Recovery IntelligenceStart with reproducible synthetic cases, learn the shape of historical public records, then evaluate against an authorized partner cohort with adjudicated outcomes.
Generated contracts, invoices, payments and credits with planted exceptions and legitimate lookalikes. Deterministic seed 42; no customer data.
The FY2022-25 Purchases and Contracts report is a useful starting point for ingestion, agency/vendor mapping and data-quality checks.
Ask a recovery firm or enterprise AP team for a bounded, authorized historical category with adjudicated outcomes.
Run the same deterministic engine used by the command-line benchmark. Processing stays in your browser; no paid API calls.
Export rows from 1,000 base invoices
Synthetic contracts with dated amendments
Posted full or partial credit records
Planted review candidates
Ready. Rules check paid price variance and separate duplicate payments.
Candidate cases
Synthetic candidate value
Repeated paid export rows ignored
Contract data issues
| Invoice | Rule | Candidate amount | Supporting calculation |
|---|
First 12 candidates displayed. Download the full queue for all source references.
50 uncredited price variances, 50 legitimate amended prices, 50 fully credited differences, 50 partially credited differences, 50 additional duplicate payments, 50 canceled invoices, 50 unpaid invoices and 650 clean base invoices. Twenty repeated export rows test ingestion; eighteen are paid and pass through payment deduplication.
The checked-in regression result finds 150 expected cases, with 0 false positives, 0 false negatives and 0 amount mismatches. Because the fixtures and rules were developed together, this score is a software check. Independent specialist-authored cases and a blind historical cohort are required to assess generalization.
Ground truth is stored in the repository outside the published site. The engine reads only transaction records. The fixture has simplified one-line, single-currency invoices and one paid event per normal invoice; installments, tax, freight, FX and complex credit allocation require further rules.
| Record | Minimum fields | Purpose |
|---|---|---|
| Contract & amendments | Contract/vendor/item IDs, terms, unit, effective dates and document references | Establish the applicable price or obligation |
| Invoice lines | Invoice/vendor/item IDs, date, quantity, unit price, tax and freight | Compare like units against the contract |
| Payments & credits | Stable payment IDs, invoice allocations, posting date and realized credits | Separate invoice exceptions from paid exposure |
| Incumbent dispositions | Already worked, accepted/rejected, dispute, posted cash and reversals | Freeze baseline and evaluate incremental outcomes |
Propose one closed category, six to twelve months of history, stable pseudonymous identifiers and linked contract/invoice/payment/credit records. Agree a manageable sample size after measuring missing documents. Ask the partner to retain the answer key for blind scoring.
A bounded 100-row sample and retrieval manifest are checked into the repository. Its fields include record ID, purchase order, agency, vendor and purchase-order total. Next, inspect the dictionary and coverage, then reconcile counts before seeking transaction-level access.
Public JSON sample endpoint ↗npm install npm run data:oregon # bounded 100-row public procurement sample npm run data:generate # seed 42; fixtures + separate answer key npm run data:run # analyze + score; fails on mismatches npm test # edge cases + cost arithmetic npm run dev # Astro development server npm run build # static output in dist/
The demo implements structured-data rules. OCR, language-model extraction, specialist adjudication and collection remain production workstreams. Next: add independently authored holdout cases, then document renderings and an extraction evaluation.