Skip to content

Validation and audit

Commands that prove a Designed Database tree is internally consistent before you hand it to training, storage, or review teams.

audit structured

Implementation — builds a structured audit report, then walks each doc_id directory.

When {doc_id}.content.txt exists — the tool splits segments on ====, runs SchemaValidator, LaTeXValidator, EOS checks, root tag presence, counts of poros_section_* open/close pairs, duplicate singleton sections, tag-density heuristics (for example TagDensity ≤ 25%), and related rules.

{doc_id}.content.json — requires doc_id, content, pure_text_stream; the plain stream must not contain poros_* substrings.

{doc_id}.structure.json — requires doc_id, title, sections with non-empty sections, non-empty section titles, no legacy section_type == "section", chemical-formula noise checks, and formulas as a list without suspicious truncation patterns.

Multimodal slice — index file presence, required fields per entry, files under images/, asset_copied all true, and similar disk-level checks.

Outputs — human-readable console summary plus audit_report.json at the audit --root_dir. Summary counts include total_docs, passed, failed, warnings, docs_passed, docs_failed; each document entry carries category, passed, message, and details.

Exit code — 0 if summary.failed == 0, otherwise 1.

validate structured

Lightweight JSON checks in _validate_structured_file:

  • Top-level keys exist.
  • content and pure_text_stream are arrays of strings (not a single concatenated string).
  • Joined content ends with </s> (or the configured EOS), includes <poros_doc>, and must not contain <br>.
  • pure_text_stream is non-empty after join.

Gap note — some docstrings claim <br> is stripped during aggregation, but a repository-wide search may show no replacement step. If upstream HTML still carries <br>, this validator will fail until the input or a preprocessing step removes it.

validate multimodal

Ensures each {doc_id}.assets.index.json exists and that every element includes fig_id, caption, and mentions (list). Logs caption and mention coverage statistics; stricter than a quick smoke test but lighter than the full structured audit path for assets.

validate acceptance

Randomised sampling: up to five documents and up to three index entries per document. Intended as a weak final gate—failures are rare unless the corpus has no valid multimodal rows.

validate delivery

Calls the same report builder as audit structured, then writes delivery_acceptance_report.json under the log root. Use it when you want a named delivery artefact separate from audit_report.json.