# Independent proof audit — pass 2

Verdict: **SUFFICIENT for delivered S1 and P1 rows only.** No recapture required for those rows.

PR [#24611](https://git.cafe/pingdotgg/t3code/pulls/24611), head [43cb23920c](https://git.cafe/pingdotgg/t3code/commit/43cb23920cf601caaf146c75477abc5dfee95cb2). HEAD independently matches. Base-to-head diff is only String.ts and String.test.ts, 31 additions and 1 deletion. No code edits, revision switches, commits, PR writes, server/browser lifecycle changes, or live userdata access during this audit. Existing paired Chromium/Xvfb session used under explicit authorization, with agent-browser and ghdrop skills. Prior capture provenance is corroborated by artifacts and supplied history; this audit did not rerun the base app.

## Row findings

| Row | Verdict | Independent checks |
|---|---|---|
| S1 composer title seed and server persistence | SUFFICIENT | Opened send-before.png, send-after.png, send-Diff0.png and send-DiffT.png individually. Before shows replacement glyph in breadcrumb/sidebar; after clean dots; full prompt retains emoji and both replies show OK. Transcript inputs match exactly across revisions: 49 i characters + intact U+1F600 + identical harmless prompt. Before JSON/event title has lone D83D; SQLite has FFFD. After both title stores contain exactly 49 i + three dots. Real provider and checkpoint effects succeeded and runs completed. Independently reproduced through UI at head. |
| P1 plan implementation title | SUFFICIENT | Opened plan-before.png, plan-after.png, plan-Diff0.png, plan-DiffT.png, plan-source-before.png and plan-source-after.png individually. Both real proposed plans retain intact emoji and identical markdown. Implement + space is 10 ASCII units; 39 i characters put emoji at units 49–50, split by cut 50. Before implementation title contains lone D83D in JSON/events and FFFD in SQLite; after has 49 ASCII units + dots. Implementation prompts match across revisions, retain emoji, receive OK and complete. |

Source confirms distinct shared-helper consumers: runtime threadTitle.ts prompt branch and useProposedPlanSend.ts plan title. No shared-helper server caller is claimed. This proves client-generated title persistence through the server, not server-side execution of truncate.

## Independent image arithmetic

Recomputed from original full RGB viewport arrays without using diff.py. Both viewports are 1440×1000. Published Diff0 is byte-for-byte equal to clipped absolute channel difference ×8; DiffT is byte-for-byte equal to per-channel difference >16 mapped to 255.

- S1: 1,734 changed pixels, 1,059 above 16.
- P1: 315 changed pixels, 276 above 16.
- Before duplicate stills: zero changed pixels. After duplicates: 17 changed pixels, consistent with the disclosed caret blink. Published primary frames show no active caret.

The index’s unmasked-difference list matches the inspected images: S1 sidebar title rasterization, now/1m age and small submit shading; P1 title/dots, prior S1 title and 1m/2m age. Main prompt/plan body, intact source emoji, OK reply and layout match. Source-plan stills additionally show the implementation dropdown focus outline after; the index does not claim a zero-difference source-plan pair.

[Independent metrics](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/2f7e61fe6b7e2255/audit-metrics.json).

## Independent UI reproduction

Ran python3 /tmp/ev-24611/capture.py send --label audit-send-after --out /tmp/ev-24611/audit-reproduction against existing session ev24611 / CDP 19246. It created a new thread via UI, filled the exact harmless prompt and submitted. Read-only DOM reports h2 = 49 i + three dots, original intact emoji in prompt, and OK assistant paragraph. Read-only SQLite extraction confirms new thread a706a3b5-07d8-46ea-9afa-f5fca7024e39, same title in SQL and JSON projection, real user/provider events, completed run and successful provider/checkpoint effects. The extracted title-generation effect was still running at extraction; do not describe every effect in this independent reproduction as completed. The original delivered rows have succeeded title-generation effects.

![Independent head reproduction: clean title, intact prompt emoji and actual OK reply](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/92daadb2e29cbb5c/audit-send-after.png)

[Reproduction server transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8df67b21db9a7bc7/reproduction-transcript.json) · [UI command transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7895f8657f45ce45/driver-transcript.txt).

## Truthfulness and limits

The [current evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1cdeea8c73e45883/caller-evidence-index.md) is appropriately scoped. No blocking unsupported rendering or persistence claim found. Its deliberately unavailable text-generation model is a disclosed fixture limitation; successful ordinary automatic title generation is not proven. An outbox effect marked succeeded does not itself establish that the model call succeeded or failed, because handled fallback can finish successfully; the published transcript alone must not be used as proof of a model error. The source and original fixture history support that setup, and observed retained seed titles establish the relevant path result. “Fallback-title runtime branch” in the unverified list refers to deriveThreadTitleSeed’s fallback-label branch, distinct from automatic title-generation failure retaining a seed.

Unverified: native mobile (no adb/emulator), native desktop (no Electron session), multi-model send, attachment and fallback-label branches, other provider adapters, and remote/relay/tunnel. The shared helper protects UTF-16 surrogate pairs; this evidence does not establish grapheme-cluster preservation or fixes for separate server truncators. No broader verdict is implied.

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