Know what a test result actually means.
A passing test is only useful if you know what was checked. Intent Evals keeps the definitions behind the results clear and versioned. Intent Labs runs the tests and explains what happened.
The approach covers AI skills, agent loops and complete workflows. It is not tied to one model provider. Each result still applies only to the setup and checks recorded in that run.
Versioned result definitions
These technical identifiers are for systems that produce or verify evidence. If you only want to understand a result, start in Labs.
Normative means the definition is adopted. Draft means it is still being developed. Reserved means the name is held for future use. These labels describe the definitions, not whether a system passed its tests.
| Predicate | Status | Asserts |
|---|---|---|
gate-result/v1
|
NORMATIVE | One deterministic or behavioral gate's outcome for one subject: the gate's identity and version, a four-state decision (pass / fail / advisory / error), the reasons, the policy the gate ran under (hash-pinned), and the exact inputs it saw. |
retraction/v1
|
NORMATIVE | An append-only, signed record that a previously published attestation has been retracted, with a closed-set reason class. |
skill-refiner-pass/v1
|
NORMATIVE | One Skill Refiner improvement pass over a SKILL. |
dashboard-render/v1
|
DRAFT | An attestation that a specific dashboard render was produced from specific verified inputs — the sign-your-own-homework layer for the labs reports hub. |
human-review/v1
|
DRAFT | A curated human verdict about a skill (thumbs, score text, annotation) captured as a first-class signal, kept separate from automated grades and never folded into them. |
validation-result/v1
|
RESERVED | Reserved for future structured validator output. Its body is not yet specified. |
skill-snapshot/v1
|
DRAFT | A content-addressed snapshot of a skill artifact at a point in time — the immutable subject other predicates reference. |
runtime-receipt/v1
|
DRAFT | A receipt that a specific runtime execution happened with specific inputs and outputs, for replay-fidelity claims. |
regression-pack/v1
|
DRAFT | A versioned bundle of sacred regression cases a skill must keep passing; a regression on a sacred case blocks release regardless of average improvement. |
eval-verdict/v1
|
RESERVED | Reserved for a single judged criterion verdict as a standalone signed predicate. |
Why this host stays separate
A result definition cannot drift whenever the dashboard changes. Predicate URIs stay on this host. Clarifications may be added, but a breaking change creates a new version instead of rewriting the old meaning. The zone is DNSSEC-signed and CAA-pinned to Let's Encrypt.