Evidence Taxonomy

AGILAB evidence should be easy to inspect without rerunning the workflow. This taxonomy defines the first shared vocabulary for run manifests, proof bundles, notebook exports, UI robot evidence, MLflow handoff records, and release proof.

Common Event Envelope

Evidence events should use a small common envelope when they are serialized into JSON evidence bundles:

Field

Meaning

schema_version

Version of the evidence event schema.

event_type

One of the event names listed below.

run_id

AGILAB run, proof, robot, release, or agent-run identifier.

seq

Monotonic sequence number within the evidence bundle.

created_at_utc

UTC timestamp recorded by the producing tool.

artifact_sha256

SHA-256 of the referenced artifact when the event has one.

prev_event_hash

Optional previous event hash for future tamper-evident chains.

event_hash

Optional hash of the canonicalized event envelope and payload.

payload

Event-specific metadata. It must not contain secrets or large raw artifacts.

Event Types

Event type

Purpose

run_manifest_event

Records the top-level run manifest, selected app, environment, and command boundary.

stage_transition_event

Records a workflow, DAG, or pipeline stage state transition.

artifact_event

Records a produced file, directory manifest, or content hash.

notebook_export_event

Records an exported notebook or notebook export manifest.

mlflow_handoff_event

Records an MLflow tracking or registry handoff when that integration is enabled.

ui_robot_event

Records screenshots, traces, HAR, video, aggregate JSON, and replay commands from UI robot validation.

agent_run_event

Records a coding-agent or assistant-backed command through the AGILAB agent-run evidence surface.

policy_check_event

Records a deterministic policy or promotion gate decision.

release_proof_event

Records release tag, package, docs, CI, coverage, SBOM, audit, or provenance evidence used in release proof.

Redaction Rules

Evidence payloads must not store secrets, raw prompts that contain credentials, full notebook outputs with sensitive data, or large artifact bodies. Prefer:

  • content hashes over raw content

  • stable reason codes over free text when a code is enough

  • local file references over embedded blobs

  • explicit redacted markers when a field was intentionally removed

Verifier Scope

A verifier consumes this taxonomy to check evidence without rerunning work. It may validate:

  • schema versions

  • required event fields

  • monotonic sequence numbers

  • artifact hash matches

  • reference closure

  • release-proof metadata consistency

It must not validate facts outside the evidence bundle, such as legal compliance, model correctness, production suitability, or whether an external auditor accepts the evidence.

Roadmap Boundary

Optional prev_event_hash and event_hash fields make room for a future tamper-evident chain. Until that verifier is shipped and referenced from the release proof, public wording must stay at “hash-backed evidence” or “designed toward tamper-evident chains”, not “tamper-proof”.