Reproducible scoring evidence · MIT + CC BY 4.0

Validate the ranking before routing leads

A plausible explanation does not make a score trustworthy. This fixture rejects unsupported criteria, future data, bad totals, scored missing values, and duplicate records before it ranks anything. The surviving records enter a queue sized by review capacity rather than an invented universal threshold.

Every lead, signal, score, and outcome is synthetic. The run tests the evaluator, not a language model or a production conversion rate. It contains no names, email addresses, social profiles, or CRM records.
Sixteen synthetic score records pass through evidence and leakage gates; ten enter evaluation and four enter the capacity-limited review queue

Files

lead_score_validator.ts Strict score validation and holdout metrics fixture.ts Sixteen synthetic score candidates and outcomes lead_score_validator_test.ts Twenty-one validation and metric tests run_fixture.ts Deterministic manifest runner lead-scoring-validation-manifest-2026-08-28.json Reviewed fixture result render_manifest.ts Dependency-free SVG renderer lead-scoring-validation-2026-08-28.svg Diagram generated from the manifest evidence_artifact_test.ts Saved manifest, SVG, and license checks evidence-run-2026-08-28.md Commands, results, hashes, and boundaries README.md File map and independent rerun commands

Evidence or zero

Each observed criterion cites a supplied signal. A missing criterion gets zero points instead of a convenient midpoint.

No future leakage

Evidence must predate scoring, and the outcome must follow it. Otherwise the record cannot enter holdout metrics.

Capacity, not folklore

The example selects the top four because that is the declared review limit. It does not present 75 or any other cutoff as universal.

Verified scope

Twenty-five tests cover score ranges, component totals, evidence ownership, timestamps, leakage, missing criteria, duplicate candidates, stable ordering, holdout metrics, saved JSON, generated SVG, and license files. In the reviewed fixture, ten of sixteen candidates entered evaluation and four entered the capacity-limited queue.

MIT license CC BY 4.0 license Read the lead scoring guide