What stands behind the numbers.

Read this before relying on the engine for anything filed. Every claim below names the test set it is a claim about, and every test set is in the public repository.

Exact versions

The hosted endpoint reports the root of whatever corpus it is serving on every response, so you can confirm the deployed corpus matches the one you tested.

Test commands and golden fixtures

git clone https://github.com/Invaro/opentax-engine && cd opentax-engine
pnpm install && pnpm build
pnpm test                                                   # 30 files, 2,059 tests
pnpm -F @invaro/opentax-corpus-us-federal report:coverage   # the coverage dashboard
node packages/corpus-us-federal/scripts/coverage-matrix.mjs # the machine-readable matrix

• 1,581 golden fixtures. Each is a hand-computed expected value for one rule at one date, written from the statute or the printed form before the rule was trusted, with the arithmetic in its description.
• Printed-table parity. The engine must reproduce every cell of the agency's own tax table at both ends of every row: Delaware (1,162 rows), Hawaii (2,000 × 3), Kansas (2,000 × 2), Oklahoma (2,000 × 2 plus the 2020 EIC table), Rhode Island (2,000), West Virginia (1,702), Nebraska (777 × 4), North Dakota (4,780 cells), Vermont (3,000 cells), Maine, Arkansas (every whole dollar to $100,000 plus five low-income tables), Connecticut, New Mexico, and a 333-row sample of the 2025 IRS Tax Table. Each suite states the rounding convention it proved and every exception the printed table carries.
• Composer tests with hand-computed line sets for every composed state; contract tests that every input fact has a CLI flag and a schema entry; a staleness test that the declared TY2026 gaps match the rule windows exactly.

Differential test against PolicyEngine US

Federal individual, TY2025, 572 scenarios across wages, filing status, children, age, interest, capital gains, tips, overtime, self-employment and QBI, student loan interest, and Schedule A.

The 33 are triaged in harness/known-differences.json: the CTC phase-out threshold for a qualifying surviving spouse, the § 224 tips deduction (modeled there as an uncapped exclusion), the whole-$1,000 phase-out step in §§ 224 and 225. Scope of the claim: federal individual, TY2025 only. There is no PolicyEngine comparison for TY2026, for states, or for business entities.

What "0% error margin" means, precisely

It is a statement about specific test sets, not a warranty about every return:

1. Zero unexplained disagreements over $1 against PolicyEngine US on the 572 federal TY2025 scenarios.
2. 100% of the 1,581 hand-computed golden fixtures reproduce.
3. 100% of every printed tax-table cell in the parity suites reproduce, at both ends of every row.

It does not mean agreement with a filed-return population, coverage of part-year or nonresident returns, coverage of business return line sets, or TY2026 state figures the Departments have not published. Those refuse.

TaxCalcBench

On TaxCalcBench TY25, 50 complete returns scored line by line, the engine run deterministically against this corpus with no model in the loop produces 47 of 50 strictly correct returns and 95.94% of scored lines. The three misses are two benchmark-side figures and one document-classification ambiguity. A Claude Sonnet agent using the engine over MCP scores 48 of 50, the figure on the landing page.

Known mismatches and exclusions

• harness/known-differences.json: every divergence from PolicyEngine, with authority.
• packages/corpus-us-federal/src/staleness.ts: every rule with no TY2026 version yet, per state, with the publication that unblocks it; enforced by test.
• The disclosed approximations on the coverage page.
• Where an agency's whole-dollar table and its own formula differ by $1, the rule documents which it encodes. Source conflicts are encoded as the form prints them and disclosed in the rule.
• Refusals, never approximated: CAMT, retained trust capital gains, kiddie preferential income, a fourth simultaneous AOTC student, § 199A with the § 68 haircut, part-year and nonresident state returns, state business returns, Forms 990 and 1065.