NMR spectrometer commissioning

A collaborative agent took an uncalibrated diamond-sensor NMR setup to a first scored run in 3.2 h.

~1 Hz · live hardware

Photo of the NV-NMR apparatus, and the figure of merit of the 34 scored experiments over 20.8 hours on live hardware, with the running best from the first scored run at 3.2 hours. — Live hardware · photo and plot from the preprint's Fig. 6 (plot traced from the figure)

3.2 hfrom uncalibrated to first scored run

0.97 / 1.47 Hzbest / median, 10 runs at final settings

151experiments on live hardware in 20.8 h

The problem

NV-NMR detects NMR with nitrogen-vacancy centers in diamond instead of a pickup coil, so it can study sub-nanoliter samples. Setting one up (find resonances, calibrate pulses, tune acquisition) takes an expert's focused attention.

What we did

Researchers set the goal, a 36-h budget, safety limits, a figure of merit fixed in advance and an 8-step protocol with no parameter values. On live hardware the collaborative agent ran 151 experiments in 20.8 h, tuning one parameter at a time with controls.

Result

Proton NMR signal detected at 61 min; precision runs from 3.2 h. Best run: 0.97 Hz uncertainty (1.1 ppm) in 10 min; median of ten runs at the final settings 1.47 Hz. Best readout delay fell in the hand-found range.

Keep in mind

Drift, not the sensor, limited precision. An independent audit re-checked the scores yet rejected the agent's claim of a replicated record; the median is quoted too.

In the preprint

Sec. IV.E, “Nuclear magnetic resonance (NMR) experiment commissioning,” and Fig. 6 of arXiv:2610.17193 describe the campaign and its checks in more detail.

Fig. 6 of the preprint, full size — Fig. 6 of the preprint: top, the NV-NMR apparatus (inset: sample tube, diamond, coils and magnets); middle, all 151 experiments by type against their start time; bottom, the figure of merit of the 34 scored experiments, each marked CRB- or drift-limited, with the running best, over 20.8 h. Click to enlarge.

Data: Live hardware
Models (as stated in the preprint): Claude Fable 5.1, Claude Opus 5.5, GPT 6 Astra (the preprint does not say which model did what)
PhysKit version (as stated in the preprint): 0.12.1, an earlier version
Where: Sec. IV.E, Fig. 6

We list the versions the preprint reports so readers can trace each result to its tools, not as a recommendation: for new work, use the latest release.