Report JSON
schema_version is 27. It moves when this shape moves,
which is the field to branch on when reading a report programmatically.
The same shape reaches you from report.to_dict() in Python and from
complydoc audit --print-json on the command line.
{ "schema_version": 27, "top_level_keys": [ "run", "documents", "skipped", "cost", "aggregate", "overall", "quick_wins", "loader", "loader_comparison", "routing", "verification", "ignores", "concepts", "categories", "evals", "chunks", "trace", "limitations", "staleness_warnings", "signal_weights", "config_masking" ], "run": { "components_run": "list of cost | readiness | sensitive; empty for a chunks run", "offline_guard": "armed | not_armed", "content_sent_to": "hosts sent document text or page images, empty unless a hosted classifier or a vision model ran", "verify_model": "the vision reader pages were checked against, null unless --verify", "verify_scope": "flagged | all, null unless --verify", "classifier_missed_workers": "documents a registered classifier could not reach", "classifier_calls": "calls a registered classifier made, 0 unless one ran", "classifier_failures": "of those, calls that raised and so produced no score", "reveal_used": "bool \u2014 true means values are NOT masked", "page_images_used": "bool", "extracted_text_used": "bool", "report_detail": "summary | full \u2014 summary leaves out documents[].cost.models, cost.documents and, unless --page-images, documents[].previews; full writes every field", "jobs": "worker processes used", "timeout_seconds": "seconds each document was given, when --timeout was used", "documents_read_after_worker_failure": "documents re-read after a worker stopped", "config_digest": "identifies the config that produced these numbers" }, "documents[]": { "relative_path": "str", "sha256": "str", "format": "pdf | image | docx | xlsx | pptx | html | markdown | text | email | other", "cost": "pages, page_count, text layer and coverage for this document", "cost.models[]": "full only: every priced model against this document, with text and vision token counts and USD", "previews[]": "page geometry and pictures the viewer draws pages from; full, or summary with --page-images. A picture has the identifiers found on the page blacked out; image_withheld counts those that could not be placed, for which the picture was left out", "readiness.signals[]": "id, value, rating, weight, why, status", "readiness.score": "value 0-100, higher is better; label; low_confidence", "sensitive.matches[]": "category, page, line, column, masked, severity, fingerprint, and evidence: confirmed | corroborated | pattern | model", "sensitive.unreadable_pages": "pages that were not searched at all", "sensitive.models_used": "category -> the detector and model that answered for it, where a model rather than a pattern read it", "extractions[]": "one per reader asked for; the first is the one kept", "extracted_text[].costs": "reader -> usd, basis (local | actual | estimated | unpriced), model, input_tokens, output_tokens: what each reading of the page cost", "extracted_text[].seconds": "reader -> seconds it took on this page, where timed", "extracted_text[].tokens": "reader -> tokenizer -> text tokens in that reading of the page", "extracted_text[].image_tokens": "vision formula -> image tokens for the page", "extracted_text[].masked_text": "with --reveal only: the page's text with every identifier covered, beside `text`, which holds the values; also masked_ocr_text and masked_readings", "extracted_text[].vision_estimate": "what the cheapest priced vision model would cost for this page, estimated", "verification": "null unless --verify: model, scope, pages[] (number, status: agrees | disagrees | filled | failed | not_rendered, why, similarity, coverage, missing, cost), unreadable_pages[], sent_to[]", "metadata_findings[]": "key, category, severity, evidence, masked (from metadata)", "content_findings[]": "page, visibility (visible | not_measured | suspected | confirmed), instruction (confirmed | pattern | model | none), severity, excerpt, hidden_reasons[], instruction_reasons[], score, in_loader_output, fingerprint", "concept_findings[]": "with --judge-concepts only: page, concept, label, severity, score \u2014 a page a judgement model said holds one of your concepts", "ignored[]": "findings an ignore file set aside, left out of every count and rule: fingerprint, kind (identifier | content), reason, by, until, and the finding itself as identifier or content", "visibility_checked": "bool, null when the scan did not run", "path_exposures": "metadata keys holding an absolute filesystem path" }, "cost": "null unless cost ran: currency, headline_resolution, resolutions[], volume, prices_as_of (the date of the oldest price used), models[] (the folder's cost per model and per path: text_layer, text_ocr, vision, each with folder_usd, per_1000_usd and annual_usd; and the model's input_per_mtok_usd, supports_vision, vision_formula and tokenizer, to price one page), and, full only, documents[] (every document against every model)", "overall": { "score": "global readiness 0-100, content and cost path and exposure", "factors[]": "name, score (null when not measured), weight, why", "bands": "documents per band \u2014 the composition the mean hides" }, "quick_wins[]": "id, title, detail, documents[], actor (complydoc | you), effect", "loader": "null unless documents came from an external loader: name, documents_returned, seconds, network_allowed, network_attempts[], error, tags[], metadata_keys[]", "loader_comparison": "null unless compare_loaders ran: baseline, loaders[] (per-loader totals, network, scores, failures, facts_found, parser_usd, tags[]), facts[] (found, scores, and nearest: the passage that came closest where a fact was missed), identifier_differences[] (found_by[], missed_by[]), metadata_keys (key -> loaders returning it), documents (path -> loaders), recommended (the loader to use, null where the run cannot tell), verdict (what decided it), ranked[], formats[] (per file type: label, documents, loaders[] with documents, failures, seconds, similarity and facts_found, skipped_by[], recommended, verdict), baselines (document -> the loader it was measured against, where not the first)", "chunks": "null unless complydoc chunks ran, which writes no documents: one entry per splitter with chunker, stats (count, tokens_total, tokens_min, tokens_median, tokens_p95, tokens_max), flag_counts, repeated_identifiers, facts[] (fact, status: whole | split | missing, chunks[]), retrieval[] (question, fact, status, rank, answer_chunks[], top_chunks[]), top_k, and chunks[] (index, document, page, tokens, identifiers[], hidden, flags[], preview, all masked, and start, end, page_characters: where it sits in its page's text)", "trace": "null unless a pipeline ran inside cd.observe: name, scan (patterns | full | off), seconds, overhead_seconds, libraries, error, connections_outside[], and stages[] in the order they started (index, kind: load | transform | split | embed | custom, component, module, method, tags[], parameters, documents_in, documents_out, characters_in, characters_out, sources[], scanned, identifiers[] (fingerprint, label, masked, severity, occurrences, evidence), hidden, metadata_keys[], metadata_keys_added[], path_keys[], connections[], hosts[], vectors, dimensions, chunks (index into chunks), parent, started, tokens_in, tokens_out, usd, usd_basis, previews[] (source, page, characters, tokens, text, metadata, all masked), finished, error, seconds)", "aggregate": "folder totals: cost, signal_distribution, sensitive_by_category, content_matrix (instruction -> visibility -> passages)", "verification": "null unless --verify: model, scope, min_coverage, pages_total, pages_checked, pages_agree, pages_disagree, pages_filled, pages_failed, pages_unreadable, usd, usd_basis, headline", "concepts": "null unless a concepts file was read: file, concepts[] (id, label, description, pattern, severity, judge, found); a concept with a pattern is found as the identifier category concept_<id>", "categories": "null unless a categories file was read: file, changes[] (category, label, enabled, severity, shipped_enabled, shipped_severity), unknown[] (names the file gave that are no category). A category with enabled false is not reported; its values are still masked", "evals": "null unless an evals file was read: file, readers[] (each loader, or the audit's own reader), checks[] (fact, document, found, scores, pages, documents, nearest, as loader_comparison.facts), each keyed by reader", "ignores": "null unless an ignore file was read: file, rules[] (finding, reason, by, until, paths[], what, matched, expired)", "limitations[]": "area, statement, affected[], severity (info | important)"}JSON Schema
Section titled “JSON Schema”That is a summary. The whole shape, every key with its type, is a JSON Schema generated from the code that writes the report:
- This version:
report-27.json - The newest, whichever it is:
report.json - From the tool you have installed:
complydoc schema --json-schema
Each schema version keeps its own address, so a report can be checked against the schema of the version that wrote it.