System architecture

Five roles.
One evidence chain.

Models interpret. Code calculates. Filings provide evidence. Validators verify. Humans decide.

FinTrace workflow showing four specialist LLM calls and one second-pass LLM call, with deterministic aggregation, calculation, retrieval, quote validation, and human review
01

What the LLM does

Four scoped prompts identify candidate revenue, cash-flow, related-party, and communication-pattern issues. A strict second pass classifies only against supplied evidence.

02

What Python does

Python recomputes growth, free cash flow, margins, and direct values using filed inputs and explicit tolerances. Model arithmetic is never authoritative.

03

What retrieval does

The retrieval stage ranks multiple passages from statements, footnotes, MD&A, policies, segment notes, and reconciliations for each stable finding.

04

What validation does

Every proposed explanation quote is matched against the filing and checked for a direct relationship. Numerical explanations must bridge the actual gap.

05

What the human does

An analyst reviews the complete evidence chain and records accept, reject, or needs-review for every finding. The decision updates the integrity digest.

EXPLAINED requires all four gates

Specific filing evidence+Direct connection+Valid quotation+Numerical reconciliation=EXPLAINED
If any gate fails, the result is UNRESOLVED. Example B12-F2: a $3m cash payment cannot reconcile a $5m free-cash-flow gap.