output-validator

Catches a broken output before a human — or a teacher's Office/PDF app — does. Stdlib-first: the core checks run with zero optional libraries; richer checks (JSON Schema, deep format conformance) activate when their library is present and are reported as skipped, never failed, when not. This is a safety net, not a nicety: it is how the ecosystem catches its own mistakes.

What it validates

• Governed JSON artifacts — tools/validate_outputs.py: required keys, human_review_required: true, no empty/fabricated standards, no real PII (SSN / non-555 phone), and JSON Schema (records / traversal / udom / connector-flags) when jsonschema is installed. --promote captures a failing case as a regression eval so the same defect can't recur silently.
• Produced documents — tools/validate_document.py: structural integrity of .docx/.pptx/.xlsx (OPC parts), .odt/.ods/.odp (ODF), and .pdf (header / xref / %%EOF) — the corruption classes documented by the Open XML SDK and veraPDF, with zero dependencies.

How it works — the pipeline

Follow references/method.md. Generation here is the validation pass: Analysis (pick schema/profile) → Checks (governance rules + structure) → Diagnosis → pass/fail + repair guidance. Honest limits: structural validity is necessary, not sufficient — deep OOXML schema (Open XML SDK) and PDF/A or PDF/UA conformance (veraPDF) are optional follow-ups, reported as such and never assumed (references/format-error-catalog.md). When a check can't run (a library is absent) it degrades to a labeled gap, so the validator itself always returns a usable answer.

References (thorough by design)

• references/rule-catalog.md — every governance + claim rule, with severity and rationale.
• references/format-error-catalog.md — OOXML / PDF / ODF / font error classes + spec citations + fixes.
• references/profiles.md — per-artifact required-key profiles.

Output: always emit the metadata block

Every validation report ends with the metadata block (references/metadata-schema.md) + human_review_required: true. On a failure, route the fix to skill-repair (mechanical) or the owning skill (judgment). Placeholders only; never real student data.

python3 tools/validate_outputs.py --input artifact.json --schema records
python3 tools/validate_document.py deck.pptx report.pdf