Fictional preview timing and model-cost estimates
Measured 2 November 2026. Twenty real preview requests for each published task, using fictional details only. This is a development measurement, not an accuracy claim about customer documents or independent production acceptance.
Trade documents use supplied structured fields and deterministic arithmetic. Cross-check runs use a fictional invoice PDF and bill-of-lading Word document with three planted differences; amendment runs request one natural-language edit. Both AI paths are restricted to authorised fictional tests, not customer processing.
| Task | Runs | Credits per run | Total model calls | Median seconds | p90 seconds | p90 USD model estimate |
|---|---|---|---|---|---|---|
| invoice | 20 | 2 | 0 | 0.67 | 0.79 | 0.00000 |
| packing list | 20 | 2 | 0 | 0.69 | 0.74 | 0.00000 |
| shipping instruction | 20 | 2 | 0 | 0.77 | 0.85 | 0.00000 |
| pair | 20 | 4 | 0 | 1.41 | 1.57 | 0.00000 |
| set | 20 | 5 | 0 | 1.99 | 2.11 | 0.00000 |
| cross check | 20 | 3 | 20 | 9.05 | 11.22 | 0.03123 |
| export data | 20 | 3 | 0 | 0.70 | 0.80 | 0.00000 |
| origin data | 20 | 3 | 0 | 0.71 | 0.93 | 0.00000 |
| amendment | 20 | 1 | 20 | 3.22 | 4.05 | 0.00223 |
What the figures mean
Each run is matched to the existing credit ledger and token meter. p90 is the 18th ordered result out of 20. Model cost uses configured gpt-6-astra token rates: USD 10 per million input tokens, 1 per million cached input tokens and 50 per million output tokens. This is a configured-rate estimate, not a provider invoice or independent proof of provider/model identity. Zero means no model call for that structured task, not that operating the service is free.
The model-only three-times-cost guard passed for this bounded corpus at the lowest approved pack value per credit, using a planning stress assumption of 2 AUD per USD. That assumption is not an exchange-rate quotation. Hosting, payment fees, staff time and legal review are excluded; no full-service profit or future-cost guarantee is made.
Limits and privacy
These small, controlled inputs do not establish real-world extraction accuracy, origin eligibility, customs acceptance or LC compliance. Source uploads and amendment results are processed in memory; generated structured drafts retain the 30-day download expiry. Customer AI processing, checkout and human review remain off.
Measured build: aul-desk-ed25a9eab65a · Read the measurement summary (JSON). Full source receipts: artifacts/COST_CALIBRATION.json in the project work folder.