Claims that fill themselves · For every employee and approver
Scan a receipt, approve it in one look
- The problem today
- Employees type every receipt into a form; approvers wade through every claim to find the few that need a decision.
- What it does
- A model reads the receipt into the claim and matches it to the card line. Policy checks run in code, so the approver sees only the claims that need a decision, with the reason.
- What you see
- Two phones: the claim filling itself in, then the approver's one-line view. Flip the model under them to compare.
Internal benchmark on synthetic data · run run-20261022172600-test · receipts in the test split
Fastest on this run: Mid-tier frontier. Lowest cost per claim: Low-cost frontier. Most claims right on every field: Mid-tier frontier.
Show figures for each modelHide figures
Dvar small model
- Every field right
- 149 of 159
- Median time to fill
- 2.48 s
- Model cost per claim
- $0.00040
Low-cost frontier
- Every field right
- 149 of 159
- Median time to fill
- 727 ms
- Model cost per claim
- $0.00023
Mid-tier frontier
- Every field right
- 150 of 159
- Median time to fill
- 541 ms
- Model cost per claim
- $0.0055
| Model | Every field right | Median time to fill | Model cost per claim |
|---|---|---|---|
| Dvar small model | 149 of 159 | 2.48 s | $0.00040 |
| Low-cost frontier | 149 of 159 | 727 ms | $0.00023 |
| Mid-tier frontier | 150 of 159 | 541 ms | $0.0055 |
Fastest on this run: Mid-tier frontier. Lowest cost per claim: Low-cost frontier. Most claims right on every field: Mid-tier frontier.
Internal benchmark on synthetic data · run run-20261022172600-test · 2026-10-22 · 3 models, same instruction
GOLDEN LANTERN RESTAURANT LIMITED 2 Bay Pde #04-23 Hong Kong Date: 2026-10-07 22:14 Pax: 4 ------------------------------ Steamed fish 824.09 Fried rice 569.83 Dim sum basket 456.08 ------------------------ T0TAL HKD 1850.00 Paid: VISA ****7430 Thank you!
Note: Team dinner, 4 people, end of sprint
No claims waiting. Scan and submit a receipt on the employee's phone.
Behind the answer
- Model
- dvar:glm-5.3-flash
- Time
- —
- Cost per claim
- —
- Scorer
- Press the button above
Press Scan receipt, then flip the model to compare. The fields fill at the speed each model replied in the frozen run, with the wait cut to 3.50 s; the measured time is shown under the claim and here.
Run run-20261022172600-test, 2026-10-22. Same instruction for every model, temperature 0. Cost is from tokens and list prices.