Theendofplausible.
AI gives you an answer. StandardProofs gives you a verdict. We check every step, expose the assumptions nobody wrote down, and settle every checkable claim with deterministic solvers. Never a maybe.
Whyweexist
No doctor acts on an answer because it sounds right. No lawyer files one because it feels right. Yet the most consequential answers in the world now come from AI that can only promise one thing: plausible.
We built StandardProofs to end the era of plausible.
Watch it catch a mistake01
Formal math tools
Bulletproof, but they only work on problems already written as math. Almost nothing is.
02
Frontier AI
Works on everything, but it cannot tell you when it is wrong. A confidence score is a feeling with a decimal point.
03
StandardProofs
Sits in between. We take AI reasoning in plain language, check every step, and settle every checkable claim with a deterministic solver.
A real answer, checked live
Watchitcatchamistake
This is a real answer being checked, live. Flip the switch to sneak an error in, and watch it get caught.
Checking live
The answer, step by step
Problem: A 20 kg child is prescribed 300 mg of acetaminophen every 6 hours. Is the dose appropriate?
- 01
The child weighs 20 kg and is dosed every 6 hours
reason: given
- 02
Pediatric dosing is 10–15 mg/kg per dose, max 75 mg/kg per day
reason: dosing reference
- 03
10 × 20 = 200 mg and 15 × 20 = 300 mg, so the per-dose range is 200–300 mg
reason: computation
- 04
300 mg lies inside 200–300 mg
reason: comparison
- 05
4 doses × 300 mg = 1,200 mg per day, under the 75 × 20 = 1,500 mg daily cap
reason: computation
The check
○ checking step 01…
VerdictChecking…
Every claim needs a reason. Every step gets checked. Nothing enters silently.
01
Answer.
A frontier AI proposes an answer. We do not take its word for it.
proposed
05 · VerdictFull audit trail
Certifiedordeclined.Neveramaybe.
Every answer comes back certified or declined, with a full audit trail. No confidence scores. A decline names the step that failed.
Answer, unpack, check, prove, verdict.
Plain language in, machine-checked steps out
FromplainEnglishtomachine-checkedsteps
Most real problems do not arrive as math. StandardProofs meets them in plain language and hands every checkable step to a deterministic solver: proofs, calculations, facts pulled from sources. Everything else is pinned to its reasons.
python
Recomputed by SymPy
# finance · loan payment, step 04 >>> P, r, n = 50_000, Rational(6, 100) / 12, 60 >>> round(P * r / (1 - (1 + r) ** -n), 2) 966.64
Machine-checked
lean4
Proven in Lean
-- healthcare · dose within range, steps 03–05 theorem dose_in_range (kg dose : ℕ) (h₁ : kg = 20) (h₂ : dose = 300) : 10 * kg ≤ dose ∧ dose ≤ 15 * kg ∧ 4 * dose ≤ 75 * kg := by subst h₁ h₂; decide
Machine-checked
Measured, published, reproducible
Wepublishournumbers.
91.4%
HLE precisionHumanity's Last Exam, expert-graded97.1%
GPQA DiamondGraduate-level science, unseen+14%
More errors caught than an LLM judge94.7% vs 83.2% of planted errors5×
Decline signalDeclined answers are 5× likelier wrongNine of our certified answers on Humanity's Last Exam disagreed with the expert-written key. Independent review sided with us twice, the key once, and found the rest too close to call. We catch what the expert panels miss.
Full methodology and statistics
arXiv:2607.01223 ↗
Where wrong answers cost the most
Builtforteamsthatcannotaffordtobewrong
01
Healthcare
When AI touches a clinical decision, probably right is not a standard. We catch what does not hold up before it reaches a patient.
02
Legal and tax
Every conclusion arrives with its sources and steps attached. Challenge any line of it. That is what it is for.
03
Finance
Numbers that survive an audit, because they arrive with one.
04
Science and engineering
The mistakes that hurt are not bad arithmetic. They are silent assumptions. We make them visible.
Theledger
Today, verified answers. Next, verified work that runs for days: research programs, regulatory filings, engineering designs, drug candidates. Eventually, a norm: if a machine claims something about the world, it shows its work, or it does not get believed.
“If the shuttle could weave by itself and the plectrum could play by itself, masters would need no servants.”
Aristotle, Politics I.4, 1253b33
Bringustheanswersyoucannotaffordtogetwrong.
Early access is open. We are onboarding design partners in healthcare, legal, tax, and finance, and hiring founding engineers and researchers.
Not ready to start? Leave an address and we will send you the details.
Open by design
The paper, the code, and the benchmark results are open for anyone to inspect, reproduce, or attack.
- Paper · arXiv:2607.01223
- Code · github.com/standardproofs
- Benchmarks · medical, science, and reasoning exams