cerl — Ebolathon: muni-bdbv-gp

Hackathon project for the Ebolathon: an RL environment where an LLM agent operates a real computational-chemistry stack (OnePot analog search, Rowan docking, ADMET-AI, Boltz-2 affinity) to redesign known ebolavirus GP entry inhibitors into improved, purchasable candidates against Bundibugyo ebolavirus (BDBV) glycoprotein — a species with no approved vaccine or therapeutic.

Layout

Path	What
environments/muni_bdbv_gp/	the verifiers SandboxEnv (v0.2.0) — README
muni_api.py	shared Muni/OnePot API wrapper (disk cache, spend guard, mock mode)
agentic_loop.py	pre-environment prototype: greedy hill-climb over OnePot analog space (VP35/4IBB)
mcts_search.py	RL-free baseline: UCT Monte Carlo tree search over OnePot analog space (free tools only) → step-by-step reasoning trace + top-10 candidates
validate_candidates.py	props/PAINS → OnePot CORE → Rowan docking → ADMET validation pipeline
configs/rl/	GRPO configs (gpt-oss, gpt-oss-20b, nemotron, qwen3.5-35b)
SUBMISSION.md	hackathon submission text (methods & rationale + proof files)

Pitch deck, analysis data, and PDB structures are kept out of this repo.

Quickstart

export MUNI_TOKEN=mcli_...      # mint: muni token create --name ci --ttl 90d
export ONEPOT_API_KEY=...       # required — onepot.ai

pip install -e environments/muni_bdbv_gp
prime eval run muni-bdbv-gp -m Qwen/Qwen3-30B-A3B-Thinking-2507 -n 3