Field note: an agent buys evidence for a robot arm
The setup
Our LeRobot test bench replays recorded pick-and-place episodes from a low-cost SO-101 desktop arm. A pink Lego brick goes into a transparent box. WCGW's analysis of those recordings produced six hazard hypotheses: unexpected movement, incorrect localization, excessive force, collision, failed grasp, and loss of control. All six are unconfirmed. The dataset has no physical units, so the recordings alone can't say how hard the gripper squeezes.
That is a gap you close with evidence, not with more reasoning. So we tried something new: we let an agent buy the evidence.
Nevermined's AI Services Catalog lists 163 pay-per-call services behind one Router and an MCP server. Beta accounts receive 200,000 welcome credits. We pointed our agent at the catalog with a budget and a rule: quote everything, pay only within the cap, and record every payment.
What the money actually is
The agent spent 16,335 Nevermined credits during this experiment. A cited web search costs 621 credits, and scraping a page costs 216.
Credits pay both rails the catalog settles on (x402 on Base, MPP on Tempo). With no wallet delegation set up, a call that runs out of credits simply fails. Nothing falls back to real money.
Round 1: the obvious plan
We started generic: hazards of collaborative robot arms.
| Call | Credits | What came back |
|---|---|---|
| Perplexity search | 621 | A SafeWork NSW / UTS report on cobot risks, with an 18-row hazard table |
| Exa paper search | 1,188 | Five on-topic risk-assessment papers |
| RecallRadar "robot" | 10,800 | Twenty product-safety notices, every one a toy or a robotic pool cleaner |
| Firecrawl × 3 | 648 | Two useful documents and one wrong page |
The lesson was uncomfortable. The cheap, general services delivered the value; the one "safety" service cost 81% of the round and was the least relevant. The wrong page was the agent's own fault: it guessed a URL instead of resolving the DOI. We made that a rule for the next round.
Round 2: the test bench
This time, the agent first read the test bench's own hazard report. It then searched the catalog for free, chose three providers, and chained them: search → resolve URLs → scrape → combine. The budget guard was code, not good intentions: any call quoted over 2,000 credits, or pushing the round past its 10,000 cap, was refused before payment.
It spent 3,078 credits across six calls. Every call settled, and none was wasted.
What the evidence says about "excessive force"
The SO-101 uses Feetech STS3215 servos. The follower variant is rated at about 27 kg·cm stall torque at 7.4 V. If the gripper jaw's lever arm is 5–10 cm (our assumption, not a measurement), that is roughly 27–54 N at the jaw.
That sounds harmless next to the 140 N that ISO/TS 15066 tables commonly list for hands and fingers. But force is the wrong number to stop at. The standard also sets pressure limits, about 300 N/cm² for a fingertip pad. Spread 54 N over a flat pad and you are far below it. Concentrate it on a 1 mm jaw edge and you can be above it. The hypothesis worth testing is not "the arm is too strong". It is "the jaw edges are too sharp".
(ISO/TS 15066 is written for industrial collaborative robots. We use it here as a reference point, not as a compliance claim about a hobby arm.)
Two failure mechanisms we did not have
A practitioner's teardown of the STS3215 described two behaviours that turn into candidate hazards:
- Thermal cut-off. At a default 70 °C the servo disables torque. On a loaded joint, that means the arm drops.
- Overload protection that software can erase. The servo backs off after sitting at 80% of stall for two seconds, but a new position command clears the flag. A control loop that streams commands without reading status can keep a stalled servo pushing until it burns out. Whether LeRobot's loop does this is now a question for us to check, not an answer.
The tables disagree
One web source lists the hand limit as 75 N and 40 N/cm². Others say 140 N. In a safety-equipment maker's reproduction of the standard, "75" for hands and fingers appears among the body model's spring constants, in N/mm. Our working hypothesis: someone copied a stiffness into a force column. Whatever the cause, the practical advice holds: take your limits from the standard, not from a blog table. That includes this one.
What this was, and what it wasn't
An agent discovered services, chose them, quoted them, paid for them, and chained their outputs into a traceable evidence pack. Every claim in it is tied to a settled payment with an on-chain receipt. That part worked.
What it did not do is decide anything. A human approved both budgets. The force figure is an estimate built on an assumed lever arm. The five new hypotheses are candidates, waiting for a force gauge, a thermometer, and an engineer. That's the WCGW way: the agent widens what you can see, and the human owns the safety model.
Next on the bench: measure the jaw geometry and the force at stall, log servo temperatures during long replays, and read LeRobot's control loop for the overload flag.
Spend: 16,335 of 200,000 credits (≈ $0.15), 12 of 12 calls settled.
The data: the transaction ledger, payment receipts, quotes, the agent's decision log and both evidence packs are published as NEVERMINED-BUYER-200K-01 on Hugging Face (CC BY 4.0).