Supply trusted verification evidence for supervision

SupervisionEvidence supplies the host’s artifact revision, task constraints and verification receipt. A recommendation for completion review must rest on current evidence, not on an Agent claiming that tests passed.

Example: the previous report passed, but its revision has not been verified

An order report changes from v1 to v2. A successful v1 receipt does not verify v2 even if a model finds its contents complete. In a multiuser system, also prevent reuse of another task’s verification records. Read a consistent snapshot for each observation and bind verification to user/session/run and artifact revision. A scope or version mismatch prevents a completion-review recommendation.

1. Connect your evidence reader

import io.agentscope.core.agent.RuntimeContext;
import io.agentscope.extensions.judge.jev.supervision.JevSupervisionMiddleware;

var source = new JevSupervisionMiddleware.EvidenceSource(request ->
    evidenceRepository.readAuthorizedSnapshot(request));
var context = RuntimeContext.builder()
    .userId(userId).sessionId(sessionId)
    .put(JevSupervisionMiddleware.EvidenceSource.class, source)
    .build();

evidenceRepository stands for an application interface taking JevSupervisionMiddleware.EvidenceRequest and returning Mono<SupervisionEvidence>. It checks ACLs, reads the current artifact and verifies the record’s provenance instead of asking a model to infer success from prose. Pass context to agent.streamEvents(messages, context) for this call. See Continuous supervision for middleware registration. Evidence reading and judgment share a budget; a read timeout is never evidence of completion.

2. Bind verification to the current snapshot

SupervisionEvidence describes revision, repositoryStatus, diff, changedFiles, instructions, documentationRequired and optional Verification. It can represent code changes or versioned business artifacts such as reports.

Verification field	Meaning in report verification
userId / sessionId	User and session authorized to see this report
runId	UUID of the current supervision invocation, from request.scope
revision	Artifact version actually checked; must match the evidence revision
source	Check identifier with provenance the host can verify
passed / summary	Actual check outcome and a summary

A successful check, nonempty source, matching scope and matching revision are all required. Do not relabel an old receipt as v2 to make it match; rerun the check against v2. For code tasks the revision must cover uncommitted changes; Git HEAD alone is insufficient. Setting passed=true in JSON does not establish trustworthy provenance.

3. Consume advice while retaining responsibility for final review

Missing evidence source, stale receipt, failed verification or truncated material prevents completion review. Consume REQUEST_VERIFICATION or MANUAL_REVIEW while the task continues normally. REVIEW_COMPLETION means the task is ready for review. It does not publish, approve or change task state for you.

Run and verify

Set JEV_CP using the example build instructions:

java -cp "$JEV_CP" io.agentscope.examples.jev.JevSupervisionExample

The example records exactly one local order-tool execution. The host supplies a receipt for order-42-v1 using the identity and run UUID from the current request.scope. Scripted JEV judgments then yield REVIEW_COMPLETION with verified revision true. Full source This example verifies evidence integration and scope. Your application still owns the real build system, artifact versioning, authorization and cross-process recovery.