# Read-only proof audit, pass 2

PR [#24008](https://git.cafe/pingdotgg/t3code/pulls/24008), final [f217759](https://git.cafe/pingdotgg/t3code/commit/f21775904c4e1c647a0b4279f752013208694bb2), base [4812dca](https://git.cafe/pingdotgg/t3code/commit/4812dcad02a86874533540e6a0c4b7f5eaea297a).

The replacement evidence resolves the six wide comparison failures and the earlier independent settling failure. All twelve consolidated visual cells are sufficient within the explicitly synthetic web sequencing scope. L1 and S1 retain the prior accepted records; no tests or code review were performed in this pass.

Audited records: [original requested matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6c690f778d80f687/matrix.md), [prior audit](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3eff83de75c3c78b/audit-report.md), [consolidated index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/53acd20fa0a18eaf/consolidated-index.md), [replacement rebuild bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3e71816cdc797fa6/wide-evidence-supplement.tar.gz), and [historical recording index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/efea767beb4899d9/public-index.md). The current PR description was read through GitCafe and its head matches the pinned head. Its historical incomplete status and old wide diffs must be explicitly superseded by the parent's append; this audit made no PR edits.

## Row verdicts

Counts below are independently recomputed, full-frame, unmasked RGB max-channel comparisons. Thresholds are strict >0 and >16; no small delta is waived. Every consolidated before/after temporal pair was opened, included in a labeled visual inspection sheet, and independently compared. Both delivered amplified diff files for every row were opened and inspected. All replacement diff files additionally match an independent recomputation byte-for-pixel. Superseded old wide still pairs/diffs were also inspected; their known card mismatch remains rejected as visual parity proof.

| Cell | Source | Pixels >0 | Pixels >16 | Max channel | Verdict |
|---|---|---:|---:|---:|---|
| A1-1440-light | replacement | 74124 | 11621 | 213 | SUFFICIENT |
| A1-1440-dark | replacement | 74077 | 9953 | 235 | SUFFICIENT |
| A1-390-light | reused | 72207 | 9812 | 213 | SUFFICIENT |
| A1-390-dark | reused | 72157 | 9628 | 235 | SUFFICIENT |
| C1-1440-light | replacement | 0 | 0 | 0 | SUFFICIENT |
| C1-1440-dark | replacement | 0 | 0 | 0 | SUFFICIENT |
| C1-390-light | reused | 0 | 0 | 0 | SUFFICIENT |
| C1-390-dark | reused | 0 | 0 | 0 | SUFFICIENT |
| C2-1440-light | replacement | 0 | 0 | 0 | SUFFICIENT |
| C3-390-dark | reused | 0 | 0 | 0 | SUFFICIENT |
| C4-390-light | reused | 0 | 0 | 0 | SUFFICIENT |
| C5-1440-dark | replacement | 0 | 0 | 0 | SUFFICIENT |
| L1 | retained accepted test/provenance record | — | — | — | REUSED SUFFICIENT |
| S1 | retained accepted source audit | — | — | — | REUSED SUFFICIENT |

Every one of the 24 consolidated side temporal comparisons has zero differing pixels. New side timestamps span 1059–1319 ms; historical records require at least 550 ms. A1 changes are the bare command/additional boundary removal, instructed-message movement and corresponding minimap change. New wide A1 deltas are entirely inside physical bbox [24,480,2176,970]. Controls are exactly equal at threshold zero. No unintended overlay, hover ghost, loading spinner, clipped message, blank final surface, theme mismatch or viewport mismatch was found in the consolidated stills. Narrow title truncation is normal responsive layout; it is not lost message content.

C5 replacement stills show the completed Never opt-out stage. Reused C5 recordings establish restoration to When you return and the subsequent automatic-predecessor send. The two stages are explicitly distinguished in the consolidated index; a restored-stage still is not being substituted for opt-out parity.

## Independent affected capture

After inspecting run.py, capture.cjs, prepare.py, status.py and equal.py, ran the exact requested command with fresh `/tmp/ev-24008/auditor-homes` and separate `/tmp/ev-24008/auditor-out`. Runtime used `T3CODE_DEV_SEED=0 vp run dev --no-seed --home-dir ...`. Synthetic fixture only; no live database was read, copied or opened.

The source observation is clean exact head, tree `fd5d9adb09c04161e4e23b090993e898602cc9d1`. The rebuilt real composer sent only `/compact keep the API notes`. Sanitized peer observations show seeded resume `cc7276b2-180b-4420-b6c2-7867f180e64d`, started/result-success/completed command lifecycle, then persisted completed run and idle composer.

Two independent 2880×1800 screenshots were captured at real Unix milliseconds 1792754812777 and 1792754813920, **1143 ms apart**. Temporal RGB differing pixels: **0**. Rebuild versus delivered `A1-1440-light-after-1.png`: **0 pixels >0; 0 pixels >16; maximum channel 0; no delta bbox**. Both rebuilt frames and both resulting amplified diffs were opened and inspected. No AA or non-AA waiver was used. This replaces the prior failed 509-pixel temporal rebuild as the current independent proof.

Matched environment: Chromium 148.0.7778.96 on Linux x86_64, CSS1440×900, DPR2, desktop pointer, no mobile emulation, en-US/UTC, reduced motion, light color scheme, frozen browser Date 2026-10-23T12:00:00.000Z, document fonts loaded (both manifests report no custom FontFace entries), same system-font environment, network idle, zero timeline scroll, neutral pointer, blurred focus, sidebar/details closed through actual controls. Screenshots are uncropped and unmasked. Motion/caret suppression does not hide content. Browser UA, platform, dimensions, DPR, font status/list, theme, frozen date and version match the delivered manifest; exact whole-frame equality independently confirms the rendered environment.

## Recordings and narration

All 24 original MP4s were opened in the isolated Chromium player and played at normal rate through an ended event; the audit manifest retains duration, currentTime and timeupdate records. Whole-duration timestamped frames and final frames were inspected for every file, with additional dense extraction around C4 toggles and C5 settings changes. Narration is burned-in step text, not a voiceover. `recording-timestamps.md` records each nominal UTC-derived action offset and exact command sequence; `recording-audit.json` retains captions and sample times. Initial captions can have negative nominal offsets because they were installed before recording began. Variable-frame encoding makes these nominal offsets approximate; the actual sampled settings/toggle frames were checked directly.

All 24 raw sanitized protocol traces independently match the listed commands and seeded resume IDs. A1 base has bare then instructed compact, while head has instructed only. C1 has bare then ordinary text on both sides; C2 uppercase input is normalized to one bare compact; C3 has bare then literal prefix, with the synthetic peer response displayed. C4 and C5 have ordinary-only opt-out sends, then bare-plus-ordinary after restoration. Completion/result records and persisted run milestones corroborate the visible idle endpoint.

C4 BEFORE shows Full at 0.7–1.0 s and the restoration toggle at 23–24 s; AFTER shows Full at 1.0 s and the return to Compact at 24–25 s. C5 BEFORE shows Never at 5–6 s, restoration selection at 23 s and When you return at 24 s; AFTER shows Never at 7 s and When you return at 20 s. Both subsequently show the restored Compact send and completed sequence.

Wide historic recordings retain their source-card mismatch and prove sequencing only. C4/C5 also visibly show historical checkpoint file cards and reconnect/loading states during the explicitly narrated immutable-seed reset; these transient fixture-reset stages are not final parity captures. Replacement wide stills provide the matched visual proof. No old video has been relabeled as matched-state proof.

## Provenance, limits and release

The local origin is GitCafe. Read-only GitCafe PR metadata reports the exact requested head and observed base. Local `git merge-base origin/main <head>` returns the pinned base. Base tree is `811b2f938f2e6ebef64595507ccaffb38a4a839b`. Replacement harness switches the entire source checkout, not individual changed files; dependency lock and root/server/web manifests are unchanged between refs. Ten new side manifests directly observe their clean source; the two early A1 BEFORE sides instead rely on the disclosed pinned runner provenance. The independent head rebuild reproduces the delivered affected capture exactly. The reused narrow captures and old recordings retain their disclosed two-file baseline substitution method; their prior acceptance is not recast as a full-base runtime.

Integrity audit: all 78 replacement bundle content hashes and all 285 historical bundle content hashes match; no files missing. The downloaded published replacement tarball matches the local tarball exactly: SHA256 `bf43de051e58d2012dd70202fdce9523d6815be658a39c15072826415c436b28`. The retained L1 logs report final 45/45 passing and base with final tests 44 passing/1 intended regression failing. Those are supplied records, not fresh execution. S1 and prior source-review findings are reused without a new code review.

Scope is actual responsive-web composer/server/Claude SDK transport with an offline synthetic peer. Compaction notices, token counts and response content are fabricated fixtures. No actual Claude summarization, transcript restoration, billing, unknown-command parser acceptance, native mobile, Electron shell, remote or tunnel proof is claimed. Desktop's shared web path and other-client/provider gate analysis remain the prior S1 record.

The capture index discloses an initial unused Browser-panel artifact written into app-owned live browser storage during the supplement task. It was not read, copied, used or published for that supplement. This pass used the explicitly requested isolated harness and local delivered evidence; **no live application database was used**.

Playback friction: the first player process closed after 16 successful end records (`page.evaluate: Target page, context or browser has been closed`). The unfinished eight files were replayed individually; only successful complete ended records count. No evidence recapture or source repair was performed beyond the one required independent rebuild.

Owned dev PID/group 1969659 was stopped by the supplied runner's captured-group finally block. The group no longer exists, active PID file is absent, and owned ports 7899/15939 have no listeners. Worktree remains clean at exact pinned head. Worktree and scratch are retained for parent cleanup. Disk remained approximately 15 GB free, above the 5 GB friction threshold. No ten-minute silent shell, tests, code changes, commits, push, PR writes, watchers, other machines, other worktrees, or delegation occurred.

Published audit manifest and independent proof (raw temporal pair, deltas, matched-environment record, sanitized sequence, runtime release, full playback/end records and timestamp index):

[independent-proof.tar.gz](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c852eef06b105310/independent-proof.tar.gz)
[audit-manifest.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/887d3434ddd0ee2c/audit-manifest.json)
![A1-1440-light-after-1.png](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/06591cd60bf3d15e/A1-1440-light-after-1.png)

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

VERDICT: SUFFICIENT