Independent ML researcher. I pre-register experiments, publish code and raw results, and report what doesn't work as well as what does.
VersionOutcomes: can a model predict what a real code change does to a concrete call?
- The data: 1,973 executed before/after outcome pairs from 16 Python libraries.
- The finding: Gemma 4 12B spotted 31 of 42 real behaviour changes with two false alarms, but predicted the new result in only 19.
Cross-scroll ink detection for the Vesuvius Challenge (reading the Herculaneum scrolls):
- A scan-type conversion raises ROC-AUC on an unseen scroll by +0.024 across three seeds.
- Self-training lifts never-adapted regions by +0.018.
- The code is under review in the challenge codebase, and the checkpoints are on Hugging Face.
Luck, not efficiency: ARC Prize 2026 paper on where LLM agents lose points on ARC-AGI-3.
- Agents rarely waste moves. They fail to work out games, and whether they do is largely luck.
- A pre-registered 12-run test of a design-principles prompt was null: +2.5 points, 95% CI −3.6 to +8.7.
Dandelion: a pre-registered continual-learning research programme on a localized "petal" unit.
- It beat Gaussian RBF units by +2.04 balanced-accuracy points on 17 untouched datasets, under one shared training recipe.
- It showed no established advantage over tanh units.
Kaggle: kradle2grave · Hugging Face: TequilaYuen
Much of the implementation is done with AI coding agents under my direction; design, decisions and claims are mine.

