Artifact Reference

Twin: none — reference page (backing: []). Every kind below is emitted by a command in reference/cli.md and verified by the postcondition pattern shown in section 3.

1. What you are testing

Every agent-learn command that produces evidence writes a JSON artifact with a top-level kind field. The kind universe is closed: the docs gate rejects a page that claims to emit a kind outside V1_DOCS_ALLOWED_ARTIFACT_KINDS (src/fi/alk/trinity.py), and agent-learn release-check asserts the eleven core kinds in V1_REQUIRED_SCHEMA_KINDS are producible. The closed set is what makes postconditions one-liners — checking payload["kind"] is always sufficient to know what you are holding.

Two values look like kinds but are not artifact kinds: agent-learning.cli.v1 is the CLI payload schema_version label, and any vendored agent-simulate.* value is rewritten to its public agent-learning.* form by public_schema_value / normalize_public_payload in src/fi/alk/_schema.py before an artifact is written.

2. Run it

Produce one artifact and inspect its kind (offline, no credentials):

agent-learn run examples/run_manifest.json --no-eval --output artifacts/run.json

import json
payload = json.load(open("artifacts/run.json"))
print(payload["kind"])  # agent-learning.run.v1

3. What you built

python -c "import json; p=json.load(open('artifacts/run.json')); assert p['kind']=='agent-learning.run.v1', p['kind']; print('ok')"

The full catalog. "Core" marks the eleven V1_REQUIRED_SCHEMA_KINDS asserted by the release gate; the remainder come from the public command registry.

Kind	Core	Emitted by
agent-learning.run.v1	yes	agent-learn run — one simulation run: transcript, world/task state, optional eval attachments
agent-learning.eval.v1	yes	agent-learn eval — eval-suite verdicts over prompts and outputs
agent-learning.artifact-evaluation.v1	yes	agent-learn eval-artifact — evaluation computed over an already-saved artifact
agent-learning.task-evidence.v1	—	agent-learn eval-task — synthesized task-evidence record from task artifacts
agent-learning.redteam.v1	yes	agent-learn redteam / redteam-corpus — campaign findings and corpus-hook results
agent-learning.optimization.v1	yes	agent-learn optimize — candidate history with content-addressed lineage
agent-learning.eval-optimization.v1	yes	agent-learn optimize-eval — optimization over an eval suite itself
agent-learning.suite.v1	yes	agent-learn suite — combined multi-step suite result
agent-learning.suite-optimization.v1	yes	agent-learn optimize-suite / action-optimize — optimization over a suite
agent-learning.actions.v1	yes	agent-learn actions — the available-actions catalog
agent-learning.action-run.v1	yes	agent-learn action-run — one executed action with its result
agent-learning.release-proof.v1	yes	agent-learn release-proof — the seven-check release proof object
agent-learning.baseline.v1	—	agent-learn baseline — pinned regression baseline
agent-learning.compare.v1	—	agent-learn compare — baseline-vs-current comparison verdict
agent-learning.init.v1	—	agent-learn init — scaffold record for a preset
agent-learning.regression-promotion.v1	—	agent-learn promote-to-regression — a finding promoted into the regression set
agent-learning.attack-evolution-shrink.v1	—	agent-learn shrink — minimized counterexample from an evolved attack
agent-learning.replay.v1	—	agent-learn replay — deterministic re-execution verdict for a kept artifact
agent-learning.report.v1	—	agent-learn report — rendered report over saved artifacts
agent-learning.doctor.v1	—	agent-learn doctor — environment and module diagnostics
agent-learning.release-check.v1	—	agent-learn release-check — the full local gate matrix verdict

4. When it fails

Symptom	First-mile class	Doctor check
kind holds an agent-simulate.* value	engine — artifact written without _schema normalization	agent-learn doctor → summary.missing_engine_modules
KeyError: 'kind' reading an artifact	config fault — file is not an agent-learn artifact (or pre-v1)	re-emit with the current CLI; check the --output path
docs gate rejects a page's artifact_kinds	config fault — value outside the closed kind set	compare against the table above (the gate payload mirrors it)

5. Prove it / keep it

agent-learn release-check --project-root . asserts the required kinds and mirrors the allowed set in its evidence payload (docs_allowed_artifact_kinds), so the catalog above cannot drift silently from the code. To put that check in your pipeline, continue with prove/release-check-in-your-ci.md; to keep a specific artifact as a regression baseline, continue with simulate/regression-lifecycle.md.