# Proof audit, pass 2

Read-only audit of [#24611](https://git.cafe/pingdotgg/t3code/pull/24611), branch `takeover/pr-13413`, head [43cb239](https://git.cafe/pingdotgg/t3code/commit/43cb23920cf601caaf146c75477abc5dfee95cb2) against merge-base [843f4e3](https://git.cafe/pingdotgg/t3code/commit/843f4e35292391897089535e8f1af59b322e70f8). Local refs and refreshed GitCafe metadata match. Inspected the complete two-file diff. Audit completed within the 40-minute budget.

## Findings first

1. **Missing required native proof:** M-Android-L and M-Android-D have no originals or diffs. Native persisted T50, compact notice M-C, expanded notice M-X, isolated command M-S and E200 remain unverified. This is the disclosed, expected infrastructure gap; no native build or device lease was attempted by this auditor. It blocks a SUFFICIENT verdict for the complete six-scene matrix.
2. **Independent application recapture remains unmeasured.** Used the lead-authorized diff-only fallback. The VM reports 5G available, the brief's disk floor. The available Browser panel reports Chrome/156.0.0.0 at DPR 2; the delivered environment is Chromium headless-shell 156.0.8078.70 at DPR 1. No fresh server, seed run, native build or alternate browser was launched. Regenerated all eight diff images from the delivered originals with the uploaded, unchanged diff.py: every regenerated image is byte-identical, with **0 changed pixels and maximum channel delta 0** against its delivered counterpart. This is a diff-artifact rebuild, not a claim of an independently recaptured application screenshot.

**No contradicted claim or defective image found in the four accepted web pairs.** All eight full-resolution originals, eight diffs and four sheets were opened and visually inspected. The visible changed regions are T50/P50 replacement glyph removal and neighboring dots, including P50's breadcrumb at 1440. E50/SHORT retain their emoji. No duplicated/ghosted elements, spinner, skeleton, stale popup, missing target text or unrelated diff illumination was found. The narrow scene intentionally uses the matrix's real open sidebar drawer: obscured chat content behind that drawer is not a claim of chat-layout coverage. The empty central chat area in the wide scene is ordinary whitespace around the short completed exchange.

## Row table

Counts are source pixels with any RGB channel delta >0 / >16. Diff0 is per-channel absolute difference amplified eightfold and clipped to 255. DiffT maps each channel to 255 iff its absolute delta is >16/255. Both are unmasked full surfaces.

| Scene | Viewport / theme | Visible proof | Diff pixels >0 / >16 | Status |
|---|---|---|---:|---|
| W-1440-L | 1440×1000, light | T50/P50 sidebar, P50 breadcrumb, E50/SHORT controls | 427 / 390 | SUFFICIENT |
| W-1440-D | 1440×1000, dark | Same | 426 / 401 | SUFFICIENT |
| W-390-L | 390×844, light | T50/P50 and E50/SHORT in open drawer | 280 / 259 | SUFFICIENT |
| W-390-D | 390×844, dark | Same | 279 / 258 | SUFFICIENT |
| M-Android-L | No device viewport acquired | No native proof | — | INSUFFICIENT: missing |
| M-Android-D | No device viewport acquired | No native proof | — | INSUFFICIENT: missing |

ROW: W-1440-L — SUFFICIENT; four images inspected; correct viewport/theme/revisions; diff-only rebuild delta 0.

ROW: W-1440-D — SUFFICIENT; four images inspected; correct viewport/theme/revisions; diff-only rebuild delta 0.

ROW: W-390-L — SUFFICIENT; four images inspected; correct viewport/theme/revisions; diff-only rebuild delta 0.

ROW: W-390-D — SUFFICIENT; four images inspected; correct viewport/theme/revisions; diff-only rebuild delta 0.

## Claims and provenance

- The evidence section's before and after SHA links match the requested full base/head hashes. All PNG dimensions match their stated CSS viewport at recorded DPR 1. The archive's final environment records agree on browser version, viewport, theme, reduced motion, loaded fonts and frozen Date. Each delivered original exactly matches its final archived frame; the archived a/b settling frames have zero pixel delta. This corroborates the settling transcript, independently of the visual inspection.
- **“Nothing else changed ... at zero threshold” is supported.** Independent full-surface arithmetic matches the delivered metrics. Every changed pixel is in visually inspected title regions: T50 sidebar 132/131 pixels in light/dark, P50 sidebar 148, and wide P50 breadcrumb 147. Changed pixels outside these regions: **0 in every scene**. Ages, controls, provider reply, plan body and chrome have zero pixel delta wherever visible. Narrow scenes do not independently show the obscured provider reply or plan body.
- **Normalization did not change the fixture titles.** Read normalize.py and all eight generated SQL files. Scalar assignments target timestamp columns and payload_json; the title TEXT column is never assigned. JSON normalization replaces ISO timestamp values recursively, then sets branch and settled/visited metadata. None of the title strings matches that ISO pattern. Decoded SQL payload titles match public provenance exactly: T50/P50 base retains U+D83D and has 53 UTF-16 units; head omits it and has 52. Their 44/34 U+200B units are unchanged. E50 and SHORT are identical across revisions. The head's isolated event history independently contains the same creation titles before projection normalization. The base fixture database is no longer present, so base history was not independently reread; base SQL/provenance and the visible replacement glyph were inspected.
- Public provenance contains actual user messages, completed runs, the proposed plan and destination sourcePlanRef. Raw input shapes and title values agree with the unchanged caller paths and changed shared boundary check. The unavailable text-generation model deliberately exercises fallback retention; successful regeneration is not proved. The adapted seed-web.py recipe is explicitly disclosed as not executed end to end. No claim of a fresh seeding execution is made by this audit.
- Matrix logical rows T/P and web C controls are visually covered by W. F/L and other caller variants remain labelled caller-probe/source coverage, as the matrix allows; no real attachment-only or context-only send is claimed. Native T, M-C/M-X/M-S, E200 and native V rendering are missing. LIVE/interaction timing and OUT are explicitly excluded. No recording rows are claimed or required for these prepared static scenes; there are no recordings to inspect.
- No screenshot in this set makes a chip/ring/transition claim. Historical evidence retained in the current PR body uses other environments and is not substituted for this pass's requested four-scene evidence.

## Example inspected scene

Before and after are labelled within this unmodified delivered sheet; full-resolution originals remain in the evidence index.

![W-390-L: before at merge-base, after at requested head, zero-threshold and threshold-16 diffs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5910d8b6cee06005/W-390-L-sheet.png)

## Cleanup and scope

No edits, pushes or comments. No application server, emulator or lease started. The single Browser-panel tab opened to inspect renderer metadata was closed. The lead's worktree and scratch root are left intact; only this auditor's subdirectory is removed after upload. No live-userdata access. Model/harness: GPT-6.1-Sol through Codex in T3 Code.

FRICTION:
- infra | VM remained at 5G available, the mandatory floor; used authorized diff-only fallback and did not start a build. | More free disk and a prebuilt native client/cache.
- tool | Browser panel reports Chrome/156.0.0.0 at DPR 2 rather than delivered headless-shell 156.0.8078.70 at DPR 1; a matching-renderer application recapture was not performed. | Explicit browser executable/DPR control or a provided compatible capture session.
- tool | Python urllib downloads received HTTP 403; curl downloaded every artifact successfully. | Use curl for this artifact host.

VERDICT: INSUFFICIENT M-Android-L, M-Android-D (native persisted T50, M-C, M-X, M-S, E200)