Every rule in the corpus cites the statute, regulation, or official form instruction it encodes, and every answer ships with a machine-checkable proof tree (PROOF-FORMAT.md). This document describes how the engine's correctness is tested.
Parameters (brackets, caps, phase-outs, tables) are verified against the issuing agency's current publications; every rule carries its citation, a verbatim source excerpt, and the validity window it asserts. Where an official printed table and its generating formula can differ by $1 (whole-dollar tables built on $50 brackets), the rule documents which convention it encodes and why — see the VA tax-table and NY schedule notes in the state rules. Below $100,000 the federal engine reproduces the printed IRS Tax Table method exactly (band-midpoint formula, half-up rounding), pinned by a test suite sampled from the printed 2025 table across all four statuses.
On TaxCalcBench TY25 (50 full returns, federal + IL/VA/CA/NY), a cold Claude Sonnet agent using the opentax MCP server as its computation oracle scores 48/50 returns exact under strict scoring (96%), 805/820 scored lines (98.2%) — one attempt per case, scored with the benchmark's own evaluator. The two remaining cases contain reference-data inconsistencies that the engine's line-level reconciliation surfaced, reported upstream (tax-calc-bench#96).