MrZ’s go-discern is more than a renamed classifier wrapper. It separates provider observations from application recommendations, makes dual-provider calls explicit, and stores enough decision metadata to try a different policy without buying another inference.

I inspected the complete README, Jev and Clef adapters, request encoder, evaluator, HTTP transport, policy engine, and result types at revision 49a66e45e1fe25ba6102ad864d83d6908921aa76. The offline test suite, race-enabled suite, and fake-provider example passed locally. I did not call either live API, measure model accuracy or latency, or conduct a security audit.

Keep the answer and the permission to act distinct

  • Preserve every Choice option’s probability, yes-support values, rubric scores, and available model/usage metadata.
  • Without an explicit policy, return observations—not an invented acceptance decision.
  • By default, dual Choice acceptance requires matching choices and passing checks for both providers.
  • Disagreement, missing metrics, provider failures, and unmatched model-specific checks produce distinct non-accepting states.
  • Conflict preference and accepting a surviving provider are explicit opt-ins, not hidden fallback behavior.

Replay thresholds without replaying inference

  • Recommend applies a new policy to saved observations locally; it does not call Jev or Clef.
  • The offline example accepted its fake billing choice, then rejected it under a stricter Jev probability threshold.
  • Watch Overall.Uncovered: the same example reported overall accepted while its outage question had no_rule.
  • Require policy coverage for every safety-critical question yourself; overall acceptance covers only questions that actually have rules.

Those example probabilities are deterministic fixtures, not measured Jev or Clef answers. Provider agreement can still be a shared mistake; thresholds need validation on your own labeled workload.

Make billing and retention boundaries explicit

  • Validate and prepare selected requests before dispatch; dual-provider evaluation runs both calls concurrently and waits for both.
  • The inspected transport disables redirects and removes the replayable request body hook; no SDK retry or silent provider fallback is added.
  • Custom retrying transports are outside that boundary, and cancellation is not an exactly-once billing guarantee.
  • Result records omit state, instructions, credentials, and provider bodies, but identifiers and model outputs can still be sensitive.

The library currently accepts text and JSON, not Clef’s multimodal inputs. Its local CPU benchmarks and maintainer-reported live contract observations should not be mistaken for reproduced model-quality or end-to-end performance results.

Typed output is not a correct decision. The useful contribution is inspectable control logic around decisions, including an explicit account of what the policy did—and did not—cover.