"Publication bias and the canonization of false facts" published in eLife (2016)

Reproducible evaluation harness for hidden coordination variables in multi-agent LLM systems.

🔗 Explore Graph-Tabular Fusion to enhance Bitcoin fraud detection, revealing that rich tabular features render graph embeddings redundant.

Graph-Tabular Fusion for Bitcoin Fraud Detection - Demonstrating when Node2Vec embeddings don't improve XGBoost. Scientifically rigorous negative result validating that tabular features encode graph structure.

VSA-keyed memory layers for transformer FFN replacement — structured codebook keys via FHRR bind

Pre-registered, six-phase study of retail algorithmic trading on Binance Spot (2017-2026). Negative result: no simple strategy beats Buy-and-Hold BTC out-of-sample. Simulators, data fetchers, and paper (EN+ES).

Neuro-symbolic AI that beats end-to-end baselines on structured manipulation (Towers of Hanoi, 8-puzzle). 100% success validated with Wilson 95% CIs. MIT license.

Frozen reproducibility study of a seven-month attempt at ML-driven cryptocurrency market selection — fifteen architecture versions, ~170 experiments, terminal finding that nothing survives OOS validation. MIT.

Content-aware adaptive sampling for Stable Diffusion - dynamically optimizing inference steps based on prompt complexity for 30-50% speedup with <5% quality loss.

Why your multi-agent LLM deception experiment might be measuring prompt comprehension. An honest post-mortem.

A cost-sensitive Human-in-the-Loop framework for AML transaction monitoring — empirically demonstrating when human review increases operational cost under asymmetric error regimes.

Research prototype: category-aware layer skipping for small LLMs on Apple Silicon. Honest write-up of what worked (1.9x batch throughput on Llama 3.2 3B) and what didn't (MMLU collapse). Outperformed by LayerSkip/MoD/vLLM.