Wrong numbers come from retrieving the wrong document, extracting the wrong number from the right document, or asking a model to do arithmetic.
2
Kepler preprocesses documents with code, assigns facts unique identifiers, and lets the model plan around those identifiers instead of generating numbers itself.
3
AI systems should verify sources and arithmetic, enforce document access rules, and refuse outputs when a fact cannot be resolved.
Summary
Vinoo Ganesh argues that Wall Street's distrust of AI comes from a structural problem rather than hallucination alone. A finance system must produce work that someone can review and trust, which means its numbers need a traceable source and a clear failure mode. He breaks incorrect facts into three causes: the wrong document, the wrong number from the right document, and model-generated arithmetic. Kepler addresses these before and during generation. Code extracts facts from documents and assigns them identifiers. The model plans which facts to use, while code performs retrieval and computation. A number that has no resolvable identifier cannot appear in the output. Ganesh also describes searching widely before narrowing results, enforcing tenant and user scoping, and correcting or refusing citations at search time. His advice to AI buyers is direct: ask whether the system writes pointers, checks citations and arithmetic, prevents unauthorized files from reaching clients, and explains what happens to facts that came from nowhere.
Trust requires a person who can be held responsible for an error
Ganesh says firms need to trust AI work before they can use it in serious investment workflows. Trust means there is someone to blame when the result is wrong. He points to a gap between confidence and accuracy: 80% of finance executives trust numbers, while 31% find errors in them. Existing systems often demo like a smart intern, or work when connected to the right structured data, but still fail on the last part of the task. For Ganesh, retrieval alone is not enough. The system has to make its sources and failure modes inspectable before a firm can rely on its output.
Ganesh separates wrong facts into three failures. The system may retrieve the wrong document, such as an internal valuation instead of a trusted analyst's work. It may find the right source but extract the wrong number, such as using 2025 revenue when the question asks for 2023 revenue from a 2025 10-K. Or the model may perform mental math. He says models predict the next token rather than calculate, so arithmetic should go to deterministic tools. A citation filter that checks the answer only after generation does not solve the structural problem, because the incorrect answer has already been produced.
Kepler gives facts identifiers before the model starts answering
Kepler preprocesses unstructured documents into a structured data store. Code writes every number and assigns facts, such as a number in a 10-K, unique identifiers or pointers. The model acts as a smart analyst that plans the route through the data. It can decide which source and entity to retrieve, but code performs the retrieval and math. The model reads pointers rather than receiving permission to invent arbitrary values. Unless it can dereference a pointer, the corresponding fact cannot be generated in the output. Users can click source chips to inspect how a result was obtained.
Deterministic tools complement human investment judgment
Ganesh draws on his experience as head of business engineering at Citadel. He says portfolio managers are talented at understanding data, but they work with deterministic tools because the data must be right. Kepler stores information deterministically when possible. For information that cannot be stored that way, it uses preprocessing and hybrid storage, then combines reranking with pointer dereferencing to retrieve it. Ganesh compares this with fundamental investors finding missing information through GLG or Guidepoint calls, experts, and company management. The system chooses a path, looks up what it knows, and handles missing information separately.
A lookup has different failure properties from a search
Ganesh says a lookup cannot get a stored fact wrong, while search can. Repeating the same search with the same query can produce different results because model-based systems are probabilistic. Kepler therefore treats retrieval as a document-ranking pipeline. It uses ANN and BM25 to collect a broad candidate set, reciprocal ranking to narrow it, and cross-encoders to compare the query with document chunks. The exact counts he gives are up to 200 initial documents and about 60 after reciprocal ranking. This lets the system search broadly before narrowing, instead of allowing a model to stop at the first convenient result.
Search access must be scoped to the user and tenant
Ganesh describes security failures that can expose one firm's documents to another. A system may forget to filter the tenant, let a caller claim to be another firm, return stale information, bypass search rules through a direct document read, or query an index that is not populated. He says these problems require searches to carry their security constraints. The system must limit the documents returned for each query based on the user's and firm's access. This is separate from whether the retrieved text looks relevant. A correct document can still be an unauthorized document.
Kepler's verification process stores facts and their pointers deterministically. If a hash-map lookup does not find a value, the system receives nothing back. Ganesh describes a contract in which citations are verified, corrected at search time, or refused. A model that cannot dereference a pointer cannot print the number. He applies the same idea beyond finance, including legal precedents. A case that was never stored in the source corpus should not appear as a newly invented precedent. Post-processing remains necessary because a probabilistic model can still generate a number after reading the available pointers.
Ganesh advises builders and buyers to ask whether the model writes pointers and whether every citation and arithmetic operation is checked. They should be able to click a number and see its source, whether the value came from extraction, calculation, or both. They should also ask how the system prevents unauthorized files from reaching customers and what happens when a fact comes from nowhere. His production example shows refusals dropping sharply after pointer dereferencing was introduced, though not to zero because models can still generate values during inference. That remaining failure is why the additional post-processing check matters.