433s; After39.741s exceeds32.672033s. Similar approximately7s shifts occur elsewhere. Likely different origin for driver and recording; no truncation/product defect inferred. Supply an explicit mapping rather than treating offsets as direct media seeks.

## FRICTION and limits

curl downloads succeeded; no urllib used. One local summary script initially assumed every event had a label; corrected to optional lookup. Native media decoded through ffmpeg and inspected as local images; no browser launched. No server, browser, rebuild, tests, fetch, edits, push, comments, merge or watch performed. Rebuild deliberately not done: missing R2/R3 rows already establish bounded insufficiency, and rebuilding cannot manufacture those recordings. Live forge metadata was not refreshed because this is a fixed-SHA proof audit, not general merge readiness. PR registered with the thread. Report upload is the only external publication. Own downloaded/generated scratch removed after upload.

VERDICT INSUFFICIENT. Repair invalid/missing paired chapters and timing provenance on the requested SHAs; existing positive R1 evidence does not satisfy the full pass1 acceptance matrix.

Model/harness: GPT-6.1-sol / Codex in T3 Code.