If Your Math Agent Can't Check Itself, It's a Poet

If Your Math Agent Can’t Check Itself, It’s a Poet

Most math “reasoning” demos are storytelling.

They sound convincing. They look sophisticated.

And then you do the one thing Olympiads—and reality—require:

you check the result.

If your math agent can’t do exact arithmetic, validate constraints, and run sanity checks…

it’s not a solver.

It’s a poet wearing a lab coat.

Math Olympiad “Reasoning” Isn’t the Secret. Checks Are.

🐉 The Reasoning Dragon (what it really is)

The “reasoning” failure mode isn’t that models can’t think.

It’s that they produce un-auditable chains with no witnesses:

• hidden assumptions
• sloppy arithmetic
• missing boundary cases
• no falsifiers
• no verification step

Humans get fooled by long explanations. Math doesn’t care.

The Stillwater rule

Reasoning must be checkable.

Not just plausible.

In practice, that means:

1. Exact computation where possible (integers/fractions, not floating vibes)
2. Multiple test cases (don’t trust one example)
3. Witness-first steps (intermediate claims you can verify)
4. Falsifiers (what would break this?)
5. Receipts (the checks you ran)

Poet vs Solver (the only split that matters)

Poet (looks smart)

• one-shot solution
• confident tone
• long narrative
• no tests
• no invariants

Solver (is correct)

• defines constraints clearly
• derives intermediate witnesses
• checks arithmetic exactly
• validates edge cases
• outputs a verdict + receipts

If you want reliable math, build a Solver.

A concrete template (copy/paste into any math workflow)

Use this every time:

1. Claim: what are we solving?
2. Given: list constraints + domain (integers? reals? nonnegative?)
3. Plan: the smallest method that could work
4. Witnesses: intermediate results you can verify
5. Checks: substitute back / boundary conditions / dimension sanity
6. Falsifiers: what would disprove this solution
7. Verdict: PASS / FAIL / NEED_INFO

This turns “reasoning” into something you can audit.

Why this wins on Olympiad-style tasks

Olympiad problems punish:

• missing cases
• arithmetic slips
• hand-wavy leaps
• “it seems” arguments

Checks catch all of those.

And checks are cheap.

That’s the punchline.

MrBeast-style challenge (participation loop)

Give me any hard problem you care about (contest / interview / real project).

I’ll reply with:

• a checks-first solution structure
• the exact witnesses to compute
• the sanity checks to run
• and what would falsify the answer

Comment CHECKS and I’ll drop a one-page “Witness-First + Checks” card you can reuse forever.

Tower placement (why this is Floor 3)

In the Stillwater Tower, this is Floor 3: Proving.

Because the goal isn’t to “sound like you reasoned.”

The goal is to produce an answer that survives a skeptic.

The point (one line)

Math agents don’t win by talking smarter. They win by checking harder.

— Phuc Vinh Truong