Data Provenance

Synthetic-only disclosure — no real insurer data

Provenance

Training and evaluation data are either public corpora (distributional / stylistic characteristics only) or synthetically generated from randomized claim skeletons — never real insurer files.

No real insurance company data

Training and evaluation data are either public corpora (distributional / stylistic characteristics only) or synthetically generated from randomized claim skeletons.

Usage · Changelog · Sample corpus

Public sources (prime examples)

Source	What we take	What we do not take
FUNSD (form understanding)	Field label lexicon, key-value layout patterns, OCR noise shape	Raw forms as training labels for our taxonomy
DocLayNet (layout)	Layout class frequencies; legal/regulatory prose texture from `laws_and_regulations` pages	DocLayNet category labels as classifier targets
RVL-CDIP (document images)	Surface characteristics of form/letter/memo/invoice/email classes for style conditioning	RVL-CDIP class labels as our taxonomy
Public insurance claim tables	Histograms for loss amounts, loss types, state mix (shape only)	Individual real claim rows as documents
Legal writing samples (e.g. public opinions / pile-of-law style text)	Vocabulary, discourse markers, IRAC-ish reasoning templates	Legal document types as classification labels

Legal text is injected only into narrative sections of insurance documents (claims correspondence, supporting evidence, memo reasoning). The classifier label set remains insurance-only.

Synthetic generation

Skeleton sampling — randomized ClaimSkeleton objects validated against data/schemas/claim_skeleton.schema.json, seeded by data/profiles/insurance_distributions.json
Stage A — document text (LLM via OpenRouter when configured, else deterministic templates conditioned on profiles)
Stage B — adjuster memo text from skeleton + Stage A
Noise injection — OCR garble using data/profiles/ocr_noise_profile.json

Every record is appended to data/provenance_log.jsonl with stage, source, model (if any), and prompt/profile version.

DICIE application fixtures

The paper-aligned DICIE path (src/docie/) uses the same synthetic-only rule. Committed examples in tests/fixtures/sample_docie_documents.jsonl are hand-written fictional HCFA / UB-04 / LOG / sales texts for CI — not real claimant documents. Application label sets live in taxonomy/medical_bills.yaml and taxonomy/salvage_claims.yaml. Batch DICIE runs append provenance rows with stage=docie_pipeline.

Sample medical + salvage document corpus

The queryable corpus store (src/storage/, see sample_document_corpus.md) houses larger sets of fictional medical bills and salvage documentation (Letters of Guarantee, salvage sales receipts, towing/storage attachments) patterned after AmFam-style intake surfaces. No proprietary insurer files are ingested. Seed/import events log stage=sample_corpus_seed / sample_corpus_import.

Skeleton schemas:

data/schemas/medical_bill_skeleton.schema.json
data/schemas/salvage_document_skeleton.schema.json

RVL-CDIP public index

The queryable RVL-CDIP SQL store (src/rvl_cdip/, see rvl_cdip_sql.md) indexes the public aharley/rvl_cdip label lists. Hub downloads and the SQLite DB are confined to .venv/rvl_cdip/ (covered by the existing .venv/ gitignore). Build events log stage=rvl_cdip_build. The ~38 GB image archive is never fetched unless explicitly opted in.

Committed vs gitignored

Committed: schemas, taxonomy (ACORD + medical bills + salvage claims), characteristic profiles (data/profiles/*.json), tiny test fixtures (including DICIE samples), sample-corpus seed exports under data/sample_corpus/seeds/
Gitignored: bulk data/raw/*, synthetic JSONL outputs, provenance log, trained model weights, pipeline/DICIE caches under data/pipeline/, sample corpus SQLite DB + regenerable exports under data/sample_corpus/, RVL-CDIP artifacts under .venv/rvl_cdip/

Reproducibility

Characteristic profiles are versioned JSON committed to the repo so synthetic generation can run without re-downloading multi-GB corpora. Corpus ingest scripts remain available to refresh profiles from Hub samples when needed.

--- title: "Data Provenance" subtitle: "Synthetic-only disclosure — no real insurer data" --- ::: {.hero-kicker} Provenance ::: ::: {.hero-lead} Training and evaluation data are either **public corpora** (distributional / stylistic characteristics only) or **synthetically generated** from randomized claim skeletons — never real insurer files. ::: ::: {.callout-important} ## No real insurance company data Training and evaluation data are either **public corpora** (distributional / stylistic characteristics only) or **synthetically generated** from randomized claim skeletons. ::: [Usage](usage.md) · [Changelog](CHANGELOG.md) · [Sample corpus](sample_document_corpus.md) ## Public sources (prime examples) | Source | What we take | What we do **not** take | |--------|--------------|-------------------------| | FUNSD (form understanding) | Field label lexicon, key-value layout patterns, OCR noise shape | Raw forms as training labels for our taxonomy | | DocLayNet (layout) | Layout class frequencies; legal/regulatory prose texture from `laws_and_regulations` pages | DocLayNet category labels as classifier targets | | RVL-CDIP (document images) | Surface characteristics of form/letter/memo/invoice/email classes for style conditioning | RVL-CDIP class labels as our taxonomy | | Public insurance claim tables | Histograms for loss amounts, loss types, state mix (shape only) | Individual real claim rows as documents | | Legal writing samples (e.g. public opinions / pile-of-law style text) | Vocabulary, discourse markers, IRAC-ish reasoning templates | Legal document types as classification labels | Legal text is injected only into narrative sections of insurance documents (claims correspondence, supporting evidence, memo reasoning). The classifier label set remains insurance-only. ## Synthetic generation 1. **Skeleton sampling** — randomized `ClaimSkeleton` objects validated against `data/schemas/claim_skeleton.schema.json`, seeded by `data/profiles/insurance_distributions.json` 2. **Stage A** — document text (LLM via OpenRouter when configured, else deterministic templates conditioned on profiles) 3. **Stage B** — adjuster memo text from skeleton + Stage A 4. **Noise injection** — OCR garble using `data/profiles/ocr_noise_profile.json` Every record is appended to `data/provenance_log.jsonl` with stage, source, model (if any), and prompt/profile version. ## DICIE application fixtures The paper-aligned DICIE path (`src/docie/`) uses the same synthetic-only rule. Committed examples in `tests/fixtures/sample_docie_documents.jsonl` are hand-written fictional HCFA / UB-04 / LOG / sales texts for CI — not real claimant documents. Application label sets live in `taxonomy/medical_bills.yaml` and `taxonomy/salvage_claims.yaml`. Batch DICIE runs append provenance rows with `stage=docie_pipeline`. ## Sample medical + salvage document corpus The queryable corpus store (`src/storage/`, see [sample_document_corpus.md](sample_document_corpus.md)) houses larger sets of fictional medical bills and salvage documentation (Letters of Guarantee, salvage sales receipts, towing/storage attachments) patterned after AmFam-style intake surfaces. No proprietary insurer files are ingested. Seed/import events log `stage=sample_corpus_seed` / `sample_corpus_import`. Skeleton schemas: - `data/schemas/medical_bill_skeleton.schema.json` - `data/schemas/salvage_document_skeleton.schema.json` ## RVL-CDIP public index The queryable RVL-CDIP SQL store (`src/rvl_cdip/`, see [rvl_cdip_sql.md](rvl_cdip_sql.md)) indexes the public [aharley/rvl_cdip](https://huggingface.co/datasets/aharley/rvl_cdip) label lists. Hub downloads and the SQLite DB are confined to `.venv/rvl_cdip/` (covered by the existing `.venv/` gitignore). Build events log `stage=rvl_cdip_build`. The ~38 GB image archive is never fetched unless explicitly opted in. ## Committed vs gitignored - **Committed:** schemas, taxonomy (ACORD + medical bills + salvage claims), characteristic profiles (`data/profiles/*.json`), tiny test fixtures (including DICIE samples), sample-corpus seed exports under `data/sample_corpus/seeds/` - **Gitignored:** bulk `data/raw/*`, synthetic JSONL outputs, provenance log, trained model weights, pipeline/DICIE caches under `data/pipeline/`, sample corpus SQLite DB + regenerable exports under `data/sample_corpus/`, RVL-CDIP artifacts under `.venv/rvl_cdip/` ## Reproducibility Characteristic profiles are versioned JSON committed to the repo so synthetic generation can run without re-downloading multi-GB corpora. Corpus ingest scripts remain available to refresh profiles from Hub samples when needed. ::: {.see-also} ### See also [Sample Document Corpus](sample_document_corpus.md) · [RVL-CDIP SQL Index](rvl_cdip_sql.md) · [Architecture](architecture.qmd) · [About](about.qmd) · [Bugfix audit](bugfix_audit_round2.md) :::