Evidence,
not explanations.
Establish a checkable account of what happened.
Agent logs describe part of a task. They do not automatically connect the requested work, the actor, the authority granted, and the external result.
Evidence is the False Systems capability area for preserving those relationships, with records that distinguish direct observation, inference, and uncertainty.
Discuss the evidence you need ↗“Done” is a claim.
What supports it?
An agent’s explanation can be useful, but it is not independent evidence of its own actions. Independent evidence needs a source separate from that self-report, with a stated trust boundary: who controls the observer, what it captures, and whether the worker can alter its records.
A runtime record may establish that a request was sent. A receipt from the destination may support a different claim. Neither should be silently promoted into proof of a complete, correct business outcome.
Observability helps collect and query signals. Evidence adds the question: how do these signals relate to the work, its authorization, and the conclusion being drawn? These responsibilities can overlap.
Follow the claim
back to its source.
- Requested workWhat was asked, and by whom?
- Authorized delegationWhich grant, scope, and issuer?
- Agent or workerWhich actor and execution context?
- Observed actionWhat did the observer actually capture?
- External outcomeConfirmed, inferred, or still unknown?
- Evidence and verificationWhat supports the account, and what can be checked?
Keep the distinctions.
An illustrative account, not an API schema or a genuine production capture.
- Attribution
- The human, agent, workload, or infrastructure identity associated with the observation—and the source of that association.
- Authority
- The grant, issuer, scope, and authority chain relevant to the action. A valid signature alone does not establish permission.
- Continuity
- The work identifier, execution context, earlier records, and handoffs needed to understand where the event belongs.
- Integrity
- Provenance and integrity material that a verifier can check for alteration, together with the required trust information.
- Observation
- The observer, captured action, supporting artifacts, and capture limits. A record of an attempted request says exactly that.
- Uncertainty
- Separate inferred conclusions, missing observations, and unconfirmed outcomes from facts directly captured.
Illustrative example / A deployment request
- Observed
- The worker sent an authorized deployment request; a runtime observer recorded the attempt.
- Inferred
- The agent reports that the new version is serving traffic. That report needs independent support.
- Unknown
- No destination confirmation was captured. The deployment’s external outcome remains unconfirmed.
A valid record.
A bounded conclusion.
A reviewer with the records, relevant trust material, grants, and supporting evidence should be able to check the account independently of the agent. Verification is limited by what those materials establish and what the reviewer is authorized to access.
A signature authenticates a record and helps detect alteration. It does not make every assertion true, establish complete observation, or prove the action was correct.
A missing event does not prove that nothing happened. A request sent does not prove its remote effect completed. Uncertain outcomes need checking before retrying an action with external effects.
Evidence names a platform responsibility backed by observation, storage, and verification components—not a new standalone service. It does not claim that all infrastructure actions are captured.
Decisions. Execution.
A supporting account.
Control Plane connects work to organizational decisions and delegated authority. Habitat manages Operations, workers, runtime history, and retained results. Evidence relates observations to that context.
Publication of shared records is separate from local execution and runtime durability. Gaps in published evidence must remain visible, rather than implying that execution never occurred.
Inspect the architecture ↗What would you need
to establish?
Bring a sensitive action, an uncertain outcome, or an account that currently depends on the agent’s word.