phionyx/airep-evidence-cases

AIREP Evidence Cases

Source pinning

Frozen, source-pinned publication — not a live mirror of the canonical repository's `main`.

Later commits on the canonical repository are not reflected here unless a new publication says so. Byte identity with the snapshot is a distribution-integrity property, not independent verification.

A public reference/conformance corpus, not a blind benchmark, field measurement, model leaderboard, or certification. AI governance and AI safety researchers can inspect runtime evidence, agent governance, verification, reproducibility and tamper evident records. Model evaluation is a possible context, not a result measured by these first-party examples. The train name is a Dataset Viewer transport split only: no learned model or random train/test split is provided.

v01_reference contains 6 policy/configuration cases over three exact frozen v0.1.2 inputs (one standalone JSON and two JSONL chains). v02_beta_lifecycle contains eleven canonical beta lifecycle cases. Each row is an input + operator configuration, not a unique real-world incident. Do not merge deterministic identities across fixtures. complete and unproven-independent have the same bytes and different independence policy; both have distinct case IDs.

source/ retains signed bytes unchanged with fixture_sha256 in the index. No dataset metadata was added to signed JSON; no artifacts were re-signed. expected_disposition is a declared bounded expectation with a pinned source, while actual-results/ and observed_* record executed reference outputs. test_passed reports a regression assertion, not positive evidence validity. The complete beta case remains INCOMPLETE without intended-target inventory. Missing receipt/execution records do not establish non-delivery/non-execution. Accepted keys are supplied assumptions, not verified people; distinct demo keys are not independent organizations. No input or evidence URL is fetched.

Canonical source: AIREP. Pins: v0.1.2 44387bd43cc06ba656eaa7ff670be5c8e3220aca; beta 8a6c01ecce457aa94330c0ed7219e4c56ebfe771; current publication-source review ff5c3551052251726c0ed878dcc23a44e305bd93. Source lock and per-file manifest travel with this distribution. The frozen v0.1 parser/signature-encoding limitations are preserved; beta validation rules are not retroactively applied.

Interactive Lab · External evidence with version/role bounds. The external v0.1.2 producer and frozen v0.2 consumer results must not be combined into same-version interoperability. These are links to the public registry, not new external implementation runs by this dataset.

Code/examples are Apache-2.0; normative source text/schemas are CC-BY-4.0. Original license texts and patent statement are retained. Public deterministic test key material has no production trust significance. There are no user uploads, private collaboration materials, emails or external unlicensed bundles.

Use exact revisions: load_dataset("phionyx/airep-evidence-cases", "v02_beta_lifecycle", revision="<full published HF commit>"). Select one config; raw sources and actual-output JSON are not extra Viewer rows. Cite the canonical AIREP research artifact and the explicit dataset revision separately; this mirror is not a new independent research finding. Export date is not a new field-study date.

Paper

AIREP: A Protocol for Per-Decision Evidence in AI Runtime Governance https://arxiv.org/abs/2608.21363

This dataset is a source-pinned evidence/conformance corpus. Linking the paper does not make the dataset an independent validation or interoperability result.