The receipts behind the math.
Delta Prime is built on one hard commitment: the AI never does the math. A deterministic calculation engine does — and the fair question is whether that engine is actually right. This page is the answer, in the engine's own terms.
We validate it the way you'd audit a calculator: run the same case through a free, authoritative outside reference and compare the numbers. Every check below runs automatically on every build, against answers frozen in our test suite, so the agreement can't quietly drift. Where a check is weaker, or doesn't exist, we say so.
How we check it
A deterministic engine does the arithmetic.
Every number in your plan comes from ordinary rules-based software that follows fixed formulas and published tax rules — the same kind of code that runs tax-prep and actuarial systems. Same inputs, same answer, every time. The AI only interprets your question and explains the result; it never computes.
We check it against outside references.
For each part of the engine we run the same case through a free, authoritative outside reference — an independent calculator, or the government's own published tables — and compare. This is differential testing: agreement with a source we don't control, not just agreement with ourselves.
The checks run on every build.
Each reference answer is frozen into our test suite, and the engine is re-run against all of them automatically, every time we ship. Hundreds of cases, continuously. A number can't drift out of agreement without a test turning red first.
A financial planner might sanity-check a handful of numbers once. These checks run hundreds of cases across every part of the engine, on every build, and never expire quietly. For the plain-language version of why the engine isn't AI, see How It Works.
The receipts
Each row is a part of your plan, the outside reference we check it against, and how close the engine lands. “To the cent” means the largest difference we observe across the test battery is under one cent.
| Part of your plan | Checked against | How close |
|---|---|---|
| Federal income tax — brackets, long-term capital gains, qualified dividends, the 3.8% net investment income tax, taxation of Social Security, and the standard deduction (including the age-65 addition) | NBER TAXSIM — an independent academic tax model used in hundreds of published studies | To the cent |
| Social Security — the reduction for claiming early, the delayed-retirement credit, and spousal and survivor benefits | The Social Security Administration's own published benefit-adjustment tables (Pub. 05-10147) | Within 0.1 point |
| Required Minimum Distributions from retirement accounts | The IRS Uniform Lifetime Table (Publication 590-B) | Divisors match exactly |
| Medicare IRMAA premium surcharges | The published CMS 2026 income tiers and surcharge amounts | Thresholds & surcharges match exactly |
| ACA marketplace subsidies — the required-contribution percentages and the 400%-of-poverty cliff | IRS Rev. Proc. 2025-25 and the 2025 HHS poverty guidelines | Match exactly |
| Alternative Minimum Tax | IRS Form 6251, worked line by line | To the cent |
| Self-employment tax, the QBI deduction, and solo-401(k)/SEP contributions | Schedule SE, Form 8995, and IRS Publication 560, worked by hand | To the cent |
| Rental real estate — depreciation, §1250 recapture, and passive-loss limits | 2026 tax law, worked by hand | To the cent |
| Founder / early-employee stock — the §1202 (QSBS) exclusion | The 2026 statute, worked by hand | To the cent |
| Federal student-loan repayment — income-driven, RAP, ICR, and standard plans | Federal repayment rules (34 CFR §685.209, P.L. 119-21) | To the cent |
| Estate & trust planning factors — GRAT / CRAT annuity and remainder values (§7520) | IRS Publication 1457, Table B | To four decimal places |
| Longevity — the life-expectancy table behind survival projections | The SSA Period Life Table's published life expectancy at 65 | Within 0.05 year |
| Retirement success rates — the Monte Carlo simulation | The published Trinity / Bengen safe-withdrawal studies, over 1928–2025 market history | Reproduces the 4%/30-yr figure within ~1 point |
The federal income-tax check (highlighted) is the flagship: more than 200 sample households spanning every tax bracket, the capital-gains rate zones, the net investment income tax, and the taxation of Social Security, all reproduced to the cent against an independent calculator we don't control. It also caught a real bug — Social Security taxation that omitted realized capital gains — which the comparison surfaced and we fixed.
What this does and doesn't cover
Trust copy is only worth anything if it's honest about the edges. Here is where the checks are strong, and where they stop.
- State income tax is an approximation — and it is not oracle-validated.
The federal check above is the rigorous one. State income tax is modeled as an estimate — a few states use their real brackets and standard deductions, and the rest use a flat top-marginal rate — but none of it is checked against an independent tax calculator the way the federal numbers are. Treat the state portion as a reasonable estimate, not a validated figure. - The strongest check is federal tax; the rest are checked against official tables or worked by hand.
Only federal income tax runs through a fully independent calculator (TAXSIM). The others are validated against the official published figures from the IRS, SSA, CMS, and HHS — and where no public calculator exists, we work the IRS form or statute by hand and have a second reviewer re-derive it. That's genuine external ground truth, but it isn't all the same kind of check, and we'd rather say so. - The tax mechanics are validated at 2023; the current-year dollar amounts are checked separately.
The public TAXSIM deployment computes federal tax through 2023, so that's the year we run the head-to-head comparison. The engine's formulas don't change year to year — only the dollar thresholds do — so 2023 validates the mechanics, and the current-year brackets and constants are checked separately against the published IRS figures. Tax constants track current law and are updated as the law changes. - Projections are modeling, not guarantees.
Monte Carlo results are drawn from historical market data using standard actuarial and statistical methods. They describe a range of possible outcomes; they are not a forecast, and nothing here promises any particular investment return. Past market history does not guarantee future results. - Longevity uses a healthier-than-average default, which you can change.
By default we use an annuitant mortality table, which assumes a longer-than-average life and so is slightly optimistic about how long a portfolio has to last. A general-population SSA table is one setting away for a more conservative view — and it, too, is validated against SSA's published life expectancy. - This validates arithmetic, not your decisions.
These checks prove the engine computes the tax and planning math correctly. They are not a substitute for filing your taxes, and Delta Prime is not a registered investment adviser. When a question needs a licensed professional — a CFP, CPA, or attorney — we tell you plainly and point you there.
Projection, not a guarantee. Educational illustration, not financial advice. Delta Prime AI is not a registered investment adviser.
Join the beta