Broadening coverage. Deepening verification.
The federal alpha brings together a tax-rule dictionary, document reading and printed-form output. This page shows what each part covers and how the project checks its work.
At a glance
- Available now: tax year 2025, federal returns only: Form 1040, 15 of its schedules and 65 other IRS forms, filled by ytax-prep.
- Read from your documents: 14 kinds of tax form, read from their PDFs; figures from other forms are entered by hand.
- What's next: state returns, tax year 2026 and more situations on the forms listed.
- How it is checked: built with AI assistance and tested from several angles (Testing and review).
Explore what is available
The release includes Form 1040, related schedules and additional IRS forms. ytax-prep reads supported document types automatically and accepts manually entered figures for other supported inputs.
Coverage is measured four ways: what the tax dictionary models, what ytax-prep reads automatically from documents, what it fills on printed forms, and what has been exercised by tests and separately implemented checks. Each tells a different part of the story.
Coverage in ytax-facts alpha-0.1.0
Counted from the release's federal tax dictionary for tax year 2025, beside the IRS Direct File dictionary for tax year 2024, with the same definitions. A writable fact is one a return can supply; a derived fact is computed from others.
The fact dictionary (XML)
| yTax, ytax-facts alpha-0.1.0 (tax year 2025) | IRS Direct File (tax year 2024) | |
|---|---|---|
| Fact XML modules | 123 | 36 |
| Facts | 20,917 | 3,030 |
| Writable facts (inputs) | 7,611 | 616 |
| Derived facts (computed) | 13,306 | 2,415 |
A fact defined both as an input and as a computed value is counted in both rows, so the two rows can add up to more than the facts (IRS Direct File: 1 fact).
Forms
| Count | |
|---|---|
| Output: Form 1040 and its schedules | 16 |
| Output: other IRS forms | 65 |
| Input: read from a PDF | 14 |
| Input: payer forms the dictionary models | 55 |
Form 1040 and its schedules (16)
Form 1040; Schedules 1, 1-A, 2, 3, 8812, A, B, C, D, E, EIC, F, H, R, SE
Other IRS forms (65)
Forms 172, 1116, 1116 Schedule B, 2106, 2120, 2210, 2210-F, 2441, 2555, 3800, 3800 Schedule A, 3903, 4137, 4255, 4562, 4563, 4684, 4797, 4835, 4952, 4972, 5329, 5695, 6198, 6251, 6252, 7206, 8283, 8332, 8396, 8582, 8606, 8615, 8801, 8814, 8815, 8824, 8828, 8829, 8839, 8853, 8859, 8862, 8863, 8880, 8888, 8889, 8915-F, 8919, 8936, 8936 Schedule A, 8938, 8949, 8958, 8959, 8960, 8962, 8978, 8995, 8995-A, 8995-A Schedule A, 8995-A Schedule B, 8995-A Schedule C, 8995-A Schedule D, 8997
Read from a PDF (14)
W-2, W-2G, 1099-INT, 1099-DIV, 1099-B, 1099-R, 1099-MISC, 1099-G, 1099-SA, 1099-OID, 1098, Schedule K-1 (Form 1065), Schedule K-1 (Form 1120-S), Schedule K-1 (Form 1041)
Payer forms the dictionary models (55)
W-2 and 1099 variants, 1098 series, 1095 series, SSA-1099, K-1s and others
How we strengthen the calculations
AI writes the rules, and independent AI agents test and verify the results; the calculation itself is done by the rule engine, the same way every time.
Automated regression checks
The development process includes authored calculation cases, dictionary invariants, structural checks and checks that generated artifacts stay consistent with their source. They keep every covered case stable as the rules grow, so an improvement in one place does not quietly change another.
Separate, answer-blind calculations
A separate verification implementation derives results from IRS publications before seeing yTax's answers. Because the two are written apart, agreement between them is meaningful, and a difference points straight at something worth examining. Its reports record exact agreement, explained method differences and anything not yet checked.
Cross-engine comparisons
The project has compared yTax with other tax calculation engines on synthetic federal returns, including scenarios from the IRS Assurance Testing System and test returns adopted from IRS Direct File.
A comparison counts only when the other engine receives the inputs needed for the line being checked. Cases that cannot be mapped are not treated as passes, even when their results happen to agree. Disagreements are analyzed against the cited instructions, and open questions are kept in view until they are resolved.
Testing the application workflow
Application tests also examine the path from documents and answers to calculated results and filled forms. Comparing the application's results with the engine's own confirms that nothing is lost between a document and the form it fills.
Turning findings into improvements
Verification finds real problems, and each one makes the rules stronger. One engine comparison found that a qualified-business-income calculation omitted the self-employed health-insurance deduction. The project added a regression case, corrected the rule and separately rechecked the result against the instructions and another engine.
Finding, explanation, regression test, fix and recheck: that is how each important correction is made and kept.
Scope of the checks
The checks are project-run and AI-assisted. They cover selected results on synthetic households and returns, and selected Form 1040 lines in the engine comparisons; a form marked verified means its destination line was checked. A form's presence does not mean every situation on it is supported, and fact counts show the dictionary's scale rather than its correctness. None of this is an external tax-professional audit, tax-agency certification or endorsement by the IRS or another engine's provider, and the alpha release is for evaluation, not for preparing or filing a real return.
What comes next
Tax year 2026 and the remaining federal forms, state returns for California, New Jersey and New York, and professional review. The aim is to serve former Direct File users and more taxpayers through yTax or software built on it. Focused professional review and new test scenarios are valuable ways to help extend the project.