- A self-driving lab can spend all night running a recipe that was never going to work. Berkeley's A-Lab hit a 71% success rate, so ~29% of its attempts never yielded the target (A-Lab, Nature 2023). The fix isn't a smarter model. It's where the safety veto lives. 🧵
- In our Autonomous Lab AI demo, agents advise, code decides. A Lab Director LLM narrates strategy and a Critic flags drift. Neither can approve a synthesis. The go/no-go sits in a deterministic digital-twin gate that runs outside the agents and can't be overridden.
- Every candidate the optimizer proposes hits five real, published chemistry checks before a single reagent is touched: Goldschmidt tolerance window, octahedral factor, charge neutrality, Sn(II) humidity-oxidation hazard, and the organic-precursor annealing ceiling.
- A charge-imbalanced Bi substitution, or a Sn(II) mix that oxidises in humid air, is doomed before it runs. The gate returns BLOCKED, names the violated rule, and reports the exact reagent dollars and robot-hours it just saved. No model gets a vote on that.
- Over one seed-411 campaign: 38 doomed syntheses blocked before they ran. $4,419 in reagents and 90.1 robot-hours saved. Zero infeasible experiments executed. That's the A-Lab 29% waste leak closed by code, not by trusting an agent's self-report.
- The closed loop still finds the material. A lead-free perovskite meeting a hard 1.2 to 1.5 eV bandgap spec in 75 experiments, where random screening on the identical hidden objective would need ~1,491 (about 20x fewer). Real GP surrogate, target-directed BO.
- Honest scope: the objective is a physically-motivated synthetic benchmark, not real DFT or lab data, and no material was actually synthesised. The engine is a numpy/scipy GP with ParEGO acquisition; the production swap is BoTorch + Ax.
- Then the receipt. Every decision, syntheses and blocks alike, becomes an append-only ALCOA+ record with a SHA-256 hash chained to the one before it. 113 records this run. Edit one field and the chain visibly breaks at that record. Provenance mapped, not yet GxP-validated.
- For anyone building autonomous labs: should the feasibility/safety veto ever sit inside the LLM, or does regulated autonomy require it in deterministic code the model can't touch? Where do you draw the line? #selfdrivinglabs #labautomation #materialsinformatics #AIgovernance
- We build the brain, safety gate, and provenance layer on your existing instruments. Watch the gate block a doomed synthesis and the audit chain break at the tampered record, live: veriprajna.com/demos/autonomous-lab-ai
Published on X · August 18, 2026
On social media
See this post on its original platform
In our archive