How to test
The tour as a checklist
Each stop says what to do and what you should see. The guided tour walks the same list with a bar at the bottom of each page.
1. A team dinner over the limit
DoPress Scan receipt, watch the claim fill in, then press Submit claim.
CheckMerchant, amount, category, GL code and people fill in one by one, and a note says how far over the Hong Kong meal limit the dinner is per person. The approver sees the same flag.
2. A Chinese receipt, from the approver's side
DoRead the approver's card, then flip the model in Behind the answer.
CheckOne line, a Receipt matched chip, ready to approve. Under it: each model's speed, its cost per claim, and whether every field matches the answer key.
3. A cash taxi next to a look-alike card charge
DoPress Scan receipt. Then pick another model and scan again.
CheckThe ride was paid in cash. A same-amount card charge from another day is in the list; a model that matches the claim to it is marked critical, because it would hide a cash reimbursement in a card line.
DoRead the answer, then press Show the query.
CheckA one-line answer and a chart, computed from the ledger. The model wrote only the small JSON query, so it cannot make up a figure.
DoRead the table, then switch models above the Exceptions queue.
CheckShare of items closed with no review, critical failures (any one blocks the model), accuracy per task, latency, and cost per company-month, all computed from the frozen run.
6. The bill whose bank details changed
DoRead the bill and each model's reading.
CheckHowever the bill is read, the account differs from the vendor record, so code sends it to a call-back. The model reads; code decides.
DoRead the Compare screen, then open a model's row.
CheckDvar's evals product on the claims workflow: each model against today's model on the same gold set, its p99 time and cost, and the four release checks with the reason each one fails or passes.
DoSkim the scorer rules and what is real.
CheckNo AI judge: field-by-field scoring after fixed normalisation rules, a fixed list of critical failures, and a list of what is written for the demo.