Validation
Dated synthetic tests, bounded real model checks, and deployment proofs are recorded separately. Source implementation does not by itself prove a live rollout or general reliability.
On this page
Recorded validation evidence
The table keeps the environment, observed result, and qualification together. “Undated” means the source records the check without assigning it a calendar date; it is not a new date estimate.
| Date | Check | Environment and scope | Observed result | Limits and source |
|---|---|---|---|---|
| Isolated container checks | Shared source in an isolated ARM Linux Compose project under Colima; synthetic fixtures; Python 3.12, Node.js 22.22, pinned PostgreSQL/Qdrant images, and Ollama Cloud gpt-oss:20b. |
Preparation, server import, and correct Atlas release answers passed. PostgreSQL and Coordination restart preserved existing message IDs; a custom-format dump restored the same four messages. | Checkpoint persistence does not establish clean-target restoration, crash behavior, or exactly-once effects. Pinned validation source | |
| Undated | Contract and retrieval fixture checks | Python contract suite, deterministic model double, in-memory checkpointer, real MCP server, and a separate real-BGE smoke test over three fictional documents. | Preparation, exclusions, hashes, GCS reads through a double, imports, Qdrant round trips, request accounting, routing failures, MCP boundaries, and invalid requests were covered. Hit@1, Hit@5, and MRR@5 were 1.0 for three labeled questions. | These checks use synthetic documents and model doubles; the metrics are not a production reliability score. Pinned validation source |
| Bounded real-model follow-up | Isolated Ollama Cloud gpt-oss:20b check, followed by a private glm-5.3-flash route check. |
Two fresh preparation questions and one follow-up passed in isolation; two fresh gateway questions and one follow-up passed after the private correction deployment. | This bounded sample does not establish general reliability. Representative answer quality, correction behavior, other hosts, resource and failure limits, and additional client identities remain unverified. See known limits below. |
Known limits and bounded follow-up
Representative answer quality, correction behavior, other CPU/host targets, resource and failure limits, real GCS access, optional Docling conversion, and additional client identities remain unverified. The latest local repair and heading-whitespace correction passed two fresh preparation questions and one follow-up in an isolated Ollama Cloud gpt-oss:20b check on 2026-10-13. The correction was subsequently deployed on the private glm-5.3-flash route; two fresh gateway questions and a follow-up passed. This bounded result does not establish general reliability. Enterprise shared memory, Kubernetes/SSO, commercial scale, and measured customer AI latency are unimplemented and unverified. Tailscale private communication is selected but not established by this pass.