Different records answer different questions
The previous delivery model is being replaced. All new inquiries are paused. Read Managed Verse →.
Application logs help explain how a service handled a request. Behavior records connect a declared traffic flow to an observed response. Infrastructure logs help an operator diagnose deployment and cleanup. These records are useful, but they are not interchangeable with a signed evaluation ledger.
The local evaluation foundations record authorized requests, policy decisions, execution starts, and results in a signed chain. Model activity and effect activity share a consistent attempt record. A separate behavior ledger explains journey scheduling and outcomes.
The legacy AWS runtime logs are not automatically promoted into that offline signed evidence system. Reviewers need to know which path produced a record before drawing conclusions from it.
What exact replay means in the reference
The replay reference verifies a sealed transcript and reconstructs its recorded decisions and public results. It checks the manifest, source chain, policy and adapter bindings, relevant artifacts, and the selected verification profile. Model bearing transcripts require the corresponding agent plan semantics rather than a planless verification path.
Replay does not call the adapter or ask the target to perform the request again. This is important when a request could have consequences: reading a historical result should not silently create a new action.
Exact transcript replay is a local foundation. It is not a browser recording, a snapshot of every cloud resource, or a restored copy of the entire environment.
Separate public records from private material
Public effect evidence retains digests and commitments instead of raw tool arguments, idempotency keys, or HTTP response content. This allows identities and relationships to be verified without treating the public ledger as a general store of sensitive payloads.
Dynamic tool calls can retain the required canonical arguments and prepared invocation material in a bounded, encrypted and authenticated local sidecar. Private replay validates that material against the public commitments and assignment provenance before reconstructing the call.
Missing, corrupt, or inconsistent material prevents successful private reconstruction. This legacy sidecar does not retain durable private HTTP projections or complete model conversation history. Digests can also remain linkable metadata; excluding payloads is not a blanket confidentiality guarantee.
Read the proof at its actual scope
| Verified property | What it does not establish |
|---|---|
| Signed chain integrity | That the record is remotely durable or that the supplied chain is the newest one. |
| Bound request and result | That every possible target action was observed. |
| Deterministic transcript reconstruction | That a cloud environment can be restored to the same state. |
| Matching artifact commitment | That the artifact remains available at a remote location. |
| Recorded execution order | Exact wall clock timing or cloud provider certification. |
The newest record must be established through an independently trusted current reference. Local signatures alone cannot prevent an older valid history from being presented as current.
Replay is different from recovery
Replay examines historical evidence without executing a target operation. Recovery may change retained state or reconcile a partially recorded operation. That difference requires a fresh authority check: historical permission does not authorize a new mutation.
The local recovery foundations bind retained history, the current agent plan, signing trust, private commitments when present, and a fresh supervisor authority. An uncertain result is not converted into a successful operation by retrying speculatively. Missing material must remain visible as a limit.
These older mechanisms operate within local storage and process boundaries. They do not specify the planned Managed Verse recovery or evidence delivery model.
Plan the debrief before the run
Record the exact scenario and tier, deployment identity, activity configuration, participant scope, and evidence sources at the start. During the session, retain the trace and request identities needed to connect a participant observation to service behavior. Record collection gaps and uncertain outcomes while their context is still available.
In the debrief, separate the participant's account, the service observations, and the evaluator's conclusion. Each answers a different question. Preserve only the material the exercise needs, with an explicit owner and retention decision.
Capture, restoration and research views for planned Managed Verse need their own stated release scope. These older examples do not establish availability before launch. Assess these legacy agent evaluation and exact replay foundations on their documented local scope.