Method, sources, and where this model is wrong

Sources

• Federal income tax, 2026 — brackets, standard deduction, the age-65 addition, and long-term capital gains breakpoints from IRS Rev. Proc. 2025-32.
• RMD divisors — IRS Uniform Lifetime Table, Treas. Reg. §1.401(a)(9)-9(c), as published in Pub 590-B Appendix B Table III. Carried to age 120, its actual end.
• RMD start age — SECURE 2.0: 75 for those born in 1960 or later, otherwise 73.
• Social Security taxation — the provisional-income test of IRC §86.
• NIIT — 3.8% above $200,000 single / $250,000 joint.
• IRMAA — 2026 income thresholds and the two-year MAGI lookback; surcharge dollar amounts are the latest published (2025), pending the CMS 2026 release. See limits below.

None of these are typed from memory or estimated. Where we don't have a sourced figure, the tool doesn't show a number.

Where this model is wrong

• State income tax is a flat rate. This is the biggest known error. Most states are graduated and many exempt Social Security entirely; we apply one rate to all ordinary income and gains. In a graduated state, treat our state slice as an upper bound. We show it because ignoring state tax would be worse, not because it's precise.
• Monte Carlo returns are normally distributed. No fat tails, no crash clustering, no serial correlation, no valuation dependence. Real markets crash harder and in streaks. Our downside band is therefore optimistic — treat a bad path here as milder than a bad decade actually is.
• Markets are the only random input. Your lifespan, your spending, and your health are fixed at whatever you typed. Every probability here is conditional on those being right, and they won't be.
• Inflation is a single constant. One rate, applied to everything, forever.
• One household, one plan. We model a couple's tax filing but not two separate financial lives, divorce, or supporting a parent.
• No ACA subsidy cliff. If you retire before 65 and buy marketplace coverage, a Roth conversion can cost you subsidies in a way we don't model. That can dwarf the bracket math.
• The Roth analysis ignores your heirs' brackets and the ten-year inherited-IRA rule, both of which can change the answer.
• Inherited-IRA tax is estimated, not yours. The "what you keep" figure values a leftover traditional IRA at what your heirs would pay to draw it down over ten years (the SECURE Act rule), assuming their distributions stack on top of roughly $120,000 of their own income. Your heirs' actual brackets are unknown; a lower-earning heir keeps more, a higher-earning heir less.
• The age-65 standard-deduction addition is credited once, not twice. We collect one household age, so for a married couple we credit a single 65+ addition rather than assume both spouses are 65 — which would under-state tax whenever one is younger. If you are both 65+, your real standard deduction is slightly higher and your tax slightly lower than shown.
• IRMAA surcharge dollar amounts are the most recently published (2025) figures applied to 2026 income thresholds. CMS sets the following year's surcharge amounts in the autumn; until the 2026 amounts are published we use the 2025 ones, which understates the surcharge by a few percent. The thresholds and the two-year lookback are 2026.
• Tax law is modeled as frozen at 2026. It isn't. Brackets are inflation-indexed and Congress changes the rules; we hold today's law constant across a thirty-year projection, which is certainly wrong and we have no better option.

Rules we hold ourselves to

• Errors may not run in the flattering direction. A model that invents money you don't have is worse than no model. Where we must approximate, we approximate against ourselves.
• No number without a source. If we can't cite it, we don't display it.
• The same plan always gives the same answer. Our Monte Carlo uses a fixed seed. A tool where refreshing the page changes your odds — or where adding $10,000 of savings lowers them — is not something anyone should make a decision from.
• Refuse rather than default. When we haven't been told something, we say so instead of substituting a plausible-looking assumption and reporting it as your answer.

If you find something wrong here, it's a bug and we want to know. A calculator that's confidently wrong about someone's retirement is worse than no calculator at all.