WRITING
Three papers built from original research, with methods, data, and limitations included.
Islands of Innovation: Defense Microelectronics Contracting and the Limits of Geographic Spillovers
This study tests whether defense microelectronics R&D raises semiconductor patenting by nearby private inventors who do not receive DoD funding. It links a CPC-filtered USPTO county-year panel to USAspending obligations and BEA, BLS, and NSF controls, then estimates two-way fixed-effects and distance-band models. Lagged obligations predict private patenting at a one-year horizon, but the effect dissipates. After direct funding is accounted for, proximity to a defense innovation hub is not statistically significant.
The Convergence Mirage: The Limits of Representation Similarity in From-Scratch Music Genre Classification
Eight model families were trained from scratch on FMA-Medium’s 16-genre task under a frozen five-seed protocol. SVM-RBF on classical DSP features reached the highest macro-F1 at 0.399, and no deep-architecture pair separated after Bonferroni correction. Mean penultimate-layer CKA across deep models was 0.685, compared with a 0.673 multi-initialization floor for untrained networks. The result shows why representation-similarity claims need untrained floors, same-architecture ceilings, and feature-group ablations.
Swap or Flop: Behavioral Interpretability of Switching in VGC-Bench Agents
This paper evaluates strategic switching in competitive Pokémon VGC doubles across five agents, four team conditions, and 20,000 logged battles. It uses switch-level logging, a 10-signal Strategic Switching Score, and an outcome-linked residual diagnostic. Behavior-cloned agents show the clearest interpretable strategy in distribution. A self-play PPO agent wins its fixed training matchup while averaging about 0.2 voluntary strategic switches per battle, showing that win rate alone does not establish learned strategy.
Essays & notes shelf: opens when there is something worth shelving.