Conformance levels L0 → L5

A sender or collector is ARP-conformant at level Nif it implements all features at levels ≤ N. AgentMinds' collector requires L0, recommends L2; cross-site recommendation filtering unlocks at L2+.

L0

Bare envelope

Required fields
agentreport.summarywarnings[].message
What this unlocks

Reports stored. Nothing else.

L1

Severity + lifecycle hint

Required fields
L0 +warnings[].severitywarnings[].categorywarnings[].status
What this unlocks

Dashboard sorts your warnings; status filtering kicks in (resolved warnings stop resurfacing).

L2

Cross-site dedup

Required fields
L1 +warnings[].fingerprintwarnings[].first_seenwarnings[].last_seenwarnings[].count
What this unlocks

Cross-site recommendation filtering activates. The pool stops recommending things you already have. (Sweet spot for most adopters.)

L3

Learned-pattern lifecycle

Required fields
L2 +memory.learned_patterns[]Pattern.fingerprintPattern.confidencePattern.status
What this unlocks

Your agents contribute to the pool. What you learn becomes recommendations for others (privacy-safe; only fingerprints leave your tenant).

L4

Telemetry + skills

Required fields
L3 + (telemetry OR project_info.skills)
What this unlocks

OTel GenAI / OpenInference traces flow into the pool. Skills self-describe via SKILL.md.

L5

Verifiable identity

Required fields
L4 + W3C Verifiable Credentials binding
What this unlocks

Planned for ARP 2.0. Cross-org agent delegation with cryptographic identity proof (AGNTCY OASF Identity Service).

Quick path to L2

Most adopters should target L2— it's the level at which AgentMinds' recommendation pool becomes useful. The fastest route:

  1. Use the Python or Node SDK — both ship L2-conformant payloads automatically.
  2. Or, if you're writing your own collector, compute fingerprint = sha256(agent_name + "::" + normalize(message)) client-side and timestamp every warning.

Reference