I point agents at open scientific problems and judge what comes back. Most of it does not survive my review.
- torux: a stellarator's coils in, its vacuum flux surface out, at max |B·n|/|B| of 10⁻¹⁵ in a median 4.2 s. A Proxima Fusion scientist wrote that it "verifies vacuum surfaces at 1e-15 in about a minute, which none of our field-line or Boozer-surface routes come close to".
- Zero-bulk metamaterials: the closest anyone has come to settling Graeme Milton's open question: a material that squeezes freely yet twists as stiffly as physics allows.
- Proxima Fusion's VMEC++ #819 (merged): 16 to 100x lower interpolation error in its vacuum-field tables.
- MACE #1806 (merged): magnetic MACE's energy made continuous at the cutoff.
- Predicting, Fast and Slow: an SSD already knows which of a learned controller's guesses are wrong. A 4,096-entry table of how its recent reads came back finds them (median AUROC 0.68) where nine uncertainty estimators do no better than a coin. Code. arXiv.
- Learner history and calibration: practice histories improved forecasts of pupils' maths performance (Brier score 0.2148 to 0.2027). NeurIPS 2026 MATH-AI workshop paper, poster.
- Atom Learning Model (ALM): the learning model behind a maths tutor used by 373 pupils for seven weeks, 26k+ marked attempts. arXiv.
- Optimal-transport cartogram: population maps without rotation versus diffusion's 7 to 9 degrees median on matched tests. Code. arXiv.



