Composable computational-science methodology skills for AI research agents — pre-registration over TDD. A science-domain reimplementation of Superpowers.

Pre registration utility

Instruments that check what AI agents claim against committed bytes. The diff gate reads a PR summary against its diff and reports by default. EXTERNAL-1 upheld 23 of 100 sampled accusations, precision 0.23 (85 path claims, styxx 7.47.0; path accusations withheld since). Receipts, failures included, are in the repo.

Node.js SDK is a connector library for the insanely fast CentralNic Reseller (formerly RRPProxy) Backend API.

Pre-registered independent eval of TypeSafe Jev against a nano-class LLM, a frontier LLM, and a supervised encoder (Banking77 + CLINC150 zero-shot)

In this repository you are going to find the documents we produced to support the discussion in our Digital Methods of the Month. These documents will help you orienting yourself if you want to pickup the method in your research. Go to the readme file

Research code: parameter symmetry and weight-space learning on sine-network implicit neural representations.

A theta-driven CA1 spiking model in which the standard gamma statistic reports the drive's harmonic comb rather than the rhythm, and no screen recovers it. Code, run logs and figures for the preprint.

Command-line tools and a Claude Code plugin bundling preregistration, citation verification, and results provenance for reproducible science workflows

A Quarto preregistration template for preregistering ecological modelling studies

GO/GOLANG SDK is a connector library for the insanely fast HEXONET Backend API.

Designing, Evaluating and Testing Preregistration Templates for Ecology and Conservation

Python SDK is a connector library for the insanely fast HEXONET Backend API.

🪞 Catch false positives/negatives in AI eval claims — pre-registration, fair-baseline, small-sample CI. Zero training, zero deps.

BITSS Reproducibility Workshop, New Delhi March 2017

Historical RV15 forecasts for six U.S. stocks. Option state improves both families; the mixed B2 block improves the linear model only. Source-time assumptions, placebo and final-window limits remain material. Public aggregate checks are not an independent licensed-data replication.

Public connectomes as fixed reservoirs: load a real brain's wiring, simulate it as a spiking network, drive it, read it out — with controls and pre-registration. Experiment 001: two fly brains.

Cross-asset price impact is set-identified, not point-identified: the confounding gap has rank at most K + rank(B). A diagonal truth still yields a dense cross-impact matrix, and the trades immune to the resulting cost error form a large but measure-zero subspace. Preregistered and reproducible.

A nested, content-addressed benchmark family for vector search at 10^6-10^12 rows, with three independent notions of a correct answer. RC validation program complete: the generation-vs-memorization boundary is measured.

A method for keeping AI-assisted research honest: pre-registration + an independent agent that recomputes every claim from raw data. Forged on markets — 16 certified dead hypotheses. Runnable demo included.