The evidence report
verify never answers with a bare yes or no. It returns an EvidenceReport: the verdict, every check that was planned with its source, time and findings, and a list of what was not checked.
Fields
| Field | What it holds |
|---|---|
verdict | pass, fail or unknown |
summary | One line for a person, such as Failed: 1/1 checks completed, 4 finding(s). The checks are the answer |
target | The target as it was checked |
checks | One result for each check planned for this kind of target |
notChecked | What this verifier cannot or did not check. Always present, never empty |
generatedAt | When the report was made (ISO 8601) |
verifier | The package name and version that made it |
Each check result holds:
| Field | What it holds |
|---|---|
id | The check, such as osv.vulnerabilities |
kind | deterministic or heuristic |
status | pass, fail, error or skipped |
source | Where the evidence came from (name, sometimes url) and when (retrievedAt, ISO 8601) |
findings | What it found, each with a severity, a message and often a ref |
reason | Why it errored or was skipped |
durationMs | How long it took |
A finding's severity is info, low, medium, high or critical. Its ref points at what was found: an advisory id, a JSON pointer into the records, or a C2PA manifest field.
Verdicts
The verdict fails closed. It is worked out in this order:
- fail when any check failed, even a heuristic one or one that is not required.
- unknown when a required check is missing, errored or was skipped. Unknown is not zero.
- unknown when only heuristic checks are required: judgement can't certify.
- pass otherwise.
Treat unknown as not verified. Read notChecked before telling anyone that something is safe or true: a pass means the named checks passed against their sources at the time in retrievedAt, nothing more.
Statuses and kinds
passandfailmean the check ran to completion. The summary counts these as completed.errormeans it could not run: the source did not answer, the input was invalid, or the tool crashed. Thereasonsays which. A check never throws; an exception becomeserror.skippedmeans it did not run, on purpose: nothing is configured for it, or there was nothing it could decide. Thereasonsays which.deterministicchecks give the same result for the same input and source state: advisory lookups, schema validation, signature checks.heuristicchecks are judgements. They can fail a target, never pass one on their own.
Over MCP
The tool returns the report twice: as JSON text, for clients that only show text, and as structured content, for clients that read it as data. The tool is marked read-only (readOnlyHint), and as reaching outside the client (openWorldHint), since it looks advisories up on OSV.dev.
Rules baked in
- Evidence, not verdicts. Every check names its source and when it looked, and
notCheckedis always there, so limits are never silent. - Deterministic first. Every check is labelled
deterministicorheuristic, and heuristics can fail a target but never pass it alone. - Independent checks. There is no language model "plausibility" score. The verifier never asks the generator to grade its own work.
- Fail closed. A required check that errors or is skipped makes the verdict
unknown, neverpass. - Verify the verifier. The server is a dependency too: it is read-only, has three runtime dependencies, and is published with provenance. See the security policy.