Build for the Memo, Not the Demo — Shawn Chan, China Resources Holdings
Finance agents need claim-level provenance, explicit uncertainty, consistency checks, surfaced contradictions, and logged approval. Fluent output without those controls will not survive diligence.
Chan separates a five-minute demo from a decision memo assembled from conflicting filings, transcripts, and notes. His required controls are **claim-level source links**, visible separation of facts from estimates, automatic number checks, surfaced contradictions, and a logged human approval gate.
Build outputs to answer “where did this come from?” in one click. Preserve source trust levels and disagreement through retrieval and generation, and make the system refuse inconsistent figures instead of relying on a late manual review.
Chan separates a five-minute demo from a decision memo assembled from conflicting filings, transcripts, and notes. His required controls are **claim-level source links**, visible separation of facts from estimates, automatic number checks, surfaced contradictions, and a logged human approval gate. Build outputs to answer “where did this come from?” in one click. Preserve source trust levels and disagreement through retrieval and generation, and make the system refuse inconsistent figures instead of relying on a late manual review. These recommendations come from finance diligence experience rather than a comparative evaluation. They improve auditability but do not prove that a claim is correct, and accountability still rests with the person approving the decision.
This makes auditability an output contract rather than a final review step: every claim retains provenance, facts remain distinct from estimates, contradictions stay visible, and inconsistent numbers fail before approval. It reinforces prior finance patterns that reserve numeric mutation and validation for deterministic systems, while clarifying that traceability and formal checks verify evidence or derivation—not the truth or intended meaning of a claim.