Concordance: Differential Verification of Financial Planning Models by Independent Re-implementation

Financial planning calculators produce numbers that inform consequential decisions, and they almost never ship with evidence that they compute what they claim to compute. Unit tests are a weak control here: the expected value that makes a test is usually derived by the same reasoning as the implementation, so a misread convention passes its own test. We describe Concordance, a verification control operated in production across a catalog of 55 financial planning models. Each model is specified in prose and pseudocode, implemented once in TypeScript for the production engine, and re-implemented independently in Python from the specification text alone by a process with no access to the engine source. A harness compares the two implementations on 250 inputs per model — domain corners, hand-written edge cases, and seeded pseudorandom samples — requiring agreement to a relative tolerance of 1e-9 or an absolute tolerance of 1e-6. Disagreement fails continuous integration and blocks release. As of 27 August 2026 the two implementations agree on all 13,750 cases. Applying the method to an engine already in production surfaced defects in 12 of the 55 models, concentrated in unit and compounding-basis errors, degenerate-input branches, and piecewise functions evaluated with constants valid only at one point of their domain. We report the defect taxonomy, the cost of the control, and its limits: agreement between two implementations of one specification is evidence about implementation fidelity only, it does not establish that the specification is financially appropriate, and independently written implementations are known not to fail independently. We describe the external anchors and third-party cross-checks used to reduce, without eliminating, that correlation.