Deterministic finance math for AI agents (why LLMs shouldn't compute an IRR)
Large language models predict tokens; they do not run a numerical solver. Ask a model for the IRR of a cashflow series and it returns a plausible-looking number that is often wrong by enough to flip a decision. For pr…
The hallucination problem, Determinism as a contract, Evidence over assertions
The hallucination problem
Large language models predict tokens; they do not run a numerical solver. Ask a model for the IRR of a cashflow series and it returns a plausible-looking number that is often wrong by enough to flip a decision. For pricing, valuation, or risk, 'usually close' is a liability, not a feature.
Determinism as a contract
A deterministic endpoint is a pure function: the same inputs always return the same audited output. When CashflowLens solves an IRR it returns the rate that actually zeroes the net present value, found with a bracketed solver — not an estimate, and not a different answer on the next run.
Evidence over assertions
Every response carries per-rule findings: which checks fired, the values behind them, and a fix or note. An agent — or the human reviewing the agent — can verify how each number was produced instead of trusting an opaque guess. Nothing is fabricated.
How to wire it in
Route the math your agent can't be trusted to do to a deterministic endpoint and keep the LLM for orchestration and language. The agent calls the API, pays per request over x402, and gets a result it can defend.