Upgrade to H.A.R.D. 0.2-preview.3

Human-centered AI Readiness and Decision Protocol. Public Preview.

Exact versions: protocol 0.2-preview.3, MCP 0.2.0rc9, Skill/contract 0.2.0-rc.9.

This release makes consequential choices a reviewable unit: purpose and criteria, observed choice, documented historical or newly proposed alternatives, rationale provenance, impacts, evidence, accountable human disposition and follow-up. A sound current choice may stay. A new explanation is never proof of a creator's historical reasoning.

Pick the appropriate purpose

Artifact Review is a formal stage-bounded route for one person or a small team. It accepts a specification, prototype or code artifact. Record specified, walkthrough, implemented and runtime-tested evidence independently. A document can support a specification handoff without code; that outcome says nothing about runtime release approval. QUICK-6 and Full Profile retain their separate six-gate engineering-commitment purpose. Independent evaluation adds the conditions needed for reviewer-performance measures.

Records require explicit migration

Preserve old records, filenames, version identifiers and results. Do not replace only a version string and reuse a past pass. Start a new review from an identified artifact and criterion revision, declare artifact_population and mode, carry forward only unchanged source evidence, and recheck affected choices and validations. New consequential choice context is included in engineering validation fingerprints. A reference update is not a test rerun.

The artifact populations are ai_generated, runtime_ai, both, neither and unknown. AI-written code does not imply runtime AI. Missing population is invalid. Unknown population prevents a qualifying engineering result or pooled diagnostic calculation. Mixed populations, criteria, evaluator populations, stages or modes are not pooled.

Root Templates and Worked-Example now use schemas/csv-contract.json. Requirements and recovery use the same four-stage columns; findings retain both locked_at and adjudication fields. Source/verify_example.py checks this current contract. Archived root PDFs and spreadsheets remain legacy materials and are not included as current scorecards in the new compact distribution. The authoritative current forms are the Markdown and CSV files. Archived publications remain available under their original release paths.

Research measures and scope

Artifact coverage survives disclosed role overlap. Independent defect recall, expected-recall gap, omission recognition and false-ready acceptance require explicit eligibility. Unknown or violated conditions yield null and reason codes; null is never zero. The expected-recall gap remains expected reference-defect recall minus observed recall. Retrospective criteria can support artifact inspection but cannot be presented as a prospective frozen experiment.

Legacy MCP diagnostics retain their formulas with new population and eligibility safeguards. They identify current software and cannot establish a stage handoff from aggregate counts. Historical software and DOI records remain unchanged.

An omitted requirement is not automatically Critical. Acceptance, severity, scope and resolution are separate fields. Critical severity needs an applicable criterion or documented novel risk, an evidenced failure mechanism and consequence. A future-phase feature is not a current defect when it is explicitly outside scope and no current mandatory behavior depends on it. Unclear scope remains unresolved. No exclusion may silently shrink a frozen denominator.

Feedback and limits

The author supplied practitioner feedback describing a self-directed, AI-assisted trial on an unimplemented feature plan. It reported inconsistent CSV schemas, absent population stratification, ambiguous overlapping roles, missing recovery specifications and unclear finding acceptance. This informed usability and implementation corrections. The source attachments and line-referenced findings were not available for independent adjudication in this upgrade. Reported percentages, severity and efficacy are not treated as verified results. No new empirical validation, participant observation, endorsement or adoption claim is created.

The optional model companion has been removed from current setup, code and distribution. No model provider or API key is required by the deterministic MCP or portable Skill engine. The host assistant may receive user-supplied content under its own operation.

Software regression and package parity check implementation only. This remains a Public Preview. The archived DOI 10.5281/zenodo.22667623 identifies rc.3 alone.