JSON output

Every attest command prints one JSON document on stdout, and nothing else. All but attest schema wrap their result the same way:

{"schema_version": "2", "command": "run", "ok": true, "data": { ... }}

When a command can't run, ok is false and error replaces data:

{"schema_version": "2", "command": "role", "ok": false,
 "error": {"code": "missing_api_key", "message": "Set the anthropic API key ..."}}

error.code is invalid_input, missing_api_key, provider_config or command_failed. provider_config means the provider rejected the key (401), the key lacks permission (403) or the model doesn't exist (404); the message never includes the key. command_failed means an error the checks don't absorb stopped the command, such as a provider error during attest review; message names the error.

A usage error, such as an unknown option, a missing argument or --max-claims abc, is invalid_input too. When the command itself is missing or unknown, command is attest. Only --help prints plain text.

Exit codes

Code	Meaning
0	The command ran.
1	A required API key is missing or rejected (missing_api_key, provider_config), or the command failed (command_failed).
2	The input or an option is invalid.
3	A budget stopped attest run. The runs graded so far are still printed.

The exit code says whether the command ran, not whether the agent passed.

What data holds

Command	data
run	trajectories: one entry per graded run, each a task and its report. Then total (runs in the file), evaluated (runs graded), verdicts (how many runs got each verdict: pass, review, block), passed (runs whose verdict is pass), checks_passed (runs whose checks all passed, including runs that still need review), mean_overall_score, usage (calls and tokens), and budget_exceeded: why a budget stopped the run, or null.
tools, injection, role	trajectories: one entry per run, each a task and its result
review	A ReviewResult
demo	comparisons: one per run, with task, answer, the plain LLM judge's verdict (judge), attest's faithfulness result (attest), and outcome: caught (attest caught a false claim the judge passed), both_caught, both_passed or judge_only
stats	rate, low, high, n
models	provider, source (live or curated), models

A report holds results (one per check), overall_score, verdict (pass, review or block), passed (verdict is pass), checks_passed (every check passed, ignoring dispositions), blocking (the findings that blocked, each a check, verdict and subject), scores (each check's score by name) and claim_coverage (the share of claims faithfulness checked, or null). Each result and finding has the fields described in How it works; a finding in a report also has its disposition. See Reports. Gate on verdict or passed; an inconclusive run can have checks_passed: true and still verdict: "review".

From Python, report.model_dump_json() gives the same JSON as one report in attest run's output, and a check's result.model_dump_json() the same as one result from attest tools, injection or role.

Examples

# The failed claims, with the evidence against them
attest run run.json | jq '.data.trajectories[].report.results[].findings[]
  | select(.severity == "fail") | {subject, reason, evidence}'

# Fail a CI step unless every run's verdict is pass (review and block both fail it)
attest run runs.jsonl | jq -e '.ok and .data.verdicts.pass == .data.total'

# Fail only on blocks, and let runs that need review through
attest run runs.jsonl | jq -e '.ok and .data.verdicts.block == 0'

Schemas

attest schema <command> prints the JSON Schema for a command's output. The same schemas are published as schemas/v2/<command>.json on this site, for example run.json and error.json.

Within a schema_version, fields may be added but never removed or redefined. Version "2" changed what passed means (see the changelog); the v1 schemas are still published.