# Final independent read-only proof audit

**Evidence is INSUFFICIENT: six complete scoped pairs, not eleven. Code verdict: no new concrete defect established.** The independent affected-final rebuild succeeded; that formerly missing requirement is now satisfied.

Audited [#23193](https://github.com/pingdotgg/t3code/pull/23193), GitCafe [#23172](https://git.cafe/pingdotgg/t3code/pulls/23172), pinned head [1723431e82](https://git.cafe/pingdotgg/t3code/commit/1723431e82249a7e44a15b19ea0db4b5d09d6057), exact base [b5a226e49c](https://git.cafe/pingdotgg/t3code/commit/b5a226e49c52ecab4ddc2d678a7012d97e651a6a). Only repository accessed: `/home/ubuntu/wt/pr-23193`. Audit2026-10-22 UTC, completed before12:18. Read original PASS1 matrix, complete PASS2, capture/report.md, FOLLOWUP3 report and actual artifacts. This is pinned local proof review, not remote metadata freshness or a merge recommendation.

## Code and source identity

Read complete base→head diff: two files,25insertions/4deletions. Production predicate recognizes exact bare command case-insensitively, or lowercase `/compact` followed by JS whitespace. This prevents the redundant bare pre-compaction and forced queue for explicit lowercase instructed commands. Outer trim occurs in snapshot construction. Setup queueing stays independent. Followed send caller, multiline/collapse gate and server bare normalization. No new production change, concrete regression or simplification blocker found. Existing focused test success remains prior evidence; no tests rerun here.

Current HEAD is exactly the pinned40-character SHA. Both tracked files match git-show head and saved restored-head hashes: logic `f83b7cb716b2f626f3829c4533a2aa157d18e0614163a8400de220b650e5bd7b`; tests `1af97fe1f3f601d24d0a707f92b33435e1bfb130ce179f23568bbe7c676a1b13`. Saved baseline files byte-match git-show exact base; baseline-match.diff is empty. Since the complete base/head diff consists only of those two paths, the disclosed two-file source substitution represents this base's code. Capture-time browser/HMR provenance is supported by source-switch/restoration records and observed base behavior, not a new full checkout or independently witnessed historical switch. I made no source substitutions.

## Independently audited row ledger

Counts exclude prerequisite real OK. Base send/terminal UTC are persisted run.created and final completed run.updated, independently read from isolated `.t3/userdata/statev2.sqlite` using SQLite mode=ro. They are not click/recorder timestamps. All baseline affected runs are completed in storage; that does not repair partial media. PASS2 final times/IDs remain preserved in its full ledger; final texts/counts were reread here. Starting+queued/q1 follows a two-message send; single-message cases start directly. Literal outbound dispatchMode is not recorded: `auto` remains source-derived.

| Row | Added user messages base→final | Base send→terminal UTC | Base video | Final video | Audited pair / gap |
|---|---:|---|---|---|---|
| P1 | 2 → 1 | 11:40:40.413Z → 11:41:05.772Z | settled | settled | COMPLETE |
| P2 | 2 → 1 | 11:42:37.815Z → 11:42:59.566Z | PARTIAL | settled | base terminal |
| P3 | 2 → 1 | 11:58:18.261Z → 11:58:41.539Z | settled | settled | base expansion |
| P4 | 2 → 1 | 11:43:56.453Z → 11:44:13.861Z | settled | settled | COMPLETE |
| P5 | 2 → 1 | 11:44:50.294Z → 11:45:09.558Z | settled | settled | COMPLETE |
| K1 | 2 → 2 | 11:45:42.161Z → 11:46:06.947Z | PARTIAL | settled | base terminal |
| K2 | 2 → 2 | 11:46:45.260Z → 11:47:18.348Z | settled | settled | COMPLETE |
| K3-lower | 1 → 1 | 11:48:01.157Z → 11:48:13.347Z | settled | settled | COMPLETE |
| K3-upper | 1 → 1 | 11:49:03.206Z → 11:49:37.555Z | PARTIAL | settled | base terminal |
| K4-suffix | 2 → 2 | 11:50:40.116Z → 11:51:05.754Z | settled | PARTIAL | final terminal |
| K4-slash | 2 → 2 | 11:51:49.239Z → 11:52:22.265Z | settled | settled | COMPLETE |
| K4-prose | 2 → 2 | 11:53:02.854Z → 11:53:35.428Z | settled | PARTIAL | final terminal |
| K5-full | 1 → 1 | 11:54:53.046Z → 11:55:07.259Z | PARTIAL | PARTIAL | base terminal; final terminal |
| K5-restore | 2 → 2 | 11:55:45.249Z → 11:56:04.212Z | PARTIAL | settled | base terminal |

Exact stored texts: P1 long lowercase instructed command; P2 lowercase double spaces; P3 literal newline; P4 outer whitespace trimmed; P5 internal literal tab preserved. K1 `/Compact  keep the interface decisions`; K2 `/COMPACT keep decisions`; K3 exact lower/uppercase bare; K4 suffix `/compactness keep decisions`, slash `/compact/keep`, prose `Please /compact keep decisions`; K5 Full sends P1 only, restored Compact injects bare then `Reply OK`. All base non-bare/non-Full cases have injected bare then original. Fresh unique native threads in correct project were checked from every terminal JSON. Starting/queued/q1 states were checked against persisted events, not inferred from video text alone.

**Six complete pairs: P1,P4,P5,K2,K3-lower,K4-slash.** Eleven refers to settled original final recordings, not matched pairs. Current immutable FOLLOWUP3 report already corrects the earlier eleven-pair claim; this audit independently agrees with six. Nine baseline recordings visibly settle; only seven overlap the eleven settled finals, and P3 lacks its expansion interaction, leaving six complete scoped pairs. K4-suffix replacement is partial and cannot promote that row.

Remaining interaction gaps exactly: baseline terminal P2,K1,K3-upper,K5-full,K5-restore; final terminal K4-suffix,K4-prose,K5-full; baseline P3 collapsed-bare expansion/no dispatch. That is five baseline terminal replacements, three final replacements, and one missing baseline interaction sequence. Independent rebuild no longer belongs in this gap set. No new capture passes were attempted for those rows.

## Media, source scope and P3

Independently decoded all15FOLLOWUP3 MP4s (14baseline+1suffix replacement), each exit0, and extracted last frames. Visually inspected the complete terminal contact sheet and representative progression for P1,P3,K5restore, plus final contact sheet and independent rebuild. Other generated progression sheets are retained, not claimed as fully watched playback. Earlier final decode/complete-sequence checking is PASS2's prior independent evidence; no claim of a second full playback pass on all originals. Sparse VFR emits non-monotonic DTS warnings on null-output; decode success is not pacing certification. Actual wide source videos2880×1800; P3=780×1800. Saved wide stills1280×800. Do not confuse viewer downscaling with original media dimensions.

P3 adjustment remains valid: literal newline makes hasMultilinePrompt true, and collapsed mobile requires its negation. Valid flow is collapsed bare→expand/no dispatch→expanded multiline→send. Original final P3 proves that adjusted flow in PASS2. New baseline recording begins expanded with bare text, then multiline; failed Open composer recovery and two prerequisite messages establish no prior dispatch, but do not prove successful collapsed expansion. Its settled send is usable partial interaction evidence, not a complete pair. Source and zero-message facts do not waive missing interaction footage.

All fourteen baseline seed records preserve native/cache fields; native usage91720, synthetic eligibility250000/max1000000, staleTTL+120000ms. Real Claude Sonnet5.5/claudeAgent prerequisites are separate from seeded client eligibility. This does not certify genuine stale250k native history or semantic retention of interface decisions.

## Timing/scripts and exact differences

New driver.js and13ordinary/replacement timelines materially improve reproducibility. Recomputed every listed offset from recorder.startedAt: all actual supplied action offsets agree. P1 manual send offset1.526s; P3 manual send offset15.775s. Manual records do not include all P3 recovery boundaries. Final suffix timeline was overwritten during recovery and omits initial steps. Driver's generic toggle accessibility label differs from the actual retained chip label; actual timeline is authoritative. These are reconstruction limits, not newly completed exact automated recipes. Original final fourteen runs still lack complete per-recorder/action logs as described in PASS2. No MP4 creation-time subtraction or invented offsets used.

Independently recomputed **all28 new base/final pre/post comparisons** from RGB inputs. Every nonzero/max-channel>16 count,max,mean matches diff-metrics.json; all56 PNGs match exactly. Formula: clipped8×absRGB for amplified zero-threshold image; magenta where max-channel delta strictly>16/255, black elsewhere. Full frame, no alignment,masks,crops or exclusions. Metrics and independently checked paths are in support checks.json. P3 images780×1800; all other stills1280×800. Partial rows' post stills are useful visual comparisons but not terminal-equivalent paired video proof.

Visible unrelated differences remain titles/sidebar counts/order/ages/provider icons, clocks, source-control +4−25 baseline versus +0−0 final, token/summary totals, response prose/linewrap, Working/Worked and elapsed labels, scroll/vertical placement. P3 narrow/dark hides sidebars and wraps text. New suffix has provider-update toast. Baseline affected sends add bare bubble,queue,second command and second compaction failure; final affected sends omit these. Raw pixel differences include fixture/environment changes and cannot isolate predicate-regression pixels. Historical [70e5e384da](https://git.cafe/pingdotgg/t3code/commit/70e5e384da) and earlier-after [fafd55cbfa](https://git.cafe/pingdotgg/t3code/commit/fafd55cbfa) remain historical only, never relabeled as requested base/final.

## Independent single affected-final rebuild: successful

Used only own T3 preview tab and existing6695/14735server, exact project-owned route, header New thread in pr-23193, fresh thread `e3b39030-b6fa-446e-a145-b06da384981b`. No send to an existing captured thread. One pairing token minted/used; credentials excluded. Archived seeder/read-expression byte-match the scripts actually used. Native prompt→real OK, completed idle run, native91680tokens; seed250000/1M atTTL4200000+120000ms; Settings→Back remount.1440×900CSS,DPR2,zoom1,light; Claude Sonnet5.5,High,Full access.

Exact P1 text, Escape, Compact and send once. Recorder started12:08:51.706Z. Persisted affected run.created12:08:54.748Z starting (offset3.042s), completed12:09:05.010Z (offset13.304s). One new user message only; no duplicate bare command/queue. Successful compaction91.7K→1.50K. Immediate post-click projection still had prerequisite only; it was not used as dispatch evidence. Terminal projection and persisted events confirm completion, and independently decoded last frame visibly has successful compaction and blue Submit. Recording34.3652s/51frames/2880×1800, decode0 with VFR warnings. This satisfies one independent affected rebuild; no retry send.

Comparison against original final P1 stills, full1280×800,no masks:

| Rebuild versus original P1 | Pixels nonzero /1,024,000 | Pixels>16 | Max | Mean absolute channel delta/255 |
|---|---:|---:|---:|---:|
| pre |11913|9640|249|0.7085042317708333|
| post |13017|9664|249|0.7114182942708334|

Nonzero changes visibly include title/sidebar labels and ages/order,count, summary1.50K versus original1.85K, while command count/layout behavior agrees. Pixel delta is **MEASURED**, not zero or UNMEASURED. Recipe,pre/post,last frame,threshold artifacts,native/seed/initial/terminal/events and actual start/action records are in support package.

[Independent P1 recording](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3223cff1cde3b270/rebuild-P1.mp4)

![Independent final-head P1 rebuild: one instructed command and visibly settled successful compaction](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b8504f6f13695431/rebuild-post.png)

![Independently decoded FOLLOWUP3 last frames: nine settled baseline endings, five partial baseline endings, and partial suffix-final replacement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/986972b9ece59c1d/independent-r3-terminals.jpg)

## Artifact version and delivery

Read local FOLLOWUP3reportSHA256 `1cbb8e0d611b01ac921f4290de1245b496066263747d8e40bcbf382e9c708309`; manifestSHA256 `db5cf3434af9831775ab46795dd2746fc0ace3c31570b23eba079273421f7e61`. All234manifest entries verified with zero mismatches. Local proof-package SHA256 `b54f52d1cdaff47904ac24a626c8138b8340c564a080f20817c7ed464f101cf6`; its235entries include same report and manifest. External-seederSHA256 `af651866297e4e9c891b85bcd6b43721f509920976878c5da9e05961c741ce96`. Local inputs frozen by hashes in checks.json; no silent packaging-version substitution.

Capture author's [uploaded FOLLOWUP3 report](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6928730009f915e6/report.md) and [proof package](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4eedc216937c47ea/proof-package.zip) are identified from upload JSON. This audit verifies local package bytes/internal report, not a new download/hash certification of those remote URLs. New audit [independent support package](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b9803649921aa948/audit-final-support.zip) contains scripts,metrics,row IDs/events ledger,decode results,frames and successful rebuild. No credentials or pairing URL included.

## Surfaces and boundaries

Web is the required surface and was exercised. Desktop shares source but shell execution is unverified. Native mobile has no compact-before-send feature; responsive web is not native certification. Remote/relay/tunnel and other providers were not exercised; unchanged protocol/adapter boundaries do not add required matrix rows. No new entry point,contract,service or reverse-state implementation was added.

No code edits/comments/forge changes/watches/polls/merge/new server. T3 required PR link/list bookkeeping confirmed watching=false. Isolated fresh fixture sends/seed and one pairing are the explicitly authorized runtime mutations. Captured threads remained untouched; parent server remains running; no cleanup or process kill. Worktree tracked state stayed clean.

## FRICTION

category|what|avoid
---|---|---
Evidence counting|Eleven settled finals was promoted to eleven pairs in preliminary claim; current report correctly retracts to six|Intersect row-level base/final completeness and required actions before counting
Recorder|Five baseline endings and suffix replacement retain red Stop despite persisted completion|Judge decoded last paint; preserve partial labels; no later still substitution
P3|Expanded multiline baseline does not prove collapsed-bare expansion; narrow recovery logs incomplete|Establish actual collapsed state before recording, retain expansion/no-dispatch evidence
Script/timing|Manual P1/P3 boundaries incomplete; suffix resume overwrote initial timeline; generic toggle selector stale|Retain append-only actual calls and recorder.startedAt per run; never invent offsets
Browser API|Resize errors: Missing key at [mode], then Expected fill/freeform/preset|Use {mode:freeform,width:1440,height:900}; correction succeeded without extra send
Authentication|Fresh own tab required pairing once|Retain one CLI result privately; use same token; exclude it from artifacts
Scope|Old historical footage differs in SHA; shared desktop source is not execution|Keep exact provenance and surface limits

Model/harness:gpt-6.1-sol / Codex in T3 Code. Skills:pr-audit,test-t3-app,ghdrop; no skill-authorized messaging used.

VERDICT: INSUFFICIENT — base terminal P2,K1,K3-upper,K5-full,K5-restore; final terminal K4-suffix,K4-prose,K5-full; base P3 collapsed expansion; incomplete exact recovery/original-final recorder timelines