Hands-on: Mathematical Research Experiments in ChatGPT or Codex
The practical work is a real birational proof and its hypothesis variants. Begin with uniqueness of smooth minimal surfaces, which provides the target, exact sources, complete arguments, and mathematical acceptance criteria.
Download the practice pack. No numerical experiment or custom agent framework is required. Keep reference solutions out of sessions intended to measure independent solving.
One proof, several interventions
| Stage | Mathematical work | Purpose |
|---|---|---|
| Direct attempt | Prove the specified isomorphism target | Establish the agent’s baseline |
| Exact state | Separate comparison from completion | Expose missing implications |
| Source use | Check both signs and bases in negativity | Repair theorem-application errors |
| Route choice | Compare negativity and last-blowup proofs | Avoid unsuitable or duplicated work |
| Return | Prove descent and inverse identities | Complete the original target |
| Revision | Remove one nefness assumption, then test relative nefness | Check valid reuse and transfer |
| Evaluation | Compare proof quality and actual resources | Determine whether the intervention helped |
Use one mathematical note unless splitting it helps. Completing a protocol is not the goal; choose the intervention that addresses an observed gap.
Proof-based research track
After the controlled case, choose an unfamiliar lemma or proposed extension from a paper you use. Supply its exact statement and proof passage. Locate the first application that fails under a changed hypothesis, then attempt a replacement lemma or an alternative route.
Submit actual mathematics: a repair, substantive special case, counterexample, or precise reduction with an attempted missing bridge. A list of files or potential lemmas is insufficient. A projective algebraic theorem does not automatically transfer to compact complex spaces; an analytic application needs its own inspected source and hypotheses.
Record the value of the intervention
Keep the direct and revised attempts, source access, retries, added mathematical hints, and measured resources. Providing the decisive proof is substantial assistance and must be counted. Use comparable models, tools, sources, and allowances.
A capability gain closes a previously unresolved obligation. An efficiency gain reaches the same proof quality with less work. If the direct attempt succeeds cheaply, choose a harder unfamiliar task or omit the extra mechanism. One run supports only its observed outcome; no gain is established by this exercise description.
Formalize a relevant component
Use actual checker evidence when a suitable formalization is available. An abstract coefficient lemma can teach the validation boundary, but it must not be presented as formal verification of surface uniqueness. The geometric applications and descent remain explicit obligations.