aspireclaims & close

Month-end close · For the AI lead

How much of the month closes itself, and what it costs

The problem today
No proof, before shipping, that a cheaper or faster model is safe, or what it costs a month.
What it does
Runs one company-month through every model on one deterministic scorer: closed with no review, critical failures, latency and cost.
What you see
Scorecards per model, the exceptions queue a person would work, a drill-down into any item, and cost at your volume. The same run in Dvar's Eval Console is under Evals.

Internal benchmark on synthetic data · run run-20261022172600-test · test split

No model has a critical failure on this run, so none is blocked. Most closed with no review: Mid-tier frontier. Fastest: Mid-tier frontier. Lowest cost per item: Low-cost frontier.

Show figures for each model

  • Dvar small model

    Closed with no review
    197 of 236
    Critical failures
    0 · passes
    Median / p95 time
    2.60 s / 8.88 s
    Per 1,000 items
    $0.46

  • Low-cost frontier

    Closed with no review
    198 of 236
    Critical failures
    0 · passes
    Median / p95 time
    730 ms / 1.24 s
    Per 1,000 items
    $0.23

  • Mid-tier frontier

    Closed with no review
    199 of 236
    Critical failures
    0 · passes
    Median / p95 time
    544 ms / 755 ms
    Per 1,000 items
    $5.51

Internal benchmark on synthetic data · run run-20261022172600-test · 2026-10-22 · 3 models, same instruction

One month, 236 items, every model on the same scorer

Test split of Kestrel & Finch's October: receipts, card spend and supplier bills in SGD, HKD, AUD, USD and EUR. Each model reads every item; the policy checks then decide close or review on what it read. The answer key itself closes 210 of 236 with no review, so that is the ceiling.

A model with any critical failure is blocked, whatever its averages. Per company-month = measured cost per item × 316 items (this company's whole October). Cost from measured tokens and list prices; FX 84 INR per USD.

Exceptions queue

What a person would look at this month if dvar small model ran the close. The bank-details bill is pinned first.

No critical failures for this model on this run.

27 items to review out of 236; 1 of them the answer key would have closed (a misreading made a check fire).

Drill into one item

I261 · Supplier bill · HK

KIND: bill

SUPPLIER BILL (OCR text)
HARBOURVIEW EVENTS LIMITED
BR No: 424401533
TAX INVOICE
Invoice No.: 07-202610-941
Date: 2026-10-26
Bill to: Kestrel & Finch Pte. Ltd., Finance

Description                    Amount
Event production, deposit      36,718.77
T0TAL DUE HKD                  36,718.77

Please pay to: Pacific Rim Bank (synthetic)
Account no.: 8284833056
NOTE: our bank account has changed. Please use the new details above for this and all future payments.
Payment terms: 30 days.

Answer key against dvar small model. Switch models above the exceptions queue; a wider screen shows all three.

FieldAnswer keyDvar small model
MerchantHarbourview Events LimitedHARBOURVIEW EVENTS LIMITED
Date2026-10-262026-10-26
CurrencyHKDHKD
Total36718.7736718.77
Tax00
Invoice no.07-202610-94107-202610-941
People——
Guests00
Bank account82848330568284833056
GL code65106510
Tax codeNOTAXNOTAX
Unreadable flagnono
DecisionReviewBank details changedReviewBank details changed

Red cells differ from the answer key after normalisation (legal suffixes, punctuation, account formatting). GL 6510: Marketing: events and sponsorship.

At your volume

Type your own numbers. This only multiplies the measured unit cost and review share above; nothing here is typed in by us.

  • Dvar small modelPassesModel cost a month$0.14Items to a person a month36
  • Low-cost frontierPassesModel cost a month$0.07Items to a person a month35
  • Mid-tier frontierPassesModel cost a month$1.74Items to a person a month35

Model cost only: not the price of a product. Costs are in USD.