# Real-app proof procedure

Run each revision as a separate checkout with dependencies installed and `T3CODE_DEV_SEED=0 vp run dev`. Add an empty synthetic git workspace titled Fruit sales through `node apps/server/src/bin.ts project add PROJECT --title 'Fruit sales' --base-dir .t3`. Enable OpenCode in Settings > Providers and set its binary path to the downloaded OpenCode v2.2.5 binary. Use DeepSeek V4.1 Flash Free.

`drive.py CHECKOUT PROJECT original` sends the verbatim triage prompt through the running server's CLI. Open the returned thread in the real web app. Base: no chart. Head: both pages published by the model (it corrected its styling) are present.

`drive.py CHECKOUT PROJECT done` supplies the same fixed chart HTML on both sides to remove variation in model-generated design. Inspect the completed thread in the real app: base zero charts, head one.

On that done thread, the browser recordings use the following UI actions on each revision, at 1280×900 CSS pixels, light theme, sidebar and thread details collapsed:

1. Start preview recording.
2. Fill `[contenteditable=true][role=textbox]` with the prompt below.
3. Click `button[aria-label='Submit message']`.
4. Wait for the final reply and stop recording.

Prompt for the final discard/no-return recordings:

> Repeat the exact same HTML once, title 'Fruit sales — discarded void'. Use one execute script: preview first, then assign the awaited html_render result to a const named published. End the script with exactly `const finished = true;` after that assignment. Include NO return statement. This final declaration must discard the page result; do not leave html_render as the final expression. Do not render outside execute, and do not republish. Reply only 'Published; script returned nothing.'

The prior turn on each side used this prompt:

> Repeat the exact HTML from my previous request once. In ONE execute script, call html_preview and then html_render, title 'Fruit sales — nothing', using preview contentHeight. Discard html_render's result. Do not include any return statement: the execute script must return nothing. Do not call html_render outside that script. Do not revise or republish. Reply only 'Published; script returned nothing.'

Important observation: base's first no-return script ended in an awaited html_render expression, and OpenCode implicitly returned that reference, so it displayed a chart. Head ended with a declaration and returned null. This was not a valid comparison of discarded results. The final recording repeats the same explicit final-declaration prompt on both revisions; both actual execute outputs are null. The old chart above the latest prompt in each final recording is previous history, not a newly duplicated publication.

Run `drive.py CHECKOUT PROJECT unrelated`, `repeated`, and `reference` in separate new threads. The uploaded driver's stricter return instructions ensure the relevant result is actually discarded or directly returned. Inspect actual execute scripts and returned text (executed-scripts.json), not just the model's final claims. Base: unrelated zero charts; repeated zero; reference one. Head: unrelated one; repeated two distinct attachments with identical arguments; reference exactly one, with no extra recovered duplicate.

For stills, blur the composer and take two snapshots at least 500 ms apart; require identical PNG SHA-256 hashes. Inspect each capture before upload. Original-prompt stills use 1440×1200 CSS pixels. Fixed-chart stills and videos use 1280×900. Host labels were hidden for privacy in the original-prompt stills; later captures collapse the panels that contain them. No timeline or page content was altered.

The generic replay scenario runner cannot cover this: its agent speaks generic ACP, while the behavior depends on OpenCode 2 Code Mode metadata and its real execute protocol. No automated test result is used as proof.
