---
pretty_name: AIREP Embedded Evaluation Profile v0.1
license: other
license_name: airep-dual-license
license_link: https://github.com/halvrenofviryel/ai-runtime-evidence-protocol/blob/85f99b79fbe78377a53f03d2fafc5331eab42f71/LICENSE-CC-BY-4.0.txt
tags:
- ai-evaluation
- ai-safety
- ai-governance
- audit
- provenance
- reproducibility
- airep
---

# AIREP Embedded Evaluation Profile v0.1

> **This is not a training dataset or benchmark.** It is a Hugging Face distribution mirror of an
> experimental evaluation-evidence profile, its schema basis, registry and fixtures. The canonical
> specification history lives in the AIREP GitHub repository. Byte identity between this mirror and
> its source commit is a distribution-integrity property, not independent scientific verification.

**Experimental evaluation-evidence contract for AIREP v0.2.**

## Profile identity

| | |
|---|---|
| Profile identifier | `airep.embedded-evaluation` |
| Profile version | `0.1` |
| Carrier | AIREP wire version `0.2` |
| Schema basis digest | `sha256:ce57bd493d1a3166051bc9029972d6be215793edcd3f1b314534959659b3da8b` |
| Canonical GitHub source | https://github.com/halvrenofviryel/ai-runtime-evidence-protocol/tree/85f99b79fbe78377a53f03d2fafc5331eab42f71/spec/airep/v0.2/profiles/embedded-evaluation |
| Source commit of this mirror | `85f99b79fbe78377a53f03d2fafc5331eab42f71` — fixed source snapshot for the files in this publication (including its historical README) |
| First publication commit | `2b62580aa05302cc978c2cb5a2a55d8a0a215b55` — schema, registry and both fixtures are byte-identical to that commit; only `PROFILE_README.md` changed since |
| Status | Experimental companion profile — not a ratified standard, not a fifth AIREP artifact family |

The five files here are byte-identical to the source commit's `spec/airep/v0.2/profiles/embedded-evaluation/`
directory (the profile README is published as `PROFILE_README.md`); `MANIFEST.sha256` lists their digests.

## Publication policy and current exporter scope

**Frozen, source-pinned publication; not a continuously updated mirror of canonical `main`.**
The schema, registry, fixtures and `PROFILE_README.md` retain the exact source pin above.
Card clarifications do not refresh those source bytes. A new source edition requires an explicit
new publication revision and pin.

`PROFILE_README.md` is historical source text, not a current parser-support statement.
The current exporter parses LightEval `results_*.json`; native Inspect `.eval` and arbitrary
OpenEvals formats are not currently supported. Explicit timezone-aware start/end timestamps
are required; a timezone-naive filename is never silently labelled UTC. Profile validation is
not an assurance result or safety certification. Declared context is recorded, not established.

## What it binds

Evaluator and evaluated-party context with a declared independence basis; target model/system/
checkpoint identity; evaluator access tier, capabilities and limitations; harness/task/environment/
safeguard configuration; explicit measurement execution and observed states; content-addressed
evidence with visibility/withholding state; redactions, deviations, retries and incidents; and
optional verification references.

Three negative-state guards are enforced by the schema: `NOT_RUN` can only pair with
`NOT_MEASURED`/`NOT_APPLICABLE`; `PASS`/`FAIL` require `RAN`/`PARTIAL`; `INVALIDATED` never yields
`PASS`/`FAIL`. `example.not-measured.json` preserves `NOT_RUN → NOT_MEASURED` explicitly.

## Claim limitations

A profile `PASS` means only that the `profiles["airep.embedded-evaluation"]` payload validates
against the exact digest-pinned schema basis. It does **not** establish:

- evaluator independence or competence (`independent-declared` is a declaration, not proof);
- that the named model, revision or configuration was the one actually executed;
- truth, representativeness, completeness or reproducibility of observations;
- model safety, alignment, security, regulatory conformity or deployment suitability;
- any AIREP assurance class — Core/Authenticated/Witnessed semantics are unchanged;
- validity of a platform `verifyToken`, which the profile carries only as `platform-verification` evidence.

Without an accepted basis for the exact identifier, the correct verifier result is `NOT_EVALUATED`,
never an implicit pass.

## Files

- `embedded-evaluation.schema.json` — self-contained Draft 2020-12 schema (basis)
- `registry.json` — digest-pinned registry for the AIREP v0.2 `--profile-bases` mechanism
- `example.measured.json` — synthetic measurement that ran and failed its declared criterion
- `example.not-measured.json` — synthetic planned evaluation that did not run
- `PROFILE_README.md` — the canonical profile README at the source commit
- `MANIFEST.sha256` — digests of the mirrored files

## Related

- Exporter Space (LightEval results → profile payload + evidence manifest; current parser LightEval `results_*.json` only, Inspect/OpenEvals not parsed): https://huggingface.co/spaces/phionyx/airep-evaluation-evidence
- Exporter source (LightEval results → profile): https://github.com/halvrenofviryel/ai-runtime-evidence-protocol/tree/main/integrations/lighteval
- Collection: https://huggingface.co/collections/phionyx/phionyx-runtime-evidence-and-measurement-6aa65830536927640b962c5a
- Technical note, *Access Is Not Yet Verifiability*: https://huggingface.co/blog/phionyx/access-is-not-yet-verifiability

## Licensing

The canonical AIREP repository is dual-licensed by file type: code under Apache-2.0
(`LICENSE`), specification text and schemas under CC BY 4.0 (`LICENSE-CC-BY-4.0.txt`), with a
patent non-assertion statement. The mirrored files are specification/schema/example material and
follow the CC BY 4.0 terms; this card declares `license: other` so those file-specific terms are
not collapsed into a single repository-wide SPDX claim.