Tests, not demos
Most AI memory products show a screenshot and call it proof. Zephr's public benchmark is a CI-blocking test suite: every merged commit re-runs the same scenarios against the same fixtures, and a regression blocks the build. The goal is not a one-time audit; it is a reproducible gate that travels with the code.
TrustBench is built from fixture cards, not from ad-hoc prompts. Each card defines a corpus, a SHA-256 digest of that corpus, a set of deterministic scenarios, and a pass/fail/not-evaluated verdict. A pass means the behaviour is machine-decidable and reproducible, not that a model "feels right".
One measured feature, nine passing scenarios
Today TrustBench measures one feature: hybrid retrieval with abstention. The scenario list is concrete: hybrid golden ordering, foreign-scope exclusion, stale-evidence supersession, semantic paraphrase ranking, lexical fallback when semantic search is disabled, contradiction abstention, temporal as-of gating, clean empty-result handling, and semantic capacity-overflow abstention.
All nine scenarios pass. That is the only numeric claim the report card substantiates. The fixture vectors are hand-authored, so the result measures retrieval logic — not a production embedding model. We say that explicitly because it is the difference between an honest claim and a misleading one.
What we do not claim
Five features are marked not-evaluated: atomic idempotency, provenance/claim support, scope admission, transport adversarial resilience, and UI honesty. These are not hidden footnotes; they are listed in the public evidence report with the reason each lacks an artifact.
No contradiction-abstention precision percentage appears on the marketing site. The corpus contains a single contradiction pair, so no honest denominator exists. No end-to-end recall@k score is published, because TrustBench does not run on a labeled relevance corpus. The product surface tells you what is measured, what is not, and why — in the same breath.
A gate, not a trophy
TrustBench is wired into CI so that a regression cannot slip through. The marketing page mirrors the same data that the gate uses. If a future release breaks foreign-scope exclusion, the page will not be able to honestly claim the scenario passes, because the CI gate will have failed first.
That coupling is the point. Trust is inspectable only when the public claim and the internal gate share the same artifact.