Disciplines · Audits

Eve provider and route resilience — 2026-09

The retained live probe sent the exact model, Baidu-only tool route,

1section2 minread
Markdown source
On this page

Task 15.2 is closed at source commit e4b3501f488dd7f2141617565f4c21d23f28f51c. The machine record is eve-sota-provider-route-resilience/2026-09-16.json, validated by eve-sota-provider-route-resilience.schema.json.

The route now treats three recovery mechanisms as different products:

  • Endpoint failover stays on the exact reviewed turn model and inside the request-class allowlist. Plain calls may use Baidu, DeepInfra, or StreamLake; tool calls narrow to Baidu and require every request parameter.
  • Model fallback is disabled. The request sends no models fallback array, and a response naming any other model is rejected before buffered stream content is released.
  • Deterministic degradation is a preflight-only recovery. Three consecutive failed provider turns open a 30-second process-local circuit; the next call receives a 503 naming the deterministic message route before a model, tool, confirmation, or durable model turn starts. An in-flight failure remains an honest refusal because replay could duplicate a tool effect.

Paired faults exercise the positive and negative halves: a reviewed non-tool endpoint is allowed, an unmeasured tool endpoint is refused; the exact model and Baidu tool provider are accepted, cross-model or wrong-provider responses release only an error; and a circuit-open request degrades before work while an ambiguous in-flight turn does not replay.

Live boundary#

The retained live probe sent the exact model, Baidu-only tool route, allow_fallbacks, parameter enforcement, training denial, and ZDR to the US regional endpoint. The account returned HTTP 403 before model routing. That is useful fail-closed evidence, not a live failover success. The receipt retains no credential, prompt, response content, or tool arguments. The paired fault tests therefore carry the behavior claim; the live receipt carries only the regional admission posture.

The gateway-reported model and provider checks are application-layer evidence, not cryptographic upstream attestation. The circuit is process-local; Task 15.5 owns fleet-wide health, quarantine, canary, drift, and rollback automation.