Research bench where you can find open source software and reports: keppylab.com
evalroute Route a task to the right (model, reasoning effort) arm before you start, from a measured cost-per-verified-success table. Python library + standalone CLI on PyPI; outcomes you label grow the table.
titans-mini A streaming engine over a swappable test-time memory core (MLP-weights memory vs. generated-weights vector memory)
thomas thomas.train() — a training harness. Case→reward→train: take a Case set and a score function, get a baseline card (gonogo), run a training loop, compare before/after. Two paths: pretrain (Modal GPU) and post-train (Tinker LoRA RL or TRL GRPO + vLLM).
cotfaith CoT (un)faithfulness, study one: hint-following and confession rates on Qwen3-1.7B — pre-registered decision log, blind-labeled judge validation, byte-exact run artifacts
gonogo The eval harness from my implementations, open sourced. Scores an agent on your real cases and returns a deployment decision, including "not enough evidence yet."
Three plugins in the Hermes Agent community plugin catalog. Together they take a task from which model? to ship it or not.
hermes-plugin-evalroute
Routes each task to a (model, reasoning effort) arm, and /rate labels the outcome. Thin adapter over the evalroute library.
hermes-plugin-gonogo Turns a pilot-scale eval score into a ship / don't-ship decision. 47 out of 50 is not 94%.
hermes-plugin-thomas Fine-tunes a calibrated encoder on Modal behind Hermes' approval gate, then rates it with gonogo.
describe• Speak systems into existence
cobol-reporter 🔭 A RAG and report generator for helping to understand COBOL code
Disease Lab 🧪 Knowledge graph informed AI for disease research
WorldEnder.ai 🌎 A text-adventure about extinction events with a RAG backend