Core concepts
Five ideas explain the whole platform. Everything in the UI and API is one of these.
1. Events
An event is one thing an agent did: a tool call, a model call, a file download, a grant. The canonical form is the Agent Action Record (AAR), an open export profile. OTel GenAI spans are accepted too and mapped onto the same model. Every event carries the action, the target, the credential used, and correlation ids.
2. Emitters
An emitter is a signing identity for one agent service. Each emitter is an ECDSA P-256 keypair registered to your organization; batches are signed with the private key and the signature is verified before storage. Telemetry authenticates by signature, not by bearer token, so no secret travels with the data. Register one emitter per service (prod, staging, a single bot) so batches stay independently verifiable.
3. Chains
Within your organization, events are ordered into hash chains: per emitter batch, per connector, per day, and per agent session. Each record links to the previous one, so any later edit breaks the chain visibly. Days are sealed with roots (a Merkle-style root over the day's chains) roughly 48 hours after they close; open days are never rooted, and ingest into a rooted day is rejected.
4. Attribution
Attribution binds an event to an agent record: the workload identity that acted, the human or team that owns it, and the credential that authorized it. Exact matches on vendor + vendor_agent_id attribute at 0.95 confidence. Records that match nothing are kept as unknown, never dropped, and surface in Candidates for review. This is deliberate: in an evidence system, an unattributed event is still evidence.
5. Findings
Findings are what the chains know that dashboards don't: shared credentials across agents, reach beyond granted authority, unowned agents with live activity, custody gaps where telemetry stops mid-session. Every finding carries its full evidence trail (one click from summary to raw signed events) plus a status and history that survive across runs: open, acknowledged, resolved, or stale when a source may be silent.
How they fit together
your agents ──signed batches──▶ ingest ──▶ chains ──▶ attribution ──▶ agents + candidates
│
▼
findings ──▶ exports (auditor-verifiable)
The same evidence model serves live telemetry, Langfuse imports (LangSmith is rolling out), and identity connectors, so a backfilled trace and a streamed event are equally provable.