Evidence and limits¶
What a judge can verify¶
The strongest proof is a short chain:
- Create and approve a goal-specific map.
- View a quest without starting it.
- Begin and submit a teach-back.
- Review the score, evidence, and proposal trace.
- Confirm a gap.
- Restart and retrieve.
- Show the changed mission.
- Forget the memory and show the tombstone.
For a live run, also show the non-secret provider metadata, CockroachDB persistence, DVI filters, Managed MCP read-only trace, S3 version ID, checksum, and artifact lineage.
Claim levels¶
| Claim | Minimum proof |
|---|---|
| Product flow | Passing live journey or labeled replay journey. |
| Browser UX | In-app Browser at required widths with no overflow or console error. |
| OpenRouter inference | Live response with provider, model, request ID, and error handling. |
| Durable memory | CockroachDB restart and retrieval test. |
| DVI retrieval | Exact 1,024 vector and learner/concept/status/consent filters. |
| Managed MCP | Read-only operation and source trace. |
| S3 artifact | Version ID, checksum, lineage, and failure consistency. |
| Production path | CloudFront → API Gateway → Lambda → services smoke test. |
Known limits¶
- A replay run is not live-provider evidence.
- A model proposal is not verified truth.
- Learner name is session-scoped. Orion does not claim cross-device identity.
- Client Whisper is a local draft tool. It is not a server transcription record.
- OpenRouter billing is outside the AWS USD 20 budget.
- Bedrock/Luna is future-only in the active deployment.
- The public docs site does not expose private tickets or secrets.
Evidence files¶
The repository keeps detailed, non-secret evidence under docs/evidence/.
Use the existing evidence files for exact timestamps, commit IDs, endpoints,
and provider results. Do not copy a historical pass into a new live claim.