# Verification record

Atlas Upstream 0.2.0 was exercised on 7 November 2026 with CPython 3.12.14 on Linux. The optional protocol integration used the official MCP Python SDK 1.29.0. The product is a local alpha.

| Check | Observed result |
| --- | --- |
| Regression suite | 44 tests passed, including the official MCP client/server integration. |
| Atomic evidence | An invalid event rolls back its entire batch. Identical concurrent retries commit once. Changed or reordered replays are rejected. |
| Recovery | A new server process resumes a completed mission. A separate test recovers an open attempt and prevents a duplicate begin. |
| Tool discovery | The official client discovered seven tools, read the operating resource and received explicit schemas for four event kinds. An incomplete observation was rejected. |
| External task | A real file contained stale report text. The receiving read identified the mismatch. Upstream requested repair and then accepted a verified read. |
| Opsbook integration | Imported the observed publication and three labelled fixture cases, learned a scoped reader-verification instruction, passed both replay sets, and returned the instruction for a future session. |
| Installation | Installed the wheel into an isolated virtual environment, imported the core outside the source checkout and exercised the CLI. |
| Installed MCP package | Exercised the wheel outside the source checkout in a virtual environment using the already available official SDK dependencies. |
| Older journals | Opened and resumed a database produced by the archived 0.1.0 executable. |
| Recorded scenarios | Regenerated the handover, travel learning, local HTTP and sixteen-world paired comparison with 0.2.0. |

The protocol driver follows a written policy. A language-model-driven task trial is still pending. Physical movement in the handover scenarios is simulated. The two-person field exercise has instructions, and its actual outcome remains to be observed.

The sixteen-world comparison continues to produce 16 verified handovers per policy, 40 versus 80 information requests, and 32 versus 64 modelled human contacts. These are counts within the supplied finite cases. They do not establish a general productivity multiplier.

The MCP run records its full mission and event journal. Its SHA-256 can therefore be recomputed against the exported lesson. The Opsbook example marks its three additional cases as constructed fixtures. Their expected checks are explicit labels. They are suitable for demonstrating the integration and its replay requirement.

The MCP dependency emits a settings forward-reference warning in this test environment. It did not prevent discovery, validation, tool execution or resumption. Browser execution of the HTML replay was unavailable in the sandbox. Its JavaScript controls were exercised with a minimal DOM harness. Its underlying decisions come from the recorded engine runs.

## Reproduce

```sh
python3 -m pip install ".[mcp]"
python3 -m unittest discover -s tests -v
python3 examples/mcp_client.py
python3 examples/opsbook_bridge.py /path/to/atlas-opsbook/opsbook.py
python3 demo.py compare
```

The first command downloads the MCP dependencies when needed. The tests and demonstrations use local temporary files. The Opsbook example uses a temporary manual and the supplied path to an existing Opsbook installation.