Reference Intermediate 1 minute read Updated 2026-06-26 UTC

Manifest schemas

Lightweight schemas for reports, skill packages, evaluations, and releases.

Research statusReference schema guide Publication statePublished Reviewed byMichael Kappel Source reports2

Why schemas are small

This site avoids a database. That makes manifest discipline more important. Each file-backed registry should use small, explicit schemas that can be validated in build scripts and read by humans.

Report manifest

pseudocode
report_record = {
    slug: "ai-evolution-small-models-big-ecology",
    filename: "AI Evolution_ Small Models, Big Ecology.md",
    title: "AI Evolution: Small Models, Big Ecology",
    category: "Foundations",
    summary: "...",
    maturity: "Conceptual synthesis",
    size: 82386,
    sha256: "..."
}

Evaluation card

pseudocode
evaluation_card = {
    id: "eval.skill.math.2026-06-26",
    candidate_id: "skill.math.reasoner.q4",
    suite_id: "suite.math.v3",
    created_at_utc: "2026-06-26T00:00:00Z",
    hard_invariants_passed: true,
    metrics: {
        accuracy: 0.84,
        calibration_error: 0.06,
        p95_latency_ms: 880,
        peak_ram_bytes: 760000000
    },
    evaluator_signature: "..."
}

Release record

pseudocode
release_record = {
    alias: "math-specialist-champion",
    from_version: "skill.math.reasoner.q4.1.1.0",
    to_version: "skill.math.reasoner.q4.1.2.0",
    stage: "canary",
    traffic_percent: 5,
    rollback_target: "skill.math.reasoner.q4.1.1.0",
    approved_by: "owner",
    approved_at_utc: "2026-06-26T00:00:00Z"
}

Validation rule

A manifest schema is useful only when the build or release process refuses malformed records. Treat schema validation as a gate, not documentation.

Source reports used for this guide

These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.