Deterministic checks
For finance, correctness beats eloquence.
How we're different
- Matching and arithmetic are validated deterministically.
- LLMs draft explanations—not the underlying numbers.
- Ambiguity routes to exceptions, not silent auto-accept.
Trust rule: AI drafts, humans decide for sensitive outcomes.
Examples
- Bank vs GL matching thresholds
- Duplicate detection and date normalization
- Variance thresholds and required evidence
Confidence bands
- High: auto-suggest and log
- Medium: show rationale; reviewer confirms
- Low: exception inbox + next-step checklist
Deterministic scoring is transparent and testable.