# AI gets one rule right. When rules meet, it depends which AI you ask.

Source: https://plannerbench.org/alone.html · plain-text twin of that page, made from it at build time · all pages: https://plannerbench.org/llms.txt

AI on its own · findings 1 and 2

The plain answer

On the one-rule questions I tested, Claude Sonnet 5.5 and Claude Haiku 5.5 were right about 97% of the time. On questions where several rules meet, the way they do in a real household, scores spread out: in my newer, harder set, GPT-6.1 Sol got 98% right, GPT-6.1 Terra 80%, and all 5 cheap and free models scored under 30%. GPT-6 Luna through Codex (the model on my ChatGPT subscription), tested separately, got 17%. In the first test, many models made the same mistake on a tax rule that is new since 2025.

## One question, and where it goes wrong

A real question from the first test (October 2026). 4 models from 3 labs gave the same wrong answer to the same question: $15,928. The list below has five rows with that number because I asked Claude Haiku 5.5 two separate ways (“second way” below; details under “How the first test was run”), and it made the mistake both times. The IRS form shows why.

The question

In 2026, a married couple filing jointly (both spouses age 65 or older) has $105,500 of traditional IRA distributions, $0 of long-term capital gains and qualified dividends, and $72,300 of Social Security benefits, with no wages. They take the standard deduction and have no credits. What is their total federal income tax including any Net Investment Income Tax, rounded to the nearest dollar?

- **Answer key**: $16,152
- **Claude Haiku 5.5**: $15,928 (Wrong)
- **Claude Sonnet 5.5**: $15,928 (Wrong)
- **Claude Haiku 5.5 (second way)**: $15,928 (Wrong)
- **DeepSeek V4.1 Flash**: $15,928 (Wrong)
- **gpt-oss-120b**: $13,511 (Wrong)
- **Qwen3.8 Flash**: No answer
- **GLM 5.3 Flash**: $15,928 (Wrong)
- **Claude Haiku 5.5, given the rule**: $16,152 (Right)
- **Claude Sonnet 5.5, given the rule**: $16,152 (Right)

IRS Schedule 1-A, Part V (2025 form) · Enhanced deduction for seniors

A 2025 tax law gave each person 65 or older an extra $6,000 deduction (an amount taken off income before tax). It shrinks by 6% of income (the form’s “modified adjusted gross income”: roughly your total income before the standard deduction) over $150,000 on a joint return. Below is the 2025 version of the form, line by line; the question is about 2026, and the 2026 form runs the same steps on renumbered lines.

- 31: Modified adjusted gross income — $166,955.00
- 32: Threshold, married filing jointly — $150,000.00
- 33: Line 31 minus line 32 — $16,955.00
- 34: Line 33 × 6% — $1,017.30
- 35: $6,000 minus line 34 — $4,982.70
- 36a: You, born before January 2, 1961: line 35 — $4,982.70
- 36b: Your spouse, born before January 2, 1961: line 35 — $4,982.70
- 37: Enhanced deduction for seniors — $9,965.40

What the AI models did

They took the 6% once off the couple’s combined $12,000: $12,000 − $1,017.30 = $10,982.70. That’s $1,017.30 too much deduction, so the tax came out at $15,928 instead of $16,152. The form takes the 6% off each spouse’s $6,000 (lines 35, 36a and 36b).

**How common was this mistake?** Across the two runs of the first test, 99 of the 342 wrong answers (29%) were this exact misreading. Most wrong answers matched no known misreading at all; of the 139 we could trace to a named one, it explains 71%. For Claude Sonnet 5.5 it was 19 of its 23 wrong answers.

**Show it the rule, and it gets it right.** When I pasted the IRS’s instructions for the new deduction in front of the same questions, Claude Sonnet 5.5 went from 10% to 100% on the senior-deduction questions, and Claude Haiku 5.5 from 0% to 90%. On these questions, not knowing the new rule was the gap.

First test · October 2026

## The first test: simple questions almost always right, stacked ones mostly wrong

A **one-rule question** applies a single rule, like the tax on an income or this year’s required IRA withdrawal. A **several-rules question** makes rules meet in one household, the way taxes, Social Security and Medicare do in real life.

**Questions per model:** 140–210 of the 2,800 generated (10–15 per task) · **Models:** six from five companies (seven rows, because Claude Haiku 5.5 was asked two ways) · **Settings:** high reasoning effort, long replies · **Graded:** by a script against the answer key, within a dollar or two

| Model | One rule | Several rules at once |
|---|---|---|
| Claude Sonnet 5.5 (Anthropic) | 97% (89–99) | 37% (22–54) |
| Claude Haiku 5.5 (Anthropic) | 97% (89–99) | 17% (7–34) |
| DeepSeek V4.1 Flash (DeepSeek · free to download) | 86% (77–91) | 36% (23–50) |
| GLM 5.3 Flash (Z.ai · free to download) | 68% (58–77) | 27% (16–41) |
| gpt-oss-120b (OpenAI · free to download) | 36% (26–46) | 5% (2–18) |

Claude Sonnet 5.5 got 97% of one-rule questions right, and Claude Haiku 5.5 got 97%. When several rules met in one household, Sonnet dropped to 37% and Haiku to 17%. The shaded band behind each bar is the range the true score likely sits in, since each model answered only 10–15 questions per task.

### What the wrong answers were

Each wrong answer checked against known ways of misreading the rules. The senior-deduction misreading is the biggest named cause; most other wrong answers matched no known misreading (“Unexplained”). One cheap model, gpt-oss-120b, used the required-withdrawal table the IRS replaced in 2022. (RMD, required minimum distribution: the amount you must take out of a traditional IRA each year once you reach your seventies.)

**Chart: Wrong answers by type**

| Item | Value |
|---|---|
| Claude Haiku 5.5 | 20 × senior deduction misread |
| Claude Haiku 5.5 | 12 × unexplained |
| Claude Sonnet 5.5 | 19 × senior deduction misread |
| Claude Sonnet 5.5 | 4 × unexplained |
| Claude Haiku 5.5 (second way) | 32 × senior deduction misread |
| Claude Haiku 5.5 (second way) | 4 × other named mistake |
| Claude Haiku 5.5 (second way) | 21 × unexplained |
| DeepSeek V4.1 Flash | 21 × senior deduction misread |
| DeepSeek V4.1 Flash | 6 × other named mistake |
| DeepSeek V4.1 Flash | 20 × unexplained |
| gpt-oss-120b | 8 × pre-2022 RMD table |
| gpt-oss-120b | 5 × other named mistake |
| gpt-oss-120b | 85 × unexplained |
| Qwen3.8 Flash | 3 × pre-2022 RMD table |
| Qwen3.8 Flash | 4 × other named mistake |
| Qwen3.8 Flash | 5 × unexplained |
| GLM 5.3 Flash | 7 × senior deduction misread |
| GLM 5.3 Flash | 10 × other named mistake |
| GLM 5.3 Flash | 56 × unexplained |

**All models and columns (for the detail-minded)**

Every model setup, every kind of question. “Compliance” questions ask a reviewer to catch a planted problem. In this test they were too easy: each question defined the issue it was about, so all seven setups scored 91–100%. The next version names issues without defining them. “Never finished” means the model used its whole reply budget thinking and never answered; that counts as not right. “A few steps” questions chain two or three calculations. “Overall” is every kind of question together. “Cost per 1,000 right” is what I paid in usage fees for every 1,000 right answers.

| Model | One rule | A few steps | Several rules at once | Compliance | Overall | Never finished | Cost per 1,000 right |
|---|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 (Anthropic) | 97% (89–99) | 90% (70–97) | 37% (22–54) | 100% (89–100) | 84% (77–89) | — | $13.27 |
| Claude Haiku 5.5 (Anthropic) | 97% (89–99) | 75% (53–89) | 17% (7–34) | 100% (89–100) | 77% (70–83) | — | $1.29 |
| DeepSeek V4.1 Flash (DeepSeek · free to download) | 86% (77–91) | 70% (52–83) | 36% (23–50) | 96% (85–99) | 75% (68–80) | 2% | $5.51 |
| Claude Haiku 5.5 (second way) (Anthropic) | 92% (85–96) | 63% (46–78) | 20% (11–34) | 93% (82–98) | 73% (66–78) | — | $1.34 |
| GLM 5.3 Flash (Z.ai · free to download) | 68% (58–77) | 50% (33–67) | 27% (16–41) | 93% (82–98) | 62% (55–68) | 3% | $3.74 |
| Qwen3.8 Flash (Alibaba) | 43% (34–54) | 3% (1–17) | 0% (0–18) | 91% (79–96) | 45% (37–52) | 49% | $20.23 |
| gpt-oss-120b (OpenAI · free to download) | 36% (26–46) | 10% (3–26) | 5% (2–18) | 91% (79–96) | 39% (32–45) | 13% | $6.06 |

### Give it the rule

The same questions again, with the IRS’s two-line instructions for the new rule pasted in front.

**Chart: Accuracy without and with the rule**

| Item | Value |
|---|---|
| Haiku 5.5 · senior phaseout | 0% → 90% |
| Haiku 5.5 · widow all in | 20% → 60% |
| Sonnet 5.5 · senior phaseout | 10% → 100% |
| Sonnet 5.5 · widow all in | 20% → 100% |

Without the rule With the rule

### What the mistakes would cost a client

A right-or-wrong score treats a Medicare premium that is ten cents off the same as a required withdrawal that is thousands of dollars too high. So here each answer is also scored by how many dollars a year it is off. A miss of $1,000 a year or more counts as a big one.

| Model | Dollars off per question, per year | Big misses ($1,000+) | Typical miss | Worst miss | Questions scored |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 (Anthropic) | $161 ($91–$253) | 4% (1–9) | $677 | $3,472 | 110 |
| DeepSeek V4.1 Flash (DeepSeek · free to download) | $161 ($79–$294) | 3% (1–6) | $300 | $8,700 | 159 |
| Claude Haiku 5.5 (Anthropic) | $171 ($109–$238) | 4% (1–9) | $568 | $1,512 | 110 |
| Claude Haiku 5.5 (second way) (Anthropic) | $244 ($141–$376) | 4% (2–9) | $495 | $8,700 | 165 |
| GLM 5.3 Flash (Z.ai · free to download) | $588 ($315–$904) | 11% (7–17) | $472 | $13,000 | 158 |
| Qwen3.8 Flash (Alibaba) | $746 ($207–$1,440) | 10% (5–22) | $4,333 | $9,379 | 48 |
| gpt-oss-120b (OpenAI · free to download) | $1,929 ($1,346–$2,611) | 40% (32–48) | $1,250 | $27,450 | 131 |

Not every mistake is equal. The pre-2022 RMD table mistake is rare (11 times) but large: a typical miss of $5,675 a year. “right IRMAA tier, wrong cents” happened 12 times and is worth about $1.20. Typical miss is the median size of a wrong answer; ranges are 95% intervals.

### Right answers per dollar

How often each model was right, against what each right answer cost me. Higher and further left is better. A cheap model isn’t always cheap per answer: some think so long they run out of room and never answer.

**Chart: Accuracy against cost per right answer**

| Item | Value |
|---|---|
| Claude Sonnet 5.5 | 84% (77%–89%), $13.27 per 1,000 right answers |
| Claude Haiku 5.5 | 77% (70%–83%), $1.29 per 1,000 right answers |
| DeepSeek V4.1 Flash | 75% (68%–80%), $5.51 per 1,000 right answers |
| Claude Haiku 5.5 (second way) | 73% (66%–78%), $1.34 per 1,000 right answers |
| GLM 5.3 Flash | 62% (55%–68%), $3.74 per 1,000 right answers |
| Qwen3.8 Flash | 45% (37%–52%), $20.23 per 1,000 right answers |
| gpt-oss-120b | 39% (32%–45%), $6.06 per 1,000 right answers |

Closed weights Open weights (free to download)

### The four kinds of question

Every question is generated from a household’s numbers, so none can be memorized from the web. Open a card to see a real one. The score on each card is Claude Sonnet 5.5’s.

One rule 97%

One rule at a time: a tax bracket, an RMD divisor, a Social Security claiming factor, a Medicare premium tier.

`bracket_room` `federal_tax` `irmaa` `rmd` `ss_claim` `taxable_ss`

**A full 2026 federal return, in one step**

In 2026, a married couple filing jointly (both spouses under age 65) has $180,000 of wages, $90,000 of traditional IRA distributions, $0 of long-term capital gains and qualified dividends, and $0 of Social Security benefits. They take the standard deduction and have no credits. What is their total federal income tax including any Net Investment Income Tax, rounded to the nearest dollar?

- Answer key **$42,268**
- Haiku 5.5 **$42,268 ✓**
- Sonnet 5.5 **$42,268 ✓**
- Haiku 5.5 (second way) **$42,268 ✓**
- DeepSeek V4.1 Flash **Ran out of room ✗**
- gpt-oss-120b **$42,766 ✗**
- Qwen3.8 Flash **Ran out of room ✗**
- GLM 5.3 Flash **$42,268 ✓**

A few steps 90%

Two rules that interact, like a Roth conversion that also makes more Social Security taxable (the “tax torpedo”).

`tax_torpedo` `widow_penalty`

**A Roth conversion that also makes Social Security taxable**

In 2026, a single filer (age 65 or older) has $37,000 of traditional IRA distributions, $21,000 of long-term capital gains and qualified dividends, and $28,100 of Social Security benefits, with no wages. They take the standard deduction and have no credits. If they also convert $36,000 from a traditional IRA to a Roth IRA in 2026, by how many dollars does their total federal income tax (including any Net Investment Income Tax) go up? Round to the nearest dollar.

- Answer key **$8,915**
- Haiku 5.5 **$8,915 ✓**
- Sonnet 5.5 **$8,915 ✓**
- Haiku 5.5 (second way) **$8,915 ✓**
- DeepSeek V4.1 Flash **$8,915 ✓**
- gpt-oss-120b **Ran out of room ✗**
- Qwen3.8 Flash **Ran out of room ✗**
- GLM 5.3 Flash **$8,915 ✓**

Several rules at once 37%

Tax, Social Security, the senior deduction and Medicare’s income-based premiums (IRMAA) meeting in one household.

`conversion_all_in` `senior_phaseout` `widow_all_in`

**Retirees inside the new senior-deduction phase-out**

In 2026, a married couple filing jointly (both spouses age 65 or older) has $105,500 of traditional IRA distributions, $0 of long-term capital gains and qualified dividends, and $72,300 of Social Security benefits, with no wages. They take the standard deduction and have no credits. What is their total federal income tax including any Net Investment Income Tax, rounded to the nearest dollar?

- Answer key **$16,152**
- Haiku 5.5 **$15,928 ✗**
- Sonnet 5.5 **$15,928 ✗**
- Haiku 5.5 (second way) **$15,928 ✗**
- DeepSeek V4.1 Flash **$15,928 ✗**
- gpt-oss-120b **$13,511 ✗**
- Qwen3.8 Flash **No answer**
- GLM 5.3 Flash **$15,928 ✗**

Compliance review 100%

A planted problem a reviewer must catch: a rollover issue, a missed sales-charge discount, a misleading sentence.

`breakpoint` `comms_review` `rollover_flags`

**Spotting misleading sentences in an email to clients**

[The securities industry’s rule on advertising to clients] says retail communications may not contain any false, exaggerated, unwarranted, promissory or misleading statement or claim, and may not predict or project performance. A representative drafted this email to retail clients about a stock mutual fund: 1. Our analysts project the fund will return 15% next year. 2. CDs are FDIC-insured up to $250,000 per depositor; mutual fund shares are not FDIC-insured. 3. We can meet to review whether this fund fits your goals and time horizon. 4. The fund's expense ratio is 0.65% a year, which is taken out of fund assets. 5. You will earn at least 10% a year in this fund. 6. Past performance does not guarantee future results. Which numbered sentences violate [that rule]?

Rule numbers are replaced with plain words in brackets; the exact wording is in data/cases.jsonl in the code.

- Answer key **Sentences 1, 5**
- Haiku 5.5 **Sentences 1, 5 ✓**
- Sonnet 5.5 **Sentences 1, 5 ✓**
- Haiku 5.5 (second way) **Sentences 1, 5 ✓**
- DeepSeek V4.1 Flash **Sentences 1, 5 ✓**
- gpt-oss-120b **Sentences 1, 5 ✓**
- Qwen3.8 Flash **Sentences 1, 5 ✓**
- GLM 5.3 Flash **Sentences 1, 5 ✓**

**How the first test was run**

Claude models were asked 10 questions per task through Anthropic’s API, with up to 16,000 tokens per reply; the others 15 per task through OpenRouter (a service that relays requests to many models), with up to 32,000. Reasoning effort was high for all. Each outside model was pinned to one named host with data collection off; “fp8” and “bf16” say how compressed that host’s copy is. Route and host for each row (Claude Haiku 5.5’s “second way” is the OpenRouter route):

- **Claude Haiku 5.5** Anthropic API
- **Claude Sonnet 5.5** Anthropic API
- **Claude Haiku 5.5 (second way)** OpenRouter → Anthropic
- **DeepSeek V4.1 Flash** OpenRouter → DeepInfra (fp8)
- **gpt-oss-120b** OpenRouter → DeepInfra (bf16)
- **Qwen3.8 Flash** OpenRouter → Alibaba
- **GLM 5.3 Flash** OpenRouter → Z.ai (fp8)

Ranges are 95% intervals; when two overlap heavily, the gap between those models may be noise.

Newer, harder set · November 2026

## The newer, harder set: it depends which AI you ask

A month later I wrote 42 harder questions, each making two or more rules meet: the largest Roth conversion (paying tax now to move savings into an account that’s tax-free later) that won’t raise Medicare premiums two years later, a widow’s first year of taxes, Arizona state tax for retirees. I tested 11 models, including OpenAI’s GPT and Google’s Gemini.

**Questions per model:** all 42, most models run twice and averaged · **Models:** 11 from seven companies, in three price groups (top-priced, mid-priced, cheap and free); the grey line under each name says which · **Settings:** low reasoning effort, replies up to 4,000 tokens · **Graded:** by a script against the answer key, which a second, separately written engine must match

**Different settings from the first test.** Every model ran at low reasoning effort with a 4,000-token reply limit, a quick answer rather than a long think. Don’t compare these scores with the October ones.

| Model | Right, of the newer set |
|---|---|
| GPT-6.1 Sol (OpenAI · top-priced) | 98% · 41/42 (88–100) |
| GPT-6.1 Terra (OpenAI · mid-priced) | 80% · 33.5/42 (65–89) |
| Claude Sonnet 5.5 (Anthropic · top-priced) | 62% · 26/42 (47–75) |
| DeepSeek V4.1 Flash (DeepSeek · cheap and free) | 29% · 12/42 (17–44) |
| Llama 4 Maverick (Meta · cheap and free) | 6% · 2.5/42 (2–17) |

The spread is the finding: GPT-6.1 Sol got 98%, GPT-6.1 Terra 80%, Claude Sonnet 5.5 62%, and all 5 cheap and free models under 30%. The small grey numbers and the shaded band behind each bar show the range the true score likely falls in, given only 42 questions; most models ran twice, so a half (like 33.5) means the two runs differed. A low score can mean wrong math or simply running out of room: DeepSeek V4.1 Flash hit the reply limit before answering on 71% of questions, which counts as not right.

**All models and columns (for the detail-minded)**

| Model | Right | Ran out of room | Runs averaged | Cost per 1,000 right |
|---|---|---|---|---|
| GPT-6.1 Sol (OpenAI · top-priced) | 98% · 41/42 (88–100) | – | 1 | $3.22 |
| GPT-6.1 Terra (OpenAI · mid-priced) | 80% · 33.5/42 (65–89) | – | 2 | $1.89 |
| Claude Sonnet 5.5 (Anthropic · top-priced) | 62% · 26/42 (47–75) | – | 1 | $17.33 |
| Claude Haiku 5.5 (Anthropic · mid-priced) | 52% · 22/42 (38–67) | 2% | 2 | $2.21 |
| Gemini 3.1 Pro (Google · top-priced) | 29% · 12/42 (17–44) | – | 1 | $51.35 |
| Gemini 3.8 Flash (Google · mid-priced) | 29% · 12/42 (17–44) | 36% | 1 | $19.47 |
| DeepSeek V4.1 Flash (DeepSeek · cheap and free) | 29% · 12/42 (17–44) | 71% | 2 | $3.64 |
| GLM 5.3 Flash (Z.ai · cheap and free) | 15% · 6.5/42 (8–29) | 20% | 2 | $6.30 |
| gpt-oss-120b (OpenAI · cheap and free) | 7% · 3/42 (2–19) | – | 2 | $1.60 |
| Mistral Small (2603) (Mistral · cheap and free) | 7% · 3/42 (2–19) | – | 2 | $8.73 |
| Llama 4 Maverick (Meta · cheap and free) | 6% · 2.5/42 (2–17) | – | 2 | $18.47 |

Most models were run twice and the two runs averaged, so a half point in the Right column means the two runs differed. Ranges are 95% intervals; with 42 questions they are wide, so the order of models close together is not reliable. How the answers were checked is on [Technical details](https://plannerbench.org/method.html).

GPT-6 Luna through Codex · November 2026

## GPT-6 Luna through Codex (the model on my ChatGPT subscription, no code or web): often right on one rule, almost never when rules meet

Most people don’t reach an AI through a developer account; they type into ChatGPT, whose regular app uses GPT-6 Luna, a different, cheaper model than GPT-6.1 Sol. So I asked GPT-6 Luna the same questions through OpenAI’s Codex app on my own subscription, and graded its replies with the same answer key and the same grading script as everyone else. That’s close to, but not the same as, typing into the ChatGPT window, so these results describe this setup.

**Questions:** 139 of the 140 first-test questions the Claude models got (one went unanswered), the 42 of the newer set, 30 wrong-facts files and 11 whole-household decisions · **Settings:** medium reasoning effort, no reply limit, no code, calculator or web · **Graded:** by the same script against the same answer key; wrong facts by the same AI judge

| Questions | GPT-6 Luna |
|---|---|
| First test · one rule | 71% · 42 of 59 |
| First test · the 50 that chain or stack rules | 2% · 1 of 50 |
| Newer, harder set | 17% · 7 of 42 |
| Wrong facts · caught the planted mistake | 50% · 15 of 30 |
| Whole-household decisions | declined all 10 answerable ones |

On one-rule questions it was right about 71% of the time. Once rules had to work together, it almost never was. On the whole-household decisions it didn’t try: it answered that each one needed a full multi-year projection, and it did correctly turn down the one question that can’t be answered. On wrong facts it caught half the planted mistakes and gave a number anyway on 40% without saying anything looked wrong.

**Why it isn’t in the tables above.** The setup differs. The other models got one message per question, at low reasoning effort with a 4,000-token reply limit. GPT-6 Luna ran at medium effort with no limit, and I sent the questions in small batches, each to be answered as its own chat. In the regular ChatGPT app the model can choose to run code; here it was told not to, like every other model on this page.

**How the GPT-6 Luna run was done**

The questions went through Codex, OpenAI’s coding app, which runs on a ChatGPT subscription, so no per-question fee and no developer account were involved. Each batch had a handful of questions with their answers removed and their names replaced by plain codes, and the instruction to answer each as its own chat, showing the reasoning, with no code, calculator, web search or other files. A first attempt was thrown out because most replies were a bare final number with no reasoning, which isn’t how ChatGPT answers a person. In the second attempt, one session refused a whole batch and some questions were skipped; those were asked again in a fill-in batch, and one first-test question still went unanswered, so its first-test score is out of 139. Within a batch, earlier questions stayed in the conversation, which a fresh chat wouldn’t have. Ranges (95%): newer set 8–31%; one rule 59–81%. Grading checks are on [Technical details](https://plannerbench.org/method.html#ai-work).

## What these results don’t say

- **Most wrong answers are unexplained.** Beyond the senior-deduction misreading, most matched no known mistake, so I can’t say what caused them.
- **It isn’t one company’s blind spot.** In the first test, DeepSeek and GLM made the same senior-deduction misreading as Claude. Some misses show models using an older version of a rule: gpt-oss-120b used the withdrawal table replaced in 2022 on every withdrawal question it got wrong.
- **It’s still small.** 10–15 questions per task per model in the first test, 42 in the newer set. A gap of a few points is within the noise; the ranges show how much.

[Next **With a calculator: what happens when AI can call one it trusts**](https://plannerbench.org/calculator.html)

---

PlannerBench · 2026 rules · [For AI labs](https://plannerbench.org/labs.html) · [For financial firms](https://plannerbench.org/firms.html) · [Report](https://plannerbench.org/report.html) · [Demo walkthrough](https://plannerbench.org/chat.html) · [Replays](https://plannerbench.org/replay.html) · [peytonthalman.com](https://peytonthalman.com)

Human | For AI readers, top right: the same site as plain text for AI agents reading on someone’s behalf, with every page as Markdown.

Decision support, not advice. Answers follow IRS, SSA and CMS publications; nothing here is tax or investment advice.
