Evidence Package Specification (EPS) v0.1

Overview

The Evidence Package Specification (EPS) is a draft interchange format for AI-generated outputs that must be verified, exchanged, and preserved over time. Think of it as OpenAPI for evidence: a portable object any system can export so downstream auditors, benchmarks, and regulators can inspect what happened without depending on the original runtime.

EPS v0.1 is a draft interchange envelope aligned with RFC-001: Evidence JSON v1.0 in the Dali repository. This page is the public-facing summary; the canonical spec is RFC-001.

Repo: github.com/yenklabs/Dali · Canonical spec: RFC-001: Evidence JSON v1.0 · Related: reproducible evidence bundles

Design goals

Portable — one self-contained record, not a scattered row in a benchmark table.
Verifiable — cryptographic hashes and replay metadata so outcomes can be re-checked offline.
Domain-extensible — legal citations are the first vertical; contracts, policies, and clinical guidance use the same envelope.
Benchmark-ready — every EPS instance can be scored against the verification taxonomy.

Evidence classes

Failures are one class of evidence, not the product. EPS supports multiple evidence_class values:

Class	Description
`failure`	Documented breakdown (fabrication, unsupported proposition, etc.)
`verified`	Authority exists, proposition supported, bundle complete
`disputed`	Competing interpretations or unresolved human review
`benchmark_artifact`	Synthetic or curated eval row with golden expectations
`replay_trace`	Deterministic re-run output sealed under a policy version

The Open Evidence Corpus (open-evidence-corpus) maps to evidence_class: failure in v0.1.

EPS v0.1 envelope

{
  "eps_version": "0.1",
  "package_id": "eps-001-mata-v-avianca",
  "evidence_class": "failure",
  "created_at": "2026-06-08T00:00:00Z",
  "domain": "legal",

  "prompt": "Research whether the statute of limitations was tolled…",
  "model": {
    "provider": "openai",
    "name": "gpt-3.5-turbo",
    "config_fingerprint": "sha256:…"
  },
  "context": {
    "retrieved_sources": [],
    "runtime_state_preserved": false
  },
  "output": {
    "text": "…",
    "citations": ["Varghese v. China Southern Airlines, 925 F.3d 1339 (11th Cir. 2019)"]
  },

  "sources": [
    {
      "authority_id": "cite-001",
      "citation_string": "Varghese v. China Southern Airlines, 925 F.3d 1339 (11th Cir. 2019)",
      "source_blob_sha256": null,
      "retrieval_snapshot_uri": null
    }
  ],

  "verification": {
    "policy_version": "dali-tier1-v0.2",
    "primary_outcome": "authority_not_found",
    "outcomes_by_authority": {
      "cite-001": "authority_not_found"
    },
    "summary": "Citation does not resolve in canonical reporter indices."
  },

  "taxonomy": "dali-verification-taxonomy/v0.1",
  "replay_hash": "sha256:…",
  "review_status": "documented_incident",

  "evidence_bundle": {
    "merkle_root": "sha256:…",
    "bundle_uri": null,
    "yenklabs_investigation_url": "https://yenklabs.com/failures/001-mata-v-avianca"
  },

  "annotations": {
    "human_reviewer": null,
    "reviewer_notes": null
  }
}

Required fields (v0.1)

Field	Required	Notes
`eps_version`	yes	Spec version (`"0.1"`)
`package_id`	yes	Stable identifier
`evidence_class`	yes	See table above
`prompt`	yes	User or system prompt that produced the output
`model`	yes	Provider, name, config fingerprint
`output`	yes	Model output text and extracted citations
`verification`	yes	Outcome(s) under a named policy version
`taxonomy`	yes	Taxonomy dataset or version URI
`replay_hash`	recommended	Hash of deterministic replay under fixed policy
`evidence_bundle`	recommended	Merkle root and/or bundle URI
`sources`	recommended	Retrieved or cited primary materials
`review_status`	optional	`documented_incident`, `peer_reviewed`, `draft`, etc.

Relationship to Dali

Layer	Role
EPS	Interchange format — how evidence is exported and exchanged
Dali	Verification engine — scores packages, seals replay hashes, runs offline eval
Open Evidence Corpus	Public archive — open-evidence-corpus on Hugging Face
Verification Benchmark	Measurement — whether evidence in a package can be trusted

Roadmap

v0.1 (draft) — envelope definition, legal-vertical examples, mapping from failure records
v0.2 — dali-evidence-packages Hugging Face dataset with full seed packages
v1.0 — JSON Schema, validation CLI (dali eps validate), cross-domain examples

Citation

@misc{yenklabs_eps_2026,
  title={Evidence Package Specification (EPS) v0.1},
  author={YenkLabs},
  year={2026},
  url={https://yenklabs.com/artifacts/evidence-package-spec-v0.1},
  note={Draft interchange format for portable AI evidence}
}