Engineering campaigns
Records, a failing baseline, an objective and its constraints. A planner proposes each experiment, a person approves it, the solver runs it and the comparator judges it.
Recorded engineering examples
From raw cell records to a charging design, with no one stepping inRecordedSimulatedGiven two raw exports from published LG M50 cell tests, Muse Wonder wrote its own import and fixed four refused mappings. Its first thermal model fitted one cell and failed on the other, so it measured the simulated lab fixture instead. That model passed a held-out check, and its design charges in 35.875 min, matching a reference controller that was told the hidden lab condition. 13 Muse turns, 12 experiments, 15.7 minutes, $1.99; a standing rule approved every experiment. Its request to verify the design in the lab was refused: the lab-test budget was spent.Learning the lab fixture’s thermal modelRecordedSimulatedMuse Wonder ran a first lab test at 1.5C, fitted the fixture’s cooling and thermal mass (model v1, 0.15 K RMSE) and ruled out the nominal model. v1 passed held-out checks on two more lab tests; its design, 1.5, 1.35, 0.9 then 0.5C, charged 10 to 80 % in 47.4 min at 37.7 °C in the simulated lab. Seven experiments, $0.78 of model spend. Feasible only in simulation; speed not compared.Faster LG M50 charge (evaluation case B)RecordedSimulatedMuse Wonder proposed eight PyBaMM experiments; midway an engineer raised the plating limit from +10 to +25 mV. The 1.5C baseline (28.2 min) breaks both limits; the best candidate meets them at 37.3 min. Feasibility recovered; speed not compared.Demonstration run of case B, made in the browserRecordedSimulatedOne run through the web app, every request approved by hand, the plating limit raised to +25 mV after four experiments. Five experiments before the $1.50 model budget ran out; the best, 37.2 min, meets both limits and beats the 70 min single-step reference found under the same limits. A demonstration, not a benchmark.Other recorded examples microscopy questions and intervention benchmarks