The 30-second path: Install → Quickstart → Real input/output.

1. Project Overview

Agent Comparability is a framework-independent specification (spec/SPEC.md), a frozen conformance suite, and a thin reference implementation that adjudicates comparability of agent evaluation records — never the scores themselves.

Does it actually refuse unsupported claims? See experiments/scenarios_001/RECORD.md: four real-pipeline scenarios (real .eval logs → converter → core) covering all four verdicts — including the confounded comparison and the missing-trajectory case, both correctly refused rather than packaged as supported conclusions.

Who is this for: developers comparing agent/prompt/harness evaluation results, benchmark authors documenting what conclusions their data supports, and teams building reliability pipelines on top of existing eval frameworks.

How it relates to existing tools

2. Problem & Motivation

Three failure modes silently corrupt agent-evaluation conclusions:

These boundaries should be machine-checkable, not vibes. That is what this project ships. How this relates to benchmark-rigor and transcript-audit work (verified literature anchors): see docs/positioning.md.

3. What Agent Comparability Does

• Adjudicates a Requested Claim (e.g. score_comparison, causal, descriptive_registration) against a set of evaluation records
• Returns a Comparison Decision Record (CDR): admissibility (ADMISSIBLE / INADMISSIBLE / PARTIAL / INCONCLUSIVE), satisfied/violated/unknown conditions, supported claims AND forbidden claims, evidence refs, policy version
• Enforces evidence-trust tiers (agent-claimed vs source-verifier vs independently verified), score-scale consistency (range/unit/verifier), zero-semantics splits, and unknown-field/unknown-claim rules
• Ships a frozen conformance suite run against three separated metrics

Admissibility quick reference (full priority rules in SPEC §3.3):

4. What It Does Not Do

5. Main Features

• adjudicate(records, requested_claim) -> CDR — pure Python, zero dependencies
• Priority-ordered decision table (evidence-missing > confounded > semantics > admissible)
• Machine-verifiable CDR JSON Schema (spec/cdr.schema.json)
• Frozen A/B conformance case sets; B5 public erratum precedent protected
• Optional adapters for Inspect AI and Pydantic Evals reusing the same core

6. Installation

From PyPI (first published release: 0.5.2):

pip install agent-comparability                            # minimal core, zero deps
pip install "agent-comparability[suite]"                   # + suite basics (pyyaml, jsonschema)
pip install "agent-comparability[suite,inspect,pydantic]"  # FULL conformance run

From a clone (development):

git clone https://github.com/Rethymus/agent-comparability.git && cd agent-comparability
pip install -e .                        # minimal core, zero deps
pip install -e .[suite]                 # + suite basics (pyyaml, jsonschema)
pip install -e .[suite,inspect,pydantic]  # FULL conformance run — required by run_suite.py

Requires Python ≥ 3.10. No Docker, no model keys. Note: the core package supports 3.10+; the optional inspect/pydantic extras pull dependencies that require Python ≥ 3.11 — on 3.10 pip resolves older versions of them.

7. Quickstart

Command line (offline, zero extra dependencies — input contract below):

python -m agent_comparability examples/demo-case.json --claim score_comparison
# → CDR JSON on stdout (exit 0). INADMISSIBLE / PARTIAL / INCONCLUSIVE are verdicts,
#   not program failures. Invalid input (bad JSON / missing or mistyped fields) →
#   exit 2 with a path-precise diagnostic — never silently converted into a verdict.
#   -o cdr.json writes to a file instead. The demo case is SYNTHETIC/illustrative.

Input contract ({"id": str, "records": [{"task_id", "harness", "model", "reward": number, "reward_status", ...optional}]}); unknown record fields are ignored and recorded in unknown_conditions (SPEC §5). Add --manifest out.manifest.json to also emit an Evidence Manifest documenting artifact digests and evidence trust states (never auto-upgraded; see docs/evidence-manifest.md).

Python API:

from agent_comparability import adjudicate

cdr = adjudicate({"id": "demo", "records": [
    {"task_id": "t1", "harness": "h1", "model": "m1",
     "reward": 1.0, "reward_status": "verified"},
    {"task_id": "t1", "harness": "h2", "model": "m1",
     "reward": 0.0, "reward_status": "unresolved_zero"}]},
    "score_comparison")

8. Input/Output Example (actual output)

Note: the reference implementation's condition/claim messages are currently in Chinese; the admissibility verdict itself is locale-neutral.

The quickstart call returns:

cdr.admissibility      -> PARTIAL
cdr.supported_claims   -> ['per_status_stratified_rate(verified_zero/unresolved_zero 分列)']

A confounded group instead:

cdr = adjudicate({"id": "x", "records": [
    {"task_id": "a", "harness": "h1", "model": "m1", "reward": 1.0, "reward_status": "verified"},
    {"task_id": "b", "harness": "h2", "model": "m2", "reward": 0.0, "reward_status": "verified"}]},
    "score_comparison")

cdr.admissibility       -> INADMISSIBLE
cdr.violated_conditions -> ['same_task_id', 'same_model', 'single_variable_isolation']

cdr.to_dict() emits the full CDR (schema-validated). The quickstart call's complete CDR — actual output:

{
  "case_id": "demo",
  "requested_claim": "score_comparison",
  "comparison_scope": "2 records; models=['m1']; harnesses=['h1', 'h2']",
  "admissibility": "PARTIAL",
  "satisfied_conditions": [
    "same_task_id",
    "same_model",
    "single_variable_isolation",
    "continuous_score_preserved({'range': '0..1', 'unit': 'fraction', 'verifier': 'source'})"
  ],
  "violated_conditions": [
    "zero_semantics_uniform(verified_zero != unresolved_zero)"
  ],
  "unknown_conditions": [],
  "supported_claims": [
    "per_status_stratified_rate(verified_zero/unresolved_zero 分列)"
  ],
  "unsupported_claims": [
    "pooled_fail_rate(0 分语义混同)"
  ],
  "evidence_refs": [
    "rec:t1/m1/h1",
    "rec:t1/m1/h2"
  ],
  "policy_version": "cmp/0.4.0",
  "limitations": [
    "摘要一致仅证明未同步修改，不构成抗恶意篡改（单写信任域）"
  ]
}

9. Architecture

spec/SPEC.md                     ← single source of truth (readable standalone)
spec/cdr.schema.json             ← machine-verifiable CDR
src/agent_comparability/core.py  ← thin reference implementation
src/agent_comparability/data/    ← SPEC + schema + frozen cases (packaged)
conformance/                     ← run_suite.py (3 metrics) + independent_impl.py + cases
adapters/                        ← optional inspect-ai / pydantic-evals wrappers
tests/                           ← smoke + property tests

Adjudication flow (priority order, per SPEC §3.3):

flowchart LR
    A[Records + Requested Claim] --> B{trace available?}
    B -- missing --> C[INCONCLUSIVE]
    B -- ok --> D{task / model /<br/>score scale consistent?}
    D -- conflict --> E[INADMISSIBLE]
    D -- ok --> F{zero semantics uniform?}
    F -- verified + unresolved mixed --> G[PARTIAL]
    F -- ok --> H[ADMISSIBLE]
    C & E & G & H --> V[CDR<br/>supported + forbidden claims]

10. Conformance Suite

python conformance/run_suite.py reports three separated metrics that must never be merged:

1. Policy correctness — reference implementation vs frozen expectations
2. Spec conformance — an independent implementation written from SPEC alone
3. Integration consistency — the same core via Inspect/Pydantic adapters

A failure in each means a different defect class (implementation bug / spec ambiguity or misread / adapter bug).

The full suite run requires the adapter extras (.[suite,inspect,pydantic]) — a missing adapter is treated as FAIL, not skipped.

11. Inspect AI / Pydantic Evals Integrations

pip install -e .[inspect] && python adapters/inspect_adapter.py
pip install -e .[pydantic] && python adapters/pydantic_adapter.py

Both wrap the same core (no semantic duplication). Inspect scorer uses metrics=[] — admissibility is deliberately not aggregated into accuracy. The Pydantic evaluator checks the task's real output (ctx.output) against frozen expectations — it never re-adjudicates inside the evaluator — and the adapter's exit code derives from the full 22-pair report (cases + failures). For real Inspect EvalLog files (not frozen cases) use the converter: python -m agent_comparability.inspect_log run.eval --claim … --status … (official APIs, explicit mapping policy, no silent guessing — see docs/inspect-log.md).

12. Specification & Versioning

Three distinct lines: software tag (e.g. v0.5.0), specification cmp/0.4.0 (inside SPEC), and agent_comparability.__version__. Judgment-table changes require a spec major bump (SPEC §5). See CHANGELOG for the full tag lineage (v0.3.1–v0.3.6 are deprecated candidates kept immutably for the record).

13. Limitations

• Experimental specification; self-tested conformance, no external certification
• Single-writer trust domain: digest equality proves non-divergence, not tamper-resistance (SPEC §1)
• No repetition/cost modeling inside adjudication (handled by callers)
• Chinese/English docs kept equivalent; translation drift tracked in CI link checks

FAQ

Does passing the conformance suite certify my agent is reliable? No. It certifies that an implementation matches the specification. It says nothing about any agent.

Can I use it with my existing eval framework? Yes. The core is framework-independent (adjudicate() is a pure function over record dicts). Adapters for Inspect AI and Pydantic Evals are included; any framework whose output you can map to the record schema works.

Why not just compare scores? Because 0.0 can mean a verified failure or an unresolved zero; because model and harness changing together turns a diff into an unfounded causal claim; and because missing traces inflate success rates silently. §2's failure modes are all real-world artifacts, not hypotheticals.

Does it run agents or benchmarks? No — no network, no credentials, no Docker. It only ever sees the records you pass in.

Which versions should I depend on? Package v0.5.0+ (version lines unified and CI-enforced since then); specification cmp/0.4.0 evolves independently — see §12.

Project status

14. Contributing

See CONTRIBUTING.md. Key rule: never edit frozen expectations to make tests pass; errata follow the public B5 precedent. PRs must pass all three conformance metrics separately.

15. Security

See SECURITY.md. Private vulnerability reporting preferred.

16. License & Acknowledgements

Apache-2.0 (see LICENSE, NOTICE). Optional integrations inspect-ai / pydantic-evals are MIT and not bundled. Conformance case data is original to this project; no CC BY-NC data is distributed. Research lineage: this repo was distilled from a private research repository; internal process records live under docs/internal/.