Physical-lab autonomy for the Phase-178 autonomous-research stack (TODOS/phase-178 §178.8). Extends Phase 130 (Kalika autonomous experimentation) with cloud-lab / robotic-lab driver abstractions and a defensive safety gate.
Components#
| Module | Summary |
|---|---|
driver.py |
The unified LabDriver trait (reserve/run/observe/log/abort) with Emerald Cloud Lab (emits real Symbolic Lab Language), Strateos (emits real Autoprotocol JSON) and Opentrons OT-2/Flex (emits a real Python protocol) drivers over a fail-loud transport, plus an instrument catalogue. (§178.8.1) |
campaigns.py |
Coscientist active-learning reaction optimisation (real UCB/k-NN loop over a Suzuki–Miyaura space), A-Lab candidate → synthesis → XRD phase-ID → active-learning campaign, and ChemCrow 18-tool executive (real MW + safety-screen tools, seams for the rest). (§178.8.2) |
models.py |
FAIR-Chem/UMA property prediction and Boltz-2/AlphaFold3/ESM3 protein design as fail-loud seams, each with a real offline baseline (group-contribution molar-mass/boiling-point; Kyte–Doolittle hydropathy). (§178.8.2.4–.2.5) |
safety.py |
A defensive wet-lab hazard classifier (CWC-Schedule-1, controlled-substance, precursor-export, biosafety-level, high-energy, germline-CRISPR), a mandatory human-in-the-loop gate (PROHIBITED categories never dispatch; HIGH/BSL-3+/dual-use require approval), and a tamper-evident chain-of-custody. (§178.8.3) |
The native-program emission (SLL/Autoprotocol/Opentrons), the active-learning loop, XRD matching, offline model baselines, and the entire safety gate + chain-of-custody are fully local and tested. The cloud-lab APIs, learned chemistry/protein models and LLM tool backends are honest fail-loud seams.
Testing#
bash
python3 -m pytest tests/ -q # 25 tests
python3 -m ruff check .