# Study 002 provenance and completion amendment

Original run date: 2026-08-15 (Australia/Sydney)<br />
Completion evaluation: 2026-09-03 (Australia/Sydney)

## Status of the frozen record

The original application artifacts, handoffs, protocol, prompts, task and
frozen context have not been edited for the completion evaluation. The dated
supplement adds the missing A, B and D Chrome keyboard observations,
screenshots, a recovered MCP-call trace and an executable evidence check.

`manifest.txt` records the exact bytes that were frozen in the temporary run
workspace. During the original publication copy, a single final line-feed was
removed from each of these five text files:

| Frozen file | Published-copy SHA-256 | Reconciliation |
|---|---|---|
| `common-prompt.md` | `27539cb50d2c5aedfe7e06deb6f11b2dee4bd1feb82e74286dbfde2904e512b3` | Appending one LF produces the manifest hash. |
| `mcp-treatment.md` | `f0b495bb8f25488ffd3d0da173a480795ef71fe7c0395fe8d83759b77d352a45` | Appending one LF produces the manifest hash. |
| `FROZEN_CONTEXT.md` | `d3004db21f4f95721f1bdc8f37dd324c571c6a59c4bb3db2f22dd92845eb749d` | Appending one LF produces the manifest hash. |
| `TASK.md` | `e29ef8e60f3cfb762eef7b3b8347d2db8d065b0baedc45d23746a4446b22a03f` | Appending one LF produces the manifest hash. |
| `README.md` | `383f6c44d3774e38a6ad915baf433feefb102a9af63eb26cec18df8711e8ea43` | Appending one LF produces the manifest hash. |

This is a known publication normalisation, not an unexplained content change.
The frozen `.gitignore` bytes were:

```text
.DS_Store
RUN_OUTPUT.md

```

including the final blank line. Those bytes reproduce its manifest hash
`c5aaee992cd50ed2be253f861dca3c42349a55619f42209e5ff3f698a04dcdeb`.
The candidate standards index is still available at commit
`c93e9d30f54baac1cb5a38603b555e33d067493a` and reproduces the manifest hash.

Run `npm run check:study-002` from the repository root to repeat these checks
and verify the completion evidence is present and parseable.

## Original run invocation

All conditions used Codex CLI 0.147.0, `gpt-5.6-sol`, high reasoning,
`--ephemeral`, `--json`, `--sandbox workspace-write` and
`--ignore-user-config`. The paths below are intentionally normalised; the
original append-only coordinator record contains the absolute temporary paths.

- A ran the common prompt in `<study-root>/run-A-baseline` with no Human
  Standards server.
- B ran the common prompt plus `mcp-treatment.md` in
  `<study-root>/run-B-legacy-mcp` with required server command
  `npx --yes @humanstandards/mcp-server@0.3.0`, a 30-second startup timeout and
  a 60-second tool timeout.
- C ran the same combined prompt in `<study-root>/run-C-candidate-mcp` with the
  required local candidate server from the repository checkout at commit
  `c93e9d30f54baac1cb5a38603b555e33d067493a`, using the same timeouts.
- D ran the common prompt plus an instruction to read the exact frozen context
  in `<study-root>/run-D-frozen-context`, with no MCP server.

Each run wrote its final message to `RUN_OUTPUT.md`. That file was deliberately
ignored in the frozen workspace; the published `HANDOFF.md` files are the
preserved final handoffs.

## MCP-call evidence

`mcp-call-trace.jsonl` contains the recoverable completed Human Standards MCP
events from the original append-only coordinator record, including exact
arguments, status, returned data and a SHA-256 of each result. It contains 24
recovered calls for B and 10 for C.

The trace is explicitly **partial**. Original output-size truncation left two
malformed nested JSON lines for B, one for C and one unparsed command envelope
for each. The B and C `HANDOFF.md` files remain the complete self-reported call
inventory. This supplement improves auditability but does not pretend that a
complete raw transcript was preserved after the fact.

## Completion-evaluation method

The same frozen A, B and D artifacts were served locally at `1440 × 1000` and
`390 × 844`. Fictional details were populated with the browser test driver's
form-fill action. Navigation, selection, Retry, review, confirmation and dialog
dismissal were exercised with real Chrome Tab, Shift+Tab, Enter, Space, arrow
and Escape input. Active elements and rendered focus indicators were inspected
after each relevant action. Confirmation was reloaded on each origin to verify
the same reference persisted.

The machine-readable outcomes are in `keyboard-results-comparative.json`. The
original C-only result remains unchanged in `keyboard-results.json`.

## Boundaries that remain

This completes the evidence promised by the frozen one-run comparison; it does
not convert the comparison into a causal study. It still has one artifact per
condition, no blinded independent human review, no assistive-technology run and
no cross-browser exercise. The recovered MCP trace is partial for the reasons
above. These limits must travel with any published claim.