Ronaph3·Nov 6An AI Scientist Should Be Able to Say “I Was Wrong”Designing a Research Evolution Engine by combining ideas from Google Co-Scientist, Sakana AI Scientist, AlphaEvolve, and FutureHouse Robin.
Ronaph3·Nov 3A Better Causal Estimator Still Fails When the Confounder Is MissingFifty-four synthetic experiments in randomized assignment, observed confounding, and hidden confounding.
Ronaph3·Nov 3Latency Is a Price Assumption: Replaying a Synthetic Limit Order BookSnapshot recovery, ordered updates, taker VWAP, and why an offline execution-cost chart is not evidence of a trading edge.
Ronaph3·Nov 3A Swarm Needs a Control Plane, Not Just More AgentsAn offline experiment in dependency scheduling, fault injection, approval gates, and resumable state.
Ronaph3·Nov 2A Correct Gravity Equation Can Still Produce the Wrong OrbitComparing Euler, Velocity Verlet, and RK4 through energy drift, endpoint error, and numerical convergence.
Ronaph3·Nov 1The AMM Invariant That Should Not Stay Constant After a Fee-Bearing SwapA small exact-arithmetic model shows how to test reserve conservation, liquidity operations, and bounded execution traces.
Ronaph3·Nov 1I Built Three Search Baselines and Measured What They Actually RetrievedBM25 vs TF-IDF vs Reciprocal Rank Fusion on a small, explicitly curated technical search benchmark.A response icon1A response icon1
Ronaph3·Oct 31An Ethereum Block Dashboard That Tells You When Its Data Is SyntheticBuilding a transparent block-analysis pipeline with explicit data provenance.
Ronaph3·Oct 31LoRA on a Small Neural Network: Reproducing the Mechanism, Not the HeadlineAn experiment in parameter-efficient adaptation with a result that favors full fine-tuning.
Ronaph3·Oct 31A Walk-Forward Backtest That Didn’t Beat Its BenchmarkWhy reproducibility, execution delays, and realistic assumptions matter more than a profitable-looking chart.