Deterministic vs probabilistic calculation in finance AI
A language model samples tokens. Two identical questions can return two different numbers. That property is acceptable for drafting and unacceptable for a reported figure.
The design question is not whether to use a model but where the deterministic boundary sits. In Octopus AI, everything numeric sits on the deterministic side.
Each AFI capability documents that boundary in its published system card, so a model-risk reviewer can see exactly what the model does and does not do.
Questions and answers
What runs deterministically?
Reconciliation, consolidation, currency conversion, variance arithmetic, threshold evaluation, and forecast rebasing. All of it is code executed against reconciled source data.
What does the language model do?
Language only: reading a question, drafting a request for commentary, and explaining a computed result in plain words. It never originates a figure.
Why does this matter for SR 11-7?
SR 11-7 model risk centres on a model's output driving decisions. When arithmetic is deterministic code, model drift or a version change cannot alter a reported number, which narrows the model-risk surface to language behaviour.
Is a deterministic system less capable?
It is differently capable. Determinism removes generated arithmetic, not automation. The initiating workflow — detecting a variance, contacting the owner, logging the answer, rebasing the forecast — remains autonomous.
How is the boundary verified rather than asserted?
By reproducibility and lineage: the same inputs return the same figure, and every figure traces to source line items. Both are testable without access to model internals.
Proof
- Reference client: $523K total contract value over 3 years.
- Approximately 430 analyst hours saved.
- Forecast cycle cut from about 60 hours to under 6 hours (90% faster).
- Error rate below 0.1%.
- 2,500+ autonomous tasks completed in month one.