Compute Commons: usable robot inference

Four robot operators share 400 ms of inference time. An inference job lasts 200 ms. Robot A runs out of queued actions at 200 ms; B's target moves outside its pose tolerance at 200 ms. D and E remain usable until 400 ms. E is an eligible small operator.

Synthetic teaching model. No real robot or participant findings. Timeliness does not certify safety. A local protective controller is assumed. This snapshot is not a deployed real-time scheduler.

1. Choose a rule and predict

2. Inspect the outcome

The feasible auction chooses the highest credited value among schedules meeting robot deadlines. Winners pay their critical bid. Truthful value reporting is weakly dominant only with verified state, eligibility and deterministic timely delivery. Capacity-only admission deliberately violates that delivery assumption.

What surprised you? How would you change your decision? Do you regard the rule as legitimate, and why?

Nothing is uploaded. Do not enter names, sensitive logs or proprietary information. Export is enabled after both reflections are completed. A test click is not a research participant.

3. Explore robot-state assumptions

After committing your prediction, compare all rules using the same values. Relax one physical constraint to test whether the robot details change the outcome.

Ablation

4. Final self-challenge: what must be verified?

Can false telemetry improve a bidder's utility? The Python test checks 88,200 joint bid/deadline deviations in 300 seeded batches and finds no positive gain under fixed execution times and nonrefundable charges. This finite test is not a guarantee for all telemetry.

Try a different report: A needs 200 ms but claims 100 ms. A, B and C value timely service at 10, 9 and 8; all deadlines and the horizon are 300 ms. B and C need 100 ms each. Actual execution uses real service times.

Mechanism-design lesson: verify service duration separately. A deadline-only test cannot establish immunity to duration fraud.