01 · Begin with the evidence
Exploration is not yet a claim.
AI can produce plausible analyses faster than science can validate them. The first task is to preserve what was actually observed—including weak, ordinary, incomplete and negative results.
Researcher · Tokyo
How should AI know
what it has earned?
I study the methods that turn AI-assisted exploration into scientific and mathematical claims proportionate to the evidence.
- 01 Evidence
- 02 AlphaScience
- 03 Calibration
- 04 Mathematics
- 05 Collaborate
02 · AlphaScience
A ledger before a manuscript.
AlphaScience routes exploratory trajectories into explicit evidence-ledger states before they become manuscript-facing claims. Human supervision remains where judgment changes what the evidence can support.
Visit the project site ↗03 · The calibration turn
Match the strength of the claim to the strength of the evidence.
Claim calibration is not an editorial cleanup step. It belongs inside the research loop: support, downgrade, redirect or stop—before fluency hardens uncertainty into narrative.
The Calibration Turn in AI-Assisted Research ↗04 · AI mathematics
Beyond answers and proof verification.
Real mathematical research also depends on method selection, reusable intermediate results, novelty and resource constraints. I study how a methodology given to AI changes the path—and the value—of its mathematical work.
05 · Work together
Build AI research systems that know their limits—and extend them.
I am a researcher at Institute of Science Tokyo and a visiting researcher at the University of Tokyo. I welcome conversations on AI for science, AI mathematics and evaluator design.