Zephr
All articles

Why Zephr says 'not evaluated' instead of guessing

Trust over fluency: when evidence is missing or contradictory, the system stops rather than fabricate.

2026-07-294 min readdocs/product/vision.md

The product is inspectability, not recall

Zephr's product vision is simple: a useful result must carry provenance, temporal context, confidence, and an explanation of retrieval and trust decisions. If any of those are missing, Zephr abstains. The differentiator is inspectability, not the ability to sound confident.

This is unusual in a market that optimises for fluent answers. A model that always answers looks smarter than one that sometimes refuses. But a wrong answer with made-up confidence is worse than no answer at all when the user is about to act on it.

Three abstention scenarios

When two reviewed sources contradict each other, Zephr reports the contradiction instead of resolving it by majority vote. When no qualifying evidence exists for a query, Zephr says so rather than generalise from unrelated memories. When a semantic capacity limit is exceeded, Zephr degrades to lexical fallback and marks the result as bounded.

Each case returns a structured reason — contradiction, no qualifying evidence, capacity exceeded — not a confidence percentage that hides the problem.

Confidence is a reason, not a probability

Zephr does not emit synthetic confidence scores like 0.95. Confidence is expressed as source counts, review state, and the agreement or disagreement between sources. A claim supported by two reviewed sources that agree is labelled supported. A claim with contradictory or missing evidence is labelled abstained.

The absence of a numeric confidence score is deliberate. Percentages train users to trust a number they cannot audit. Reasons train them to look at the evidence.

Manufactured continuity is a bug

Zephr will not manufacture continuity to keep a conversation flowing. If a session ends without a valid handoff, the next session starts without those memories rather than reconstruct them from context. Honesty about what is known is the continuity policy.

That posture is backed by the verification ladder: T1 structural checks and T2 consistency checks are deterministic and CI-blocking. T3 judgement is explicitly unbuilt and advisory-only, with no authority to override a deterministic refusal.

Source evidence

This post is adapted from a source document in the Zephr repository.

Read the security model.

The documentation covers the architecture, tool contract, TrustBench methodology, and trust model in detail. Every claim links to its source.