Zero LLM Variance: Why Evaaluate Uses Deterministic Computation for All Financial Math
Large Language Models are useful tools for reading documents and drafting text. They are unsuitable for performing the financial calculations in a valuation report. This article explains why Evaaluate keeps a strict separation between tasks a language model may assist with and the deterministic computation of the valuation.
The Variance Problem
LLMs are probabilistic systems. Given the same prompt, they may produce different outputs across runs. That is acceptable for drafting prose and unacceptable for financial computation. When a professional signs a report stating a fair value per share, that number must be exactly reproducible from the stated inputs and methodology.
A model asked to carry out a multi-step DCF can also make arithmetic slips that are hard to spot in a long answer. A figure that might differ on the next run, or that cannot be traced step by step, is not one a valuer can defend.
Where LLMs Help
The distinction is not that LLMs are bad tools; it is that they are the wrong tool for arithmetic. In Evaaluate a language model has two optional uses. It can help read an uploaded statement, filling gaps the rule-based parser could not locate, and what it fills in is flagged in the extraction warnings. And it can draft the narrative sections of the research pack from sourced facts.
Neither use decides a number. Extracted figures pass through the valuer's review of the inputs, and a drafted section is printed in the report only after the valuer approves it.
Where Deterministic Computation Is Required
Every calculation that feeds into the valuation is performed by deterministic code. This includes the DCF projection, the discount rate build-up, the terminal value, the net asset value, the comparable-company multiples, the weighting of methods and the sensitivity analysis in the risk view.
These computations are implemented as ordinary functions with defined inputs and outputs. The same inputs always produce the same output, regardless of when or how many times the computation runs.
The Architecture: Separation of Concerns
The boundary runs between text and numbers. Text tasks may use a language model; numeric tasks never do. The two sides exchange structured data: sourced facts go to the drafting step, and computed results go to the report. At no point does a language model perform a calculation that affects the valuation.
Verification and Auditability
Deterministic computation makes the result auditable. For any valuation a reviewer can inspect the inputs as confirmed by the practitioner, each with its source, and recompute the result in the Excel workbook, where every computed line is a live formula over those inputs. Each computation run is stored with a hash of its inputs, and the engine is covered by an automated test suite.
The Practical Impact
When a reviewer questions a number in the valuation report, the practitioner can trace it back through the computation to the input and the source behind it. The number will not be different if the computation is run again on the same inputs.
Language models and deterministic code are complementary tools. Kept apart, the first helps with reading and drafting and the second produces figures a professional can check, explain and sign.