negative-results

llama.cpp speculative decoding measured on one RTX 3090, Qwen3.6-35B-A3B UD-Q4_K_XL, commit 3737e4137. Published figures are re-derived from the committed data by a checker that fails on drift, a coverage probe reports how many of them it actually covers, and ERRATA.md lists this study's own retracted claims.

Weights that never freeze. A continual-learning architecture, and the pre-registered experiment that falsified its central claim. https://arxiv.org/abs/2607.20792

A negative result on joint-embedding predictive architectures for prediction markets. The martingale-collapse diagnostic is sound and predicts nothing, more training makes the representation worse, and the metric everything was ranked by was mostly measuring input reconstruction. Pre-registered gates, full findings record.

Policy-Gated Governance for Operational AI Agents

JEPA agent playing Minecraft from pixels: latent world model + MPC planning, 664K params on one 8GB GPU, trained on raw gameplay with no labels. A complete lab notebook - including a 20-attempt research dead end, documented with its root cause.

One fixed 9-stage QSAR and conformal-prediction drug-discovery pipeline run unchanged across ENPP1, NLRP3, TYK2 and IRAK4 — its most valuable outputs are the two refusals.

A reproducible benchmark: do AI coding agents violate architectural import-boundary rules inferred from real codebases? Harness, pre-registration, data, and a null result.

[zenodo.20574533] CGAA: Concept Guided Adversarial Attacks

Distribution-shift teardown of openpilot v0.9.7 supercombo: does a production L2 self-driving model know when it's blind? (it doesn't, and silently)

Resample or reroute after a weak-verifier stop? Pre-registered measurements of recoverable stopping debt on MBPP+, a two-sided action-support gate on BigCodeBench that fails closed, a LiveCodeBench observability ladder, and the exchangeable-actions reference showing realized-maximum gaps carry no selector signal. Artifacts for arXiv:2607.08665v3.

A global, open-source registry and standardized schema for null findings, non-significant outcomes, and failed trials in medical research. Built to eliminate publication bias and accelerate biomedical discovery.

从20个项目中系统提取可复用方法论模式的实验记录。包含10轮正式审查实证（4后端）、58项发现、G5可追溯审计。不成熟框架，诚实的实验记录。

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

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

Does cache-aware MoE routing degrade generation, and would the usual metrics notice? 432 pre-registered generations. Every number traced to a result file.

A $0 falsification lab across two markets — crypto (~111 hypotheses) and Polymarket prediction markets (172,830 resolved markets, 1.36M trades). 184+ techniques through one committed anti-overfitting gauntlet. 0 survive to a deployable edge; the reusable validation harness is the asset. Agent-ready. MIT.

Measurement-first research on local Mixture-of-Experts inference under a hardware contract. 3 measured laws, 4 falsified ideas, and paper site.

Research platform for discovering and rigorously falsifying crypto trading strategies: event-sourced paper trading, implementation-parity verification, and a documented negative-results record.

Component-level benchmark for catastrophic forgetting in world models. Two negative results: forgetting does not follow the labelled task-distance axis, and it happens in the encoder, where the usual metrics cannot see it. 375 runs, with code, data and paper.

Does gnomAD LOEUF constraint predict drug-target safety? A negative result: LOEUF measures genetic loss-of-function tolerance well and clinical safety of inhibition poorly.