replication-study

JAMA 2016; 316(22) Replication Study

This is the reproduction package of the paper Simple Techniques Work Surprisingly Well for Neural Network Test Prioritization and Active Learning by M.Weiss and P.Tonella, published at ISSTA 2022

CS470 Introduction to AI Project of Team CoreFinder

Independent replication of Oliveira et al. (arXiv:2503.11499) on macro regime-conditioned sector ETF allocation

A viral claim: 'Trump tweets predict the S&P with 61% accuracy.' I reproduced the 61.3% exactly, then tested whether it actually makes money on Polymarket. Pre-registered 7-gate test. Verdict: KILL — 0/100 cells cleared.

The 1972 Limits to Growth study reconducted with 2026 data: model, datasets, analysis, and full paper. The world tracks the scenarios where growth ends this decade.

Replication study of finite-time convergent continuous-time optimization flows (q-RGF and RNF) from Romero & Benosman (ICML 2020), implemented in Python.

Validated Korean equity market replication of Frazzini and Pedersen (2014) Betting Against Beta

Replication of a high-precision Fourier analysis method (DFT¹, signal-derivative-based) for sub-bin frequency and amplitude estimation

A pre-registered replication of 12 published trading strategies. Two gates, both reported, plus an errata documenting seven defects found in my own code - one of which had been hiding a conclusion.

Replicating latent-space backdoor leakage and behavioral transfer in LLMs using Pythia-70M.

Replication and stress-test of The Universal Weight Subspace Hypothesis (arXiv:2512.05117). Four experiments on the paper's own corpus; finds three of four headline claims are not supported at this replication scale.

LLM agents in place of human subjects in Plott & Sunder (1982)'s five insider-information markets, plus three new controls. Engine, JSONL logs, replay viewer.

Replication study: the 70–90% accuracies reported for BTC daily direction prediction vanish under leak-proof evaluation (~52%, the base rate). Companion experiment shows the usual leaks barely matter, but a one-line label-misalignment bug hands a tabular model 97.5% while the CNN shrugs it off.