negative-result

Companion code for the MiLeTS 2026 paper "When Does Adversarial Refinement Help?" — adapting R3GAN (NeurIPS 2024) to multivariate time series imputation. A clearly-scoped negative result and open problem.

Negative result paper: the King Wen sequence has genuine anti-habituation statistical properties (confirmed via Monte Carlo against 100k baselines) but does not improve neural network training. Experiments on NVIDIA RTX 2060 (PyTorch) and Apple Silicon (MLX).

Pre-registered study: can MoE routers be trained for cache locality at no accuracy cost? A negative result at 137M scale, plus edge memory-bandwidth-wall measurements. Single RTX 3070.

A tool that tried to prove vulnerabilities were absent from shipped binaries. It does not work. Published in full: paper, corpus, and the audit trail.

A Windows 11 Bluetooth LE Audio/LC3 driver project targeting Qualcomm on x86-64 — with an explanation of why it can't work without an OEM or reverse-engineered driver for the hardware layer. Plus a tool to check LE Audio support on your PC. Not an installable driver.

Algorithmic study of sub-fp16 KV cache compression on shared-memory mobile SoCs. Documented negative result: scalar sign+magnitude (incl. outlier extraction + 2-bit) insufficient vs KIVI. Cos sim shown unreliable for KV-quant PPL.

Measuring Exa's maxAgeHours before building a freshness router. Cost is flat, content is identical, and the router should not exist.

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

Base-rate-honest benchmark showing LoRA-adapted TimesFM has no directional skill on equities: code, frozen data, committed results, and paper.

No-go for scalar free-fraction closure of the local capacity-metric bridge; LC-O2-O1 remains open.

Connectivity Contrast Index (CCI) and 17 alternative metrics across five EEG datasets — an honest non-replication report

Attempted dense-to-MoE conversion of Qwen 3.5 9B (DeltaNet hybrid) using CMoE and D2DMoE. Documents why post-hoc MoEfication fails on SwiGLU models without extensive sparsification. Negative result.

Claim-bounded failure-boundary audit of an A10-derived Hubble-tension scaffold: no physical E(z)/H(z) bridge, no likelihood comparison, and no solution claim.

Archived negative result: measuring whether routing Claude Code prompts to free models saves tokens. It doesn't — three features, three measured failures.

MoGRU: Momentum-Gated Recurrent Unit — When Second-Order Dynamics Help and When They Don't (negative-result research)

Preregistered Stage 0 report: automatic memory distillation and DEV task calibration for a local LLM code optimization harness. Negative results with mechanism, frozen gates, no test tasks opened.