The Problem: Probability vs. Determinism

Financial services require absolute numerical accuracy, yet LLMs are fundamentally probabilistic, designed for next-token prediction rather than deterministic arithmetic. Current AI implementations often rely on "citation" as a proxy for trust, which is insufficient because it acts as an after-the-fact audit rather than a guarantee of correctness. In finance, where firms may reach opposite conclusions from the same data, the goal is not a single "ground truth" but a verifiable process that respects a firm's specific internal logic and ontologies.

The Three Tenets of Verifiable AI

To move from AI-generated content to verifiable work product, Kepler employs a system that treats the LLM as a component rather than the entire engine:

• Atomic Provenance: The model is prohibited from manipulating or "understanding" numbers. Instead, it acts as a pointer, identifying the source of a data point. The system then uses deterministic tools to extract and verify that number against the source. If a number cannot be independently verified, it is stripped from the output entirely.
• Scope Determinism: This principle mandates that the LLM should never perform math. It uses its reasoning capabilities to plan the workflow—deciding which data is needed and what operations to perform—but delegates the actual computation to deterministic code (CPU cycles). This reduces costs and eliminates hallucinations in calculations.
• Derivation Chains: Every output is treated like a pull request in software engineering. By storing the chain of events—the source, the transformation logic, and the final output—the system creates a replayable audit trail. This allows firms to codify their unique methodologies (e.g., specific ways of calculating EBITDA or enterprise value) into the AI's workflow.

Shifting from Token-Maxing to ROI

Industry focus is currently trapped in "token-maxing," where companies are rewarded for high consumption rather than high-quality output. The next phase of AI adoption will mirror the shift seen in cloud infrastructure (e.g., Snowflake/Databricks), where the focus moves from simply running workloads to cost optimization and ensuring every dollar spent contributes to a verifiable, usable pipeline. The ultimate goal is to automate the "painful, repeatable tasks"—such as parsing 10Ks or earnings transcripts—to reclaim the time of high-value analysts, allowing them to focus on judgment rather than data entry.