I conduct independent research on AI agents, evaluation, and representation-level interventions. Each project is labeled by its current status.
Claude Code kept the same secret word in all 194 referee checks across ten Hangman games when private working memory carried between turns, compared with 47 of 177 checks across ten otherwise identical games when that memory was deleted. This is a one-model pilot, not a general result about agents.
Agent memory · commitment · behavioral evaluation
Tests whether replacing a concept direction with Gaussian noise during fine-tuning can reduce a model’s reliance on a learned gender shortcut without retaining inference-time hooks. The experiment covers one synthetic task and one model; broader gender-bias reduction remains untested.
Concept ablation · shortcut learning · activation noise
Tests whether document tools preserve the evidence a model needs to support an answer. The experiment compares source documents with the text and metadata exposed to an agent.
Archive tools · evaluation · evidence
The pilot tested whether agents learn to cooperate when partners can refuse to work with them. One bounded training run reduced cooperation. The experiment did not show reliable learning or transfer, and the negative result now defines the next design question.
Multi-agent learning, cooperation, negative result