Inside Auditario’s 3PL billing document corpus

What the corpus proves

The point is not to advertise a large document count. It is to state exactly which evidence belongs to each set, make excluded structures explicit, and prevent a changed upstream page from silently replacing a frozen research source.

Seven committed public source documents

The raw engineering corpus contains seven public documents fetched once on 12 August 2026 and committed beside an immutable manifest. This register names each source instead of asking readers to trust an aggregate count.

Download the public JSON manifest

The downloadable manifest contains the full SHA-256 digest, original URL, byte count, document type, and observed contribution for every entry. Tests operate without fetching the network. That makes a changed upstream page unable to rewrite a historical result silently.

Four separately documented invoice-layout variants with synthetic values

The seven public contracts and rate cards above do not provide real merchant invoices. Separately, the repository reproduces four documented export shapes and fills them with internally consistent fictional values. Their layout provenance is not the seven-file raw-source manifest:

The layouts come from verified documentation; the merchant names, invoice dates, references, quantities, and amounts are fictional. Each fixture has a hand-labeled ground truth for line grain, totals, and expected normalized fields. Once a fixture enters a golden test, corrections create a new version rather than overwriting the old file.

From public source to regression

1. Verify and freeze the source. Confirm the URL and claimed content, then record the bytes and digest.
2. Catalog the observed structure. Identify document species, rate mechanics, billing units, layout grain, and known gaps.
3. Derive a fixture or rule with provenance. Preserve real structure; keep synthetic values visibly labeled.
4. Attach ground truth. Record expected fields, citations, totals, and exclusions in machine-testable form.
5. Test both success and refusal. The harness must parse supported structures and refuse or downgrade unsafe ones.
6. Freeze regressions. A discovered false positive becomes a permanent test before the corresponding cell can be considered for certification.

This evidence-gated approach also constrains the rate-rule language. A construct unsupported by the source corpus is cut or routed to review rather than added as a convenient free-form formula.

Guardrails in the wider engineering suite

The valuable output is not the document count; it is the refusal behavior exercised across public sources, documented-layout fixtures, and separate synthetic or adversarial regressions. Not every item below derives from the seven raw files. In particular, the internal FP-8 and Pixior regressions are outside that seven-file register.

• Synthetic example (separate regression): a percentage rate can appear in a quantity field. In FP-8, one late-fee row stored the 5% rate as 0.05 and its $205,403.51 financial base in the displayed rate field. Misreading that single row as a unit-priced service produced a false $10,268.53 variance. The previously stated $13,669.02 was a two-row aggregate, not the amount of that singular row.
• Dispute timing can be unknown. Missing contract-window evidence must not fall through into claimable money.
• One order can produce several shipments. Duplicate logic needs an economic-event reference appropriate to the fee.
• Summary grain blocks line-level conclusions. A period category total cannot safely answer a per-order question without supporting detail.
• A PDF text layer can be present and still be unsafe. Number density, character quality, and layout admission are distinct checks; a malformed layer should degrade or refuse.
• Receiving quantities can use the wrong unit. Units received do not automatically equal pallets received, even when both appear on the same row.

These lessons now appear throughout the billing-error check library and audit checklist.

What the corpus cannot prove

• No merchant benchmark: the committed engineering corpus contains no private production merchant dataset.
• No prevalence claim: it cannot show how often any billing pattern occurs across providers or merchants.
• No provider allegation or acceptance: a documented layout identifies a parser target, not a provider-approved interpretation or a claim about real billing conduct.
• No universal coverage: four layout variants do not represent every export, version, customization, currency, or document quality.
• No certification by existence: passing golden fixtures does not certify the exact check/provider/parser-layout cell required for external publication.
• No recovery evidence: a synthetic result cannot demonstrate a submitted dispute, acknowledged amount, invoice credit, or landed recovery.

What comes next

The next evidence tier is a rights-cleared, privacy-preserving, held-out real-invoice corpus split by provider and parser layout. It needs explicit provenance, legal use rights, fixture extraction that removes private merchant data, per-cell labeling, and false-positive acceptance criteria before any certification can change.

Until then, the public count stays scoped: seven hashed public source documents in one raw-source manifest; four separately documented invoice-layout variants with synthetic values; and zero certified publication cells.

Version note

Corpus 2026-08-12 · article 1.1. The 14 August accuracy correction separates the seven-file raw-source register from the documented-layout and regression suites and corrects the singular FP-8 amount. A future source addition or substantive correction will receive a new corpus or article version rather than being presented as part of the original snapshot.