What MAESTRO writes down about a change
Every mission ends with a record. Its most important field is the one that says whether the change was verified — and, when it was not, why not.
Why a status of "verified" is usually worth nothing
Most AI coding agents verify a change by running the repository's existing tests. That works when there is a suite, when it runs, and when it covers the thing that changed. When any of those is untrue the agent finds nothing to run, and reports success anyway — because its notion of failure is "a test failed", and no test failed.
There is a quieter version of the same problem. A test runner can exit non-zero without a single failing test: a coverage threshold, a lint plugin, a missing interpreter, a configuration error. An agent that reads only the printed summary sees passes and no failures, and calls the run green while the runner was saying the opposite. MAESTRO treats a disagreement between the exit code and the parsed counts as a reason to distrust the parse, not to trust the summary — the run is marked unverified and the exit code is named in the reason.
In an internal benchmark across ten open-source repositories, MAESTRO produced no false green. Five changes came back marked unverified, each with the reason; the rest did not complete. Nothing was reported as verified.
The three verification modes
A mission's record carries a verification mode, and the mode is a statement about evidence rather than about outcome.
- Full. The project's own test suite ran, its output was read, and the result was compared against the baseline captured before the change. A change that introduced a new failure is not committed.
- VM sandbox. Tests ran inside the isolated microVM for the change under review, and the result is attributed to that run.
- None. Verification did not happen. This is a real, reportable outcome, not an error state — it is what honest looks like when there was nothing to run.
A mission that produced a working change it could not verify finishes
with the status completed_unverified and an amber badge, not
a green one. The pull request body says the same thing the dashboard
says.
Why verification did not happen
"No tests were run" is a useless sentence if it is the only one available, because it blames your repository for every cause including ours. The record distinguishes them:
- the test runner your project uses was not present in the sandbox;
- the run started and was killed before it finished;
- the suite collected nothing — it started and found no tests;
- the signals disagreed — the exit code and the reported counts cannot both be true;
- the suite needs a browser that the sandbox does not provide;
- the output could not be read by any parser we have;
- the suite did not start at all.
The first of those is our gap, not yours. Saying so is the point: an agent that reports "this repository has no test suite" about a repository with a test suite has not been honest, it has been convenient.
What a record contains
The record is assembled at the end of the mission and is what the dashboard, the exported PDF and the API all serve. It holds the specification you approved, the diff that was produced, the test command and its output, the verification mode and reason, the review findings and the final verdict, the commit and pull-request URL, what the mission cost, and any boundary violation the sandbox refused.
It is deliberately a record of a process, not a score. A number that is derived from a verdict rather than measured is a verdict wearing a disguise, and MAESTRO would rather show you the verdict and the findings it rests on.
Tamper evidence
Records are hash-chained. Each entry stores the SHA-256 of its own canonical contents and the hash of the entry before it, across all accounts in one global sequence. Altering a record invalidates every link after it, including links belonging to other customers — which is what makes quiet editing of a single record impractical rather than merely discouraged. The chain table accepts inserts and nothing else: the application has no permission to update or delete a row in it.
A scheduled job submits the head of the chain to an RFC 3161 timestamping authority each week and publishes the token.
Records older than the last timestamp anchor are independently verifiable. Records written since the last anchor are internally consistent only. Anyone offering you a stronger claim about an unanchored record is describing something a hash chain cannot do on its own.
Erasure is handled without breaking the chain. When an account is deleted its stored records go with it, and a tombstone entry — a hash derived from the identifier and the time, holding no personal data — takes their place in the sequence, so the links on either side still verify. A request under Article 17 does not force a choice between honouring it and keeping the chain checkable.
Where to look
Every mission page links to its own record, and the record is available as a PDF for the cases where someone outside the team has to read it. The chain position and hash for a mission are available over the API if you want to check a record against the chain yourself rather than take our word for it.
If you have not read it yet, the thirteen phases explains where in a mission each of these fields comes from, and sandbox isolation covers where the code actually runs.