DossierRelay / judge guide
One synthetic case.
Every step reviewed.
Collect a fictional client's documents, prepare a response, approve its exact content and inspect the case export. The synthetic Business exercise is the primary judging route.
Choose your testing environment. The local demo opens directly with no account or access key; cloud AI and email are disabled. For the AWS judge workspace, use the access instructions supplied privately by the project owner. Use synthetic content only, and the SES success simulator for an authorized AWS email exercise. This public guide contains no credentials or active client invitation.
WORK-04 candidate documentation · 26 October 2026. These page changes have not been deployed. The prior bounded run described below does not verify this candidate.
PRIMARY JUDGING ROUTE
Synthetic Business exercise
- Create a Business case. Open the local workspace directly, or use your private access for the AWS judge workspace. Use the name “Synthetic Judge Client” and subject “Synthetic onboarding — judge exercise.” Set the recipient to
success@simulator.amazonses.com. Record consent for this controlled exercise. Add a message asking for the company profile, project brief and unsigned engagement-letter fixture. - Prepare and inspect. Create the private case-specific client link before final approval. Draft manually, or use the available Amazon Nova analyze, summary and answer controls within the authorized test allowance. Read the output and source-message references. AI receives bounded conversation excerpts and file metadata; it does not read uploaded document bodies or decide whether documents are acceptable.
- Approve the exact response. Review the recipient, subject and full message, then approve the current version. Any later case input or consent change requires a fresh review and approval. For an explicitly authorized email exercise, send once through SES. If the result is uncertain, stop and reconcile it instead of repeating the send.
- Use the client portal. Open the existing case-specific client link privately, reply and upload the three practice files below. Keep the invitation token out of screenshots and reports. An upload remains awaiting review until the owner makes an explicit decision.
- Review each document. Open the verified preview, inspect the bytes and SHA-256 result, and save a reasoned assessment. The engagement-letter fixture is unsigned and nonbinding. Accept it only as a controlled exercise fixture, never as a real signed agreement.
- Complete and export. Once all three exercise requirements have been explicitly reviewed and accepted, complete the case and inspect its export. Check the case version, decisions, gaps and evidence limits. Recorded review controls do not establish document authenticity or qualified human acceptance.
In the local synthetic demo, draft manually and share on the client page: cloud AI and email are disabled.
This is a procedure for a separately authorized judging session. It does not authorize a new model call or email, or claim that this candidate has passed the procedure.
PURPOSE-WRITTEN TEST DATA
Practice files
These downloads contain the approved synthetic fixture bytes. They contain no real customer documents. The same downloads are available on the challenge page.
| File / size | SHA-256 of original bytes |
|---|---|
company-profile.txt425 bytes | 5671fa3fd63b2289c4abee54c44e7a2bdec7373b3dcfb8d625ea41ea8fa32fbc |
project-brief.txt468 bytes | 1e483b8c1b0272c6dcb48b98ab75ebe066b60e4d55465edad33eb9f4b23a1ea7 |
engagement-letter-example.txt431 bytes | 3c209e299fd9d7b72cc75b105ada40228d1e89cbf0b58659f6b42083f1d16dfb |
Hashes identify bytes; they do not establish authenticity, a signature or legal validity. No requester correspondence is part of this judging route.
AWS ARCHITECTURE
AI prepares. The application controls actions.
This judge release supports Amazon Nova only. The AI path is Application Lambda → AgentCore Runtime → Amazon Nova on Bedrock. Consent checks, case versions, document review, exact approval, durable send reservation and SES sending remain in the application, outside AgentCore.
The judge profile allows only the SES success simulator. A provider message ID establishes SES acceptance, not delivery to an inbox or reading. An independent event receipt, when available, is separate evidence.
The diagram describes the source architecture. An application protocol or configured runtime version numbered 2 is not a claim of “AWS platform V2.” Source configuration is not a deployment receipt.
PRIOR BOUNDED RECORD / 26 OCTOBER 2026
What was observed
The supplied acceptance report and final export record a synthetic-content case completed at version 17, with three accepted documents, zero gaps and a checked export. This is prior-run evidence, not a new candidate acceptance result.
| Step | Record | Limit |
|---|---|---|
| AI operations | Three reported actions: analyze, summary and answer. One analyze receipt retained, with distinct AgentCore and Bedrock request IDs. | Summary and answer request IDs were not captured. The answer did not explicitly cover all requested review checks. |
| Consent / version | Consent withdrawal cleared the AI draft. A stale draft save showed a refresh message. | Stale HTTP status and error code were not captured. A direct forbidden-recipient send was not tested. |
| One manual response was approved after AI draft invalidation; one SES success-simulator acceptance is recorded. | The unchanged AI draft was not sent. Pre-send approval digest/version snapshot absent. No SNS delivery, inbox or reading confirmation. | |
| Review / export | Three uploads, three explicit synthetic assessments, completion and export; fixture byte hashes match. | Review was operated by AI: qualifiedHumanReview=false. The unsigned fixture proves no agreement. |
| Synthetic scope | Content and simulator destination establish a controlled test. | case.synthetic=false is expected for an ordinary case: the flag identifies a recognized demo scenario, not whether the content is fictional or which provider runs. Original evidence bytes are unchanged. |
Quota counter deltas of 3 application model reservations, 3 runtime model reservations and 1 email reservation support bounded dispatch counts; they are not independent success or billing receipts.
Evidence basis: browser-acceptance.json, browser-evidence/final-proof-pack.json, post-e2e-quota-counters.json and LIVE-REVIEW.md in the supplied review package. These files are not exposed as public downloads here. Material assurance remains in progress; no formal human acceptance or production readiness is asserted.