All docs
Docs
Core Concepts

Evidence bundles

Evidence bundles make agent answers, proposals, and replays inspectable.

Overview

A bundle captures the rows, chunks, memory facts, redactions, and capability invocation metadata used for an output.

Evidence ids and resource records are produced by Synapsor when a capability runs with evidence enabled. Store the returned run_id, evidence_bundle_id, proposal_id when present, workflow, capability, business object ids, principal, tenant, request_id, and timestamp with your app decision.

The evidence id is one lookup key, not the only lookup path. A support lead can also start from a ticket, invoice, order, customer, workflow, capability, principal, tenant, request_id, query fingerprint, proposal, settlement, or time window.

At scale, humans should find evidence through indexed run/evidence/proposal/audit search by project, database, workflow, capability, principal, tenant, business object id, proposal or settlement id, query fingerprint, and time window. A billion reads should never mean scrolling Activity search or model logs.

For automatic approvals, settlements, or merges, reviewers should follow settlement -> proposal -> run -> evidence bundle -> query audit -> replay. That chain answers what the agent saw, what it proposed, why policy allowed it, and what changed.

Use READ RESOURCE, FROM AGENT ACTIVITY, synapsor_evidence_bundles, synapsor_evidence_resources, replay, or the workspace Activity/Evidence/Proposal views when you need to inspect the stored bundle later. The Workspace result screens show a reviewer lookup recipe after each successful capability test.

Evidence is required for audit-grade agent workflows.

Bundle ids should travel with responses, proposals, and incident reports.

Evidence bundle lifecycle

Evidence turns a capability response into an inspectable database record instead of a loose log trail.

root rowshybrid chunksgoverned memorypolicy version
evidence bundlesource ids, scores, redactions, invocation metadata
answerproposalreviewreplay
1
Context reads

Root rows, lookup rows, hybrid-search chunks, and governed memory are collected under hidden session scope.

2
Capability output

The reviewed capability returns an answer, decision, or proposal id.

3
Evidence lookup record

Synapsor stores source ids, scores, policy versions, redactions, and invocation metadata under a searchable record.

4
Review

Users, reviewers, or auditors inspect the bundle without reconstructing prompt glue.

5
Replay

The captured context can be compared with current policy, memory, and data.

Developer notes

  • Require evidence for customer-facing decisions.
  • Display source versions when an answer cites policy.
  • Store run ids, evidence bundle ids, proposal ids, business object ids, tenant/principal scope, request ids, and timestamps alongside app decisions.
  • Expose evidence lookup by indexed run/activity filters, not only by paste-a-handle forms.
  • For automatic settlements or merges, link settlement -> proposal -> run -> evidence bundle so reviewers can inspect the original read and the write decision.
  • Use replay to compare original and current behavior.