# Independent pass2 N1 / RM evidence audit

**SUFFICIENT for N1 and RM within the stated Linux/Xvfb scope.** No recapture requested. This is an evidence verdict, not a code-review or merge-readiness verdict.

Audited [#24618](https://git.cafe/pingdotgg/t3code/pulls/24618), BEFORE actual merge-base [8bcef34f](https://git.cafe/pingdotgg/t3code/commit/8bcef34f16ca2a08aa19c35dd64b71290bd4412f) versus AFTER [b7fddd87](https://git.cafe/pingdotgg/t3code/commit/b7fddd87a69bb9f868bfd54928aaf4ae881c9432). The historical c4fa baseline instruction is superseded. Prior delivered/reused rows were excluded.

[Final capture index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/afb207c5231a237e/capture-index.md) · [Full original evidence bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/563b4a672a2aed3c/evidence-N1-RM-24618.tar.gz) · [Independent reproduction bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2bf9abe4b9e3b4ac/reproduction-evidence-24618.tar.gz) · [Reproduction native metrics/trace](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/303e9d6ab142c3ca/reproduction.json) · [Independent comparison calculations](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3068099aa0455885/independent-checks.json).

## Row verdicts

- **N1, zoom 1: SUFFICIENT.** Real Electron webview, 1440×800 inner viewport, DPR2, scroll600. The original BEFORE snapshot visibly mirrors the blue left source through the right side of the glass. AFTER retains the red right source and correct stripe/label placement, agreeing with the independent headed reference. Both server snapshots are 1280×711; the 500ms repeat is byte-identical in each phase. Before/reference RGB mean absolute error 37.90510; after/reference 1.57083. [Before/after/reference and diffs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5cae38d1a15f3c83/review-N1-z1.png).
- **N1, zoom 1.25: SUFFICIENT.** Same genuine native route, inner1152×640, DPR2.5, native zoom1.25, scroll600. BEFORE again samples the wrong backdrop region; AFTER has correct blue/red composition and glass positioning at the changed zoom. The matching headed reference has the same measured zoom/DPR/layout/scroll. Server snapshots remain1280×711 with byte-identical repeats. Before/reference mean absolute error36.28044; after/reference1.56720. [Before/after/reference and diffs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/022551ee906444cb/review-N1-z125.png).
- **RM: SUFFICIENT.** The fixture contains real moving stripe animations and a prefers-reduced-motion media rule. Under no-preference, both commits report matchMedia=false, three running stripes-move animations and changing currentTime/transform. BEFORE transform translation advances41.3253→77.3406; AFTER9.33448→49.3316. Actual Emulation.setEmulatedMedia(reduce) commands then precede the server snapshots on the same attached debugger; both phases report matchMedia=true, animationName=none, zero animations and identity transform, with no intervening media reset. Stable1280×711 before/after server outputs show the blue-only wrong backdrop before and correct blue/red blur after, matching the headed reduced-motion reference. Before/reference mean absolute error41.51071; after/reference1.58655. [Before/after/reference and diffs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/86376791d02b549c/review-RM.png).

## What was independently checked

All nine unique raw server/reference captures were opened individually; all12 server original/settled captures, six reference raw/settled captures and18 difference images were independently decoded and visually inspected on labeled artifact sheets. All normalized images were independently decoded and compared to freshly computed normalization. No reused-row images were included. All nine numerical comparison objects and18 diff pixel arrays were recomputed independently and match exactly. Each settled server/reference capture equals its corresponding original bytes.

The public original bundle was downloaded independently and verified against SHA256 `7d202a364dbd6f2208f694e4e6cb17794c179ee49ecbb2057895f065939da054`; all103 payload byte counts/hashes match the published manifest. The final public index was downloaded and matches the delivered local index. Both server source-file hashes were checked directly against Git objects at their stated commits. Desktop source is unchanged between those commits; AFTER source and built-server hashes match the manifest.

Native attachment is established by real product-created webview id3, fixture URL, attached debugger, thread/tab DOM identity matching actual tool tabIds, and a guest Page.captureScreenshot within each of the12 final server snapshot request intervals. BEFORE uses scaled clip parameters (including y600/y750 for the two N1 states); AFTER uses format:png without clip. The actual output paths point into each isolated desktop backend's browser-artifacts directory. The separate web dev backend was not the snapshot source.

Both raw X11 recordings (48.2s BEFORE,150.95s AFTER) and both final phase-labeled exports (17.8s/56.7s) were opened through decoded frames across their duration, including additional frames around the motion-control interval. Both exports fully decode without error. They support interaction continuity and viewer behavior; exact target media/animation state is established by debugger/computed-state logs, not approximate video wall-clock labels.

## Fresh independent actual-server reproduction

Downloaded the public fixture generators, verified SHA256 `23d61fca95c72aef91f6b330522698e2a61f5d3bbe84ef81b96842ab9e98b74b` and dependency `1abdeee1085e1e3bd5b5280b756bb2f67d9e370410bb374442c36cb86c4eb82d`, then regenerated all3 HTML pages byte-for-byte. The blur page hash is `21eea06840eadcb8ea2c090d6e423514f04ce59b8c7c6a060800346590fac32a`.

After the capturer confirmed teardown, the auditor started fresh isolated dev/desktop homes under `/tmp/ev-24618/audit/runtime`, no seed, no VITE_HTTP_URL/WS_URL, and an owned HTTP server serving the rebuilt page at port18619. A fresh synthetic app project/thread and normal provider turn supplied the supported thread-bound MCP capability. The synthetic provider only forwarded allowed preview tool requests; the auditor controlled native geometry/media and artifact verification. No source edits, direct snapshot-helper import or fake relay were used.

Auditor-controlled N1-z125 reproduction used actual preview_status, preview_navigate and preview_snapshot(save=true). Fresh native tab `tab_1_56f893a7-05ab-4d8c-9456-1494a008cb73` maps to the real attached webview id3. Before and after the snapshot: inner1152×640, DPR2.5, zoom1.25, scroll600. The fresh trace records Page.captureScreenshot({format:png}) on that guest during the actual server call. The returned1280×711 PNG SHA256 is `ba3996322f3d4ef8ddf7c162b7b6f4cc2c3d381edcf113115e5dcf888a634ff3` and visibly reproduces the correct after composition. Mean absolute RGB difference from capturer AFTER is0.61136; from normalized headed reference1.79141. Independent runs are not claimed byte-identical.

The synthetic bridge initially failed to escape its redundant text JSON block; the audit retained the raw response and independently decoded the valid structuredContent member. The bridge later repaired that serialization. This did not replace the actual tool invocation, guest capture or server artifact. Raw transport material and the audit script are in the reproduction bundle.

![Independent N1-z125 actual-server reproduction](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/453b3ec7b8036473/reproduction-N1-z125.png)

![Capturer AFTER, independent reproduction, and headed reference, labeled top to bottom](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/68ec19f56fd68b34/reproduction-comparison.png)

## Body-ready limitations

> Evidence covers a genuine headed Electron native attachment on Linux/Xvfb; physical monitors, macOS and Windows were not tested. Reduced motion was configured through the attached guest's real CDP media emulation before actual server snapshots; this does not establish a public reduced-motion configuration API or separate headless-server RM coverage. Headed references were captured at2880×1600 and normalized to1280×711 with Pillow LANCZOS; edge, font and resampling residuals remain, so pixel identity is not claimed. X11 recordings show the product viewer, which may lag target commands; their wall-clock alignment is approximate and they do not establish exact display latency. This supplemental audit covers N1 and RM only and makes no new claim about prior delivered rows, test-suite status or merge readiness.

No source changes, PR comments/edits, watch, push or merge. No live T3 userdata accessed. Worktree remains clean at AFTER. Auditor-owned processes were stopped using captured PID/start-time identities; scratch and worktree were retained for parent cleanup. Model/harness: GPT-6.1-Sol through Codex; synthetic provider used only as the supported isolated MCP transport fixture.