# Proof audit — A6 replacement re-check

Findings:

1. **No remaining A6 sufficiency blocker. The independent full-frame mismatch is resolved.** The replacement head final `after-A6-final-b.png` is pixel-identical to the prior audit archive's independently rebuilt `proof-audit-2/rebuild/after-A6-fixed-final.png`: **0 / 2,617,920 RGB pixels changed (0.000000%)**, maximum channel delta 0, no pixels above luminance threshold 16. This is the entire native **1080×2424** frame, without crop, alignment, masking, or tolerance. Before/head, head/fresh-database head rerun, and each run's two final PNGs also have zero changed pixels. Both supplied Diff0 and DiffT regenerate exactly. [Independent measurements](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a39656f1c717b6eb/independent-comparisons.json).

2. **Minor narration precision correction, nonblocking for the required A6 behavior.** The baseline narration/subtitle says “Accept Gboard suggestion 'Hello'.” Video at approximately 5.75s does show `Hello` as the middle prediction, and the saved accessibility snapshot labels that target `Hello`. However, by approximately 6.25s the visible prediction has changed to `Geo`; the subsequent tap inserts `Geo` (visible by approximately 7.25s). The narrative should say “Tap the middle prediction, initially observed as Hello, now Geo.” The replacement section already discloses baseline `Geo \nCat` versus head/rerun `Hello \nCat`. This does not hide or omit the required suggestion acceptance, newline, full deletion, empty collapse/expand, or empty Submit, all of which the recordings prove. I accept the row's behavioral sequence while flagging the stale word label; I do not claim literal word-for-word narration accuracy.

   ![Baseline recording crops: Hello initially offered, prediction changes to Geo, Geo accepted, then a second line is typed](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c1ccd7ef5a81a9d1/baseline-prediction-sequence.png)

3. **The earlier inset explanation is consistent with the actual pixels.** New head versus the superseded delivered `after-A6-v2-fixed-final.png` still differs at **95,413 / 2,617,920 pixels (3.644611%)**; 95,407 are above y=400 and the remaining six are gesture-pill antialiasing. Shifting the old header down exactly **79 px** makes its entire 1080×110 band at old y=100–209 match new y=179–288, with zero changed pixels. The recapture's native StatusBar/SafeArea readings and AX geometry report **142 px / 54.095238 dp**; 142−79 = **63 px / 24 dp** at DPR 2.625. Thus a smaller old inset explains the placement mismatch. The old 63 px value remains an inference from images, not a runtime measurement from that earlier device; the exact overlay package remains unverified. The replacement's full-frame match to the independent rebuild closes the prior finding regardless of that unverified package name.

**A6: SUFFICIENT**, with the minor narration correction above. Before is correctly labelled [ac8dce0cc3](https://git.cafe/pingdotgg/t3code/commit/ac8dce0cc3abd8a337f2d806b5082a6834b9b81a); after/rerun are correctly labelled [f16d8770bd](https://git.cafe/pingdotgg/t3code/commit/f16d8770bda7acff940ed18a3557c050aea6651e). The checkout head and merge-base match the requested revisions. The replacement [A6 evidence section](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/41c3b2bfc5e1c8f2/a6.md) supplies native before/head/rerun recordings with audio and English subtitle tracks. All three narrated video streams match their archived raw video streams exactly.

The recordings show an empty multiline editor, native typing, a suggestion tap, Enter/newline and continuation, native Select all/Delete, an empty editor, collapse/expand retaining empty, and Submit. Each authoritative projection contains `initialAnswer:""`, `options:[]`, and a resolved request. Each Pi receipt is exactly `{"type":"extension_ui_response","id":"editor-1","value":""}`: empty string, not cancellation. Head and rerun preserve `Hello` followed by `Cat` on a second line before deletion. Baseline first loses the `l` during native typing, then accepts the changed `Geo` prediction; this intermediate difference is disclosed and does not affect the empty-state proof.

![New head final at left, prior independent rebuild at center, and black full-frame zero-difference image at right](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/001c519695ea91cf/full-frame-comparison.png)

Overall table, carrying forward the prior accepted rows as requested:

| Row | Status | Basis |
|---|---|---|
| A1 | SUFFICIENT | Prior pass retained. |
| A2 | SUFFICIENT | Prior pass retained. |
| A3 | SUFFICIENT | Prior pass retained. |
| A4 | SUFFICIENT | Prior pass retained. |
| A5 | SUFFICIENT | Prior pass retained. |
| A6 | SUFFICIENT | Required flow and empty receipt verified; new full final matches the earlier independent rebuild exactly. Minor stale prediction label in baseline narration noted above. |
| I1 | SUFFICIENT | Prior pass retained. |
| I2 | SUFFICIENT | Prior pass retained. |

Coverage: opened every supplied A6 PNG capture and diff, deduplicating pixel-identical copies (60 PNGs in the supporting archive; 32 distinct PNG contents across the direct downloads, supporting archive and selected prior-audit images). Also opened all 25 supplied review sheets. Independently used ffmpeg to sample all five distinct native recordings, including the two explicitly excluded setup attempts: 87 full-viewport samples at three-second spacing, plus 28 denser baseline prediction samples at four frames/second over 3–10s and exact-time detail frames. This is sampled recording inspection, not frame-by-frame review.

Visual observations: accepted final PNGs have no ghosting, stale toolbar, spinner/skeleton, clipped text, or unexplained unrendered region. They share the same light portrait viewport and inset. The large empty upper chat area is intentional fixture geometry and matches the independent rebuild. Native selection handles/toolbars appear during selection and disappear after deletion. During keyboard/collapse transitions, some video samples briefly clip or partially occlude the Submit button; the expanded resting card and final surface are complete. The rerun briefly shows Thinking/Working after Submit (including a roughly two-hour elapsed label under the fixed server clock), then resolves to Worked for 1ms; none survives into the accepted final captures. The provider-version warning belongs to the synthetic Pi fixture and is present identically on all finals. Clock/signal normalization, caret phase, predictions, key previews, pending-card pulse and transient submission state explain the sampled intermediate variations. No unexplained final difference remains. The excluded setup clips show Home and Choose project rather than the editor and are not counted as behavior proof.

Limits: exact 500 ms separation between native screenshot exposures remains unverified, as the replacement explicitly states; command starts are about 3–4.7s apart due to export/round-trip overhead. Both captures are identical on all three runs. No new independent device rebuild was performed in this re-check; the comparison uses the actual full-frame prior independent rebuild, extracted from the supplied prior-audit tarball. Other rows, current PR-description publication, CI and reviews were not re-audited.

Read-only cleanup: no edits to the checkout, pushes, comments, leases or servers. The checkout is clean at the requested head. All downloads stayed inside the specified audit scratch directory. Own scratch was removed after preservation uploads. [Audit artifacts, measurement script, input finals and sampled-frame sheets](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7b2784c824bfad8e/proof-audit-3-artifacts.tar.gz). Artifacts were published with ghdrop; no forge state was changed. Model/harness: gpt-6.1-sol through Codex.

FRICTION: none

VERDICT: SUFFICIENT