M4 one-path benchmark: a deliberately small result

Result at a glance

The evaluated build retained the one expected dependency candidate, classified the one expected retained edge, and reached the one expected impact target through a provenance-valid path. No target in this fixture was incorrectly converted into an unaffected result.

The rates below are written as counts before percentages because a perfect value on one case can otherwise look much stronger than it is.

• Critical candidate recall: 1/1 (100%)
• Retained-edge precision: 1/1 (100%)
• Impacted-target recall: 1/1 (100%)
• Valid path rate: 1/1 (100%)
• False-unaffected count: 0
• Release-gate result for this fixture: PASS

Dataset and ground truth

The dataset is the M4 one-path synthetic development fixture embedded in test_synthetic_graph_and_impact_results_feed_the_release_gate. It contains one material source change, one expected dependency from a claim node to one impact-object node, and one expected impacted target.

The expected dependency and target were authored by the developer as part of the executable test. They were not independently adjudicated, blinded, or held out from implementation. That makes this useful as a transparent regression and pipeline check, but weak evidence for product accuracy.

• Classification: SYNTHETIC_NON_PRODUCTION
• Cases: 1
• Gold dependency candidates: 1
• Gold retained edges: 1
• Gold impacted targets: 1
• Adjudication: developer-authored, not independent

Method

The benchmark constructs the declared graph nodes, exact synthetic source evidence, and one expected candidate key. Candidate discovery is run twenty times to verify deterministic selection. The candidate is classified through the deterministic edge provider under the recorded local policy.

The benchmark then freezes complete coverage for the two declared nodes, traverses the retained edge from one material change seed to one target, calculates the graph and impact metrics, and applies the same fail-closed release-gate function used by the M4 code.

• Network calls: none
• Provider: DeterministicEdgeProvider
• Discovery repetitions: 20
• Coverage: complete for the declared two-node policy scope
• System version: EvidView 0.1.0
• Source revision: 22dc5907f9164c0bf6c015a51b7077be83a4723e
• Evaluation date: 17 September 2026

Reproduce it

From the repository root with the locked Python environment available, run the command below. A passing run executes the named fixture and its release-gate assertions; a code, policy, dependency, or fixture change that alters the behavior must update the benchmark record rather than silently retaining this result.

• uv run pytest tests/unit/test_m4_release_gates.py::test_synthetic_graph_and_impact_results_feed_the_release_gate -q
• Machine-readable record: website/content/benchmarks/m4-one-path-0.1.0.json
• Locked dependency digest: 562951538c46a20be7c1923fd46f64f73363c102d82adee6e7969b782028b558

What this result does establish

For this exact build, fixture, policy, and deterministic provider, the implemented functions can carry one expected synthetic dependency through discovery, classification, frozen graph coverage, and impact traversal while retaining a valid evidence path. The executable gate also rejects the declared false-unaffected and unresolved-risk failure dimensions in its companion tests.

Publishing that narrow result is useful because it makes the denominator, system identity, method, and limitations inspectable. It is stronger than an unqualified demonstration and much weaker than a product-accuracy benchmark.

What this result does not establish

A 1/1 result does not estimate precision or recall on a population. It says nothing about unseen source language, larger graphs, ambiguous evidence, probabilistic provider behavior, human review burden, production infrastructure, or customer workflows.

The planned MedQMS-300 corpus requires 300 logical sources, protected qualification and holdout labels, independent adjudication, multiple change campaigns, hard negatives, and a larger metric set. That release does not yet exist. No comparison to it is implied.

• No customer or design-partner evidence
• No production or pilot evidence
• No measured time, cost, or error reduction
• No compliance, certification, safety, or completeness conclusion
• No general accuracy claim

Next evidence threshold

The next benchmark should not merely repeat this case with cosmetic variations. It should use a versioned corpus release with independently adjudicated labels, protected holdout membership, per-class denominators, disagreement records, false-negative analysis, and separate reporting for deterministic and probabilistic stages.

Until that evidence exists, CLM-012 remains suspended and the website must not convert this development check into a broad performance promise.