Problems and proofs from working mathematicians.
Reasoning models are measured on mathematics, and most public problems are already in their training data. We commission mathematicians to write new ones and grade the reasoning.
Mathematics, with proof.
Mathematical reasoning models are increasingly good at reaching the right answer by the wrong route: an unjustified step, a case left unchecked, a lemma assumed rather than proved, or circular reasoning hidden in a long argument. Meanwhile, competition and textbook problems are so widely published that scores on them say little. HUMXN commissions credential-checked mathematicians to write original problems with full solutions, check proofs and grade model reasoning step by step.
Every problem is newly written to your specification, fingerprinted at intake and checked for duplicates and near-duplicates against the dataset, then verified by a second mathematician who works the solution independently. Label provenance records whether each grade came from the author, a model suggestion or an expert reviewer, and each delivery carries a signed receipt. OpenTimestamps anchoring can show when a held-out set existed, which helps when contamination is questioned later.
For mathematicians, this is paid, remote work that rewards exactly what you trained for: posing problems with a single clean answer, writing rigorous solutions and finding the gap in an argument. Rates are shown before you accept each job, and you choose work that fits around research or teaching. Your identity is not published, and approved work is paid automatically through Stripe after identity and tax verification.
Built to your specification.
Every engagement starts from a written spec and a pilot batch. These are the most common requests we source for mathematics.
Original problems with solutions
Graduate and research-level problems with verified answers and complete worked solutions, tagged by field and difficulty.
Proof writing and checking
Rigorous proofs written by mathematicians, and model proofs checked line by line for gaps, errors and unjustified steps.
Step-level reasoning grades
Each step of a model's solution marked correct, incorrect or unjustified, with the first error identified.
Formal mathematics
Statements and proofs in proof assistants such as Lean, written or checked by mathematicians who use them.
Held-out evaluation sets
Private benchmarks of new problems, fingerprinted and timestamped, built so results reflect reasoning rather than recall.
Preference judgments
Comparisons of model solutions on correctness, rigor and clarity, with written rationales from mathematicians.
01Benchmark contamination
Published problems leak into training data. New, fingerprinted and timestamped problems give results you can defend.
02Right answer, wrong reasoning
Final-answer checks miss flawed arguments. Step-level grading by mathematicians catches where a proof breaks.
03Ambiguous problems
A problem with two defensible answers is a bad label. Independent second solving catches ambiguity before delivery.
Paid work in your field.
Remote and flexible, with the rate stated before you accept. Every job is reviewed by a second expert, and approved work is paid automatically.
- A PhD in mathematics, or doctoral study well underway
- A master's degree with substantial research or teaching experience
- Competition success, such as olympiad medals or Putnam results
- Publications or teaching in a specific field, such as algebra, analysis, combinatorics or topology
- Fluency in LaTeX, and in Lean or another proof assistant for formal work
Write graduate-level problems
Pose original problems with a single verifiable answer and a complete worked solution.
Check proofs
Read model or human proofs line by line and identify the first gap or error.
Grade model reasoning
Mark each step of a solution against a rubric and explain each judgment.
Solve another expert's problems
Work a colleague's problem independently as second reviewer to confirm the answer and wording.
Mathematics: asked often.
How do you keep evaluation problems out of training data?
Problems are written new for your request, fingerprinted at intake and checked for duplicates and near-duplicates. They are delivered only to you, and with an exclusive licence they are not offered to anyone else. OpenTimestamps anchoring records when the set existed, and signed provenance ties each problem to its author and reviewer.
What level of mathematics do you cover?
Requests range from advanced undergraduate through graduate and research level, across algebra, analysis, number theory, combinatorics, geometry, topology, probability and more. Each request is matched to mathematicians verified in that field, and every problem is solved independently by a second expert before delivery.
Do I need a PhD to join?
Not always. Many jobs ask for a PhD or doctoral study in progress, but strong competition records or a master's degree with research or teaching experience can qualify you for others. We verify identity and credentials, and may ask for a short, paid assessment problem in your field.
How do I get paid?
Rates are set by field and task and are shown before you accept. Once a second mathematician approves your work, it is credited to your balance and paid automatically through Stripe after your identity and tax details are verified. Each job appears on an itemized ledger.
How much time does it take?
There is no minimum commitment. Opportunities are offered when they match your field, each with a stated scope and deadline, and you decide which to take. Many mathematicians fit the work around research and teaching, writing problems in focused sessions rather than on a fixed schedule.
Can we commission formal proofs in Lean?
Yes, where we can match the request to mathematicians experienced with the proof assistant. Formal statements and proofs are checked by the tool and reviewed by a second expert for faithfulness to the informal statement. Deliveries come as JSONL, Parquet or Croissant, with source files and a signed receipt.
Mathematics, done by experts.
Need expert-made data in this field, or have the expertise to make it? Start here.