a 29,665-parameter decoder-only transformer implemented and trained entirely in mit scratch 3.0
a cycle-accurate cpu emulator in mit scratch 3.0 supporting subleq, chip-8, and a full mos 6502 core
in-browser local slm calibration and entropy-based safety gating via onnx runtime web & wasm
no single signal reliably separated real agreement from parallel pricing among LLM sellers.
tests whether auditors can tell real agreement between llm sellers from parallel pricing, using no-channel markets as decoys. planted agreements raised the collusion index well above baseline, but no single signal reliably separated agreement from parallel conduct.
open-weight models rarely refuse forensic work but silently fail on 17.3% of prompts.
tests whether self-hostable open-weight models can take over forensic analysis that hosted models refuse. across 2,758 graded prompts on two models, refusals were rare (0.15%) but 17.3% of prompts got silent non-answers, and partial redaction gave the best effective yield (60.2% against 38.6% for raw artifacts).
a model's own distress is represented separately from distress attributed to others.
welfare evaluations of language models often confuse internal distress directions with the sentiment of the text or with distress attributed to others. varying only the referent (the model itself, another model, or a fictional android), self-referential and other-referential distress representations separate across six open-weight models, and self-reported valence tracks the self-specific component.
open to internships and collaborations,
