What the experiments showed.
A log of my research: the question each experiment asked, how it was run, what came out, and the limits on what it means. Every number comes from the project's own code and data.
-
Anomality: crash-proofing a volatility forecaster
The model's regime gate broke down in the COVID crash because the crash was far outside its training range. Feeding it volatility on a log scale fixed that, and the model then beat a simple baseline in 29 of 30 out-of-sample periods since 2019.
29 / 30 test periods since 2019 where the model's weekly-move forecast missed by less than the simple 'last month' estimate.
-
SA-HEACS: what server-side shielding costs
A discrete-event model of server-assisted execution. A 5 ms batching window cuts GPU time about eightfold, and prefetching two steps hides most of the network round trip.
~8× less GPU time per request with a 5 ms batching window, for about 6 ms more p95 latency.
-
Potholes, front wheels and e-scooter falls in Oxford
Three instrumented rides on Voi scooters and a MuJoCo model of scooter and rider. Jolts were frequent, and in the model only front wheels of 12 in or more cut falls.
100 sharp jolts (≥ 1 g) on 6.08 km of legal riding: about one every 61 m.
-
Priority of verification: crediting humans who check AI proofs
A proposal that the first person to verify an AI-generated result, to a stated standard, may publish it as author, with the AI source named in the abstract.
722 AI-generated manuscripts in OpenAI's October 2026 repository, in 372 families of results.
- Analysis of ROC for edge detectors (arXiv). ROC-based comparison of edge detection methods.
- OpenDAG. An open dataset of causal DAGs extracted from epidemiology papers, with missing confounders flagged and checked against each paper.