The Common Pattern Behind Models, Notes, Reconciliations, and Escalations
The financial workflows in anthropics/financial-services differ in purpose, evidence, and consequence, but they can be judged through one common operating model. A DCF model, an earnings note, a GL exception report, and a KYC escalation packet all turn inputs into a bounded artifact through a named method, a set of checks, and a defined authority boundary. The useful synthesis is not a recap of each workflow. It is a practical test for comparing them without flattening their domain differences.
The invariant is a chain from evidence to accountable artifact
The first question is what enters. The repository does not treat every input as equally authoritative. Structured market data, internal ledger records, an earnings transcript, a counterparty statement, and an onboarding document have different origins and different reasons to be trusted or challenged. A workflow that does not make that distinction visible cannot explain why one number is acceptable, another must be cited, and a third should be treated only as a claim to verify.
The second question is what changes the input. A DCF applies assumptions and formulas to historical data and projections. A comparable-company analysis selects a peer set and computes consistent metrics. A reconciliation normalizes two extracts, joins them at a common grain, and classifies differences. A KYC workflow applies a trusted rules grid to a record extracted from documents and screening results. These transformations are not interchangeable. The invariant is that the transformation is named and inspectable, not that the same calculation appears everywhere.
The third question is what evidence explains the transformation. In a model, source comments and linked formulas show where inputs came from and how outputs are derived. In a comps spread, consistent definitions and a documented source hierarchy explain why the multiples can be compared. In a reconciliation, both-side values, keys, tolerances, and evidence references show why a row was classified as a break. In KYC, rule identifiers, factor tables, and document status explain why a case was routed. A result without that path may be fluent or polished, but it is difficult to review.
The fourth question is what artifact is produced. The artifact is not just a container for a conclusion. It is a bounded representation of the work: a workbook with formulas and checks, a note with a variance table, an exception report with proposed resolution, or an escalation packet with missing documents and screening hits. The artifact makes the workflow portable because it gives another person or system a stable object to inspect, revise, and route.
The fifth question is what checks constrain it. Formula auditing, balance checks, peer comparability, reconciliation critics, schema validation, and rule citations do different jobs. Some test syntax or arithmetic; some test consistency; some test evidence; some test whether the output has crossed an authority boundary. They should not be collapsed into a single idea of “quality.” A schema can constrain the shape of an extraction without proving that the extracted name is correct. A balanced workbook can still contain a poor assumption. A plausible break category can still be an unverified hypothesis.
The final question is who may decide or act. The repository consistently places professional review, approval, distribution, posting, or execution outside the agent. The artifact can be generated by an agent and checked by another worker, but the decision that changes a client status, posts a ledger, publishes a research view, or sends a client-facing document remains with the designated human or firm system.