Pay the contract rate, not the invoice.

• 0 s1 s2 s3 s4 s

  1. Received, at 0.0 s
  2. Read3.3 s, at 0.0 s to 3.3 s
  3. Matched, at 3.4 s to 3.4 s
  4. HeldLinehaul above the rate card, +$801.54+$801.54, at 3.4 s

  Every step on the record

  Sample invoice 76644482, October 17, 2026, in seconds from receipt. Recorded before Claude first went live, so Gemini 3.5 Flash-Lite read it.

• ReplayInvoice 53704390, 1 attempt

  Reconciliation failed

  1. Uploaded
  2. Reading
  3. Checking
  4. Checked

  shipment has no miles

  Failed jobs wait to be replayed

  A job that uses up its retries waits with its error, ready to replay once the cause is fixed.

• Linehaul billed twiceTern & Tide Carriers #96079756, open+$5,522.06

  Fuel surcharge above the rate cardBramblewood Express #47222915+$103.23

  Inside delivery charged, but not on the rate cardBramblewood Express #14835569+$129.40

  Findings that name the amount

  Each finding names its line and what it costs, before anyone pays.

• $0.0105

  per scanned invoice, Claude first, measured November 2, 2026

  • $0.0079Digital
  • $0.0105Scanned

  • $0.0007Digital
  • $0.002Scanned

  Reading costs a fraction of a cent

  Per invoice at paid rates on 98 test invoices; the targets are $0.02 / $0.05.

• packages/domain/src/reconcile.ts

  69  finding = { type: 'rate_mismatch', expectedCents: rate, evidence };70}71if (amountCents > finding.expectedCents) discrepancies.push({ ...finding, lineNumber, billedCents: amountCents });72if (code !== null) billed.add(code);

  Rules in code, not in the model

  The model only reads the invoice; plain code checks each charge against the rate card.

• • Let the billed amount standApprove charge
  • Dispute the overchargeDispute

  A person decides each finding

  Each decision goes into the audit trail with its note, and an invoice is approved only once every finding is decided.

• A redacted request headerSynthetic data only

  { "headers": { "x-api-key": [Redacted], "cookie": [Redacted] } }

  Keys stay out of the logs

  The demo holds only generated invoices from fictional carriers, and every service redacts keys and cookies from what it logs.

Evaluation results

98 synthetic invoices with known ground truth, read two ways. Claude first is Claude Sonnet 5.5 reading first, with the free models behind it (measured November 2, 2026). The free models alone are what CI checks on every change to extraction, and what reads once Claude's credits end (measured October 27, 2026).

On these runs Claude first is ahead on header field accuracy (digital), line-item F1 (scanned) and overcharge recall, and the free models alone on header field accuracy (scanned) and cost per invoice, paid-equivalent (digital / scanned).

Complete results with 95% confidence intervals are in the eval reports: Claude first and free models only.