# AdoptLab v0.4 upgrade evidence

Recorded on 2026-10-07. This report concerns local, automated protocol execution. It does not establish improved model performance or human adoption.

The new workspace covers fourteen workflow modules in Chinese and English. The public landing page contains three six-step case walkthroughs; these are explicitly marked demonstrations and do not execute a visitor's tools or models. The historical experiment explorer is preserved at `experiments.html` with its 48 protocol checks, 144 model episodes and older 72-episode archive.

## Real local execution

| Task | Baseline/revised runs | Independent acceptance | Release check |
| --- | --- | --- | --- |
| task-01: built-in records | 2 | Both accepted | Passed |
| docs-evidence-v04: official Filesystem document evidence | 2 | Both accepted | Passed |
| docs-config-v04: document fields and source line | 2 | Both accepted | Passed |

The six-run experiment is `a4f28e1acef94a6bace5c49f66e9c689`. Each case has feedback, a different immutable material, a same-condition retest and a recorded release check. Task, material, condition and artifact hashes are available in [the sanitized machine-readable report](workbench-evidence.json). A separate task registered through the new editor also passed a real Filesystem protocol run and artifact re-verification.

The two accepted protocol runs per case test the workflow and reference plan. They do not measure whether revised wording helps a model or a person. Historical model results have not been regenerated or merged into this denominator. Observed human sessions: 0; adoption benefits: unknown; new model expense: zero.

## Verification scope

Local automated tests cover previous contracts, read-only field preflight, immutable registration, material mismatch, migration backups, artifact tampering, missing release evidence, observation consent/deduplication/withdrawal and privacy projection. Existing cancellation, timeout, condition/responder and unresolved-cost tests remain part of the suite. The complete local Windows suite passed: 50 tests, 2 skipped (optional unavailable targets), with one dependency deprecation warning. Six public-report contract checks also passed after the final export changes. Hosted Windows/Linux/container CI must pass before merging; its exact run is linked in the release PR.

Browser checks use Microsoft Edge. The user's existing tab was checked through its Playwright interface: fourteen Chinese and fourteen English modules, three cases × six steps in both languages, developer/public paths at 1024 and 390 pixels, and an actual report download whose SHA-256 matched the committed report. A mobile grid overflow was found and fixed; a versioned stylesheet prevents stale development-cache results. Automated checks are not trial participants.

Figma has fourteen Chinese and fourteen English main editable screens, a shared variable/variant system, eleven feedback states, linked maintainer/developer/public flows, and responsive developer/public design references. The three desktop prototype journeys are linked separately in each language. Vector PDFs, screenshots and design contexts were exported locally. Cloud save remains unconfirmed after the operator stopped Computer Use; a successful plugin call alone is insufficient. Keyboard-only browser acceptance and a final presentation-mode check remain unconfirmed. The design examples and live application have semantic differences documented in the component mapping rather than an unsupported claim of exact pixel parity.

## Historical preservation and publication

The reviewed terminal upgrade runtime was imported into the existing local runtime after a SQLite backup. A read-only comparison confirmed that all 2,528 historical rows across 12 tables remain unchanged. The cost ledger upper bound remains CNY 3.963214; 22 historical unresolved requests retain their reservations. Four historical queued runs were preserved without automatic retries. All six published upgrade runs were independently reverified after the merge, including task, manifest and artifact fingerprints. The public build contains 25 allowlisted static files and preserves the 192-episode current historical report and older archive. No local backend or credential is published.

## Limits and next evidence

This is a local single-user workbench and a public static evidence site. The published site does not expose the execution backend. Observations entered locally are attested by their recorder; external identity is not automatically verified. Actual trial tasks and an observation ledger are prepared separately. Measure independent first success and revision effort only after real participants complete those tasks.