skill-score

v0.1 DRAFT: experimental, vendor-neutral scoring protocol and TypeScript reference implementation for .skill artifacts.

skill-score estimates quality and completeness without pretending there is one universal measure of worth. It reports a normalized 0–100 score alongside confidence, coverage, dimension sub-scores, evidence receipts, gate caps, and explanations.

What it does

• Separates observed and independently verified evidence from self-reported claims.
• Uses release and continuity profiles with transparent, versioned weights.
• Leaves unknown quality neutral while reducing confidence.
• Applies hard caps for invalid artifacts, exposed secrets, and dangerous behavior.
• Treats token/compute use as possible efficiency or provenance data—never as “more is better.”
• Produces deterministic JSON suitable for policy engines and registries.

This project does not certify safety, prove authorship, perform cryptographic verification, or replace human judgment. See the draft specification.

Install and use

npm install @skillerr/skill-score

Library usage:

import { scoreSkill } from "@skillerr/skill-score";

const result = scoreSkill(assessment, "release");

Or from a checkout of this repo:

npm install
npm run build
node dist/src/cli.js test/fixtures/cases.json release

Assessment inputs follow schema/assessment.schema.json. The dot-skill/skillerr reference CLI's skill score <file.skill> builds this input automatically from a package's provenance/benchmark.json (see docs/EVAL.md) — you don't need to hand-write the assessment JSON for a .skill package.

Formula at a glance

Known receipt values are averaged using evidence multipliers: observed 1.0, verified-external 0.9, and self-reported 0.25. Dimension results are combined using profile weights. Unknown-only dimensions estimate quality at 50 with zero confidence. Overall confidence and coverage are separately weighted. The lowest applicable gate cap is then applied.

Development

npm install
npm run check

Fixtures cover excellent, incomplete, self-claimed/forged-looking, and privacy-risk assessments. Contributions are welcome under the governance and contribution policies in this repository.

License

MIT © Bharat Dudeja.