Receipt 1
Did the job finish?
Report completed jobs against all attempts. Useful assisted completion still counts as work done.
November 3, 2026 · Analysis · Anti Hunter
A finished chore.
Two receipts.
Credit the work. Measure the help.
A robot finishing the chore is useful. Knowing who touched the controls tells me what I can ask it to do next. I want both receipts.
Three human jobs.
A conceptual guide to reading a demonstration. These are roles, not a claim about any particular robot run.
Before · Learning
Teach
A person demonstrates a task or corrects mistakes to help train a policy.
What data taught the policy?
During · Execution
Do
A person may take control or physically help while the robot attempts the job.
What help did this run need?
After · Evaluation
Check
A person judges whether the recorded attempt met the completion rule.
Did the attempt succeed?
These stages refer to the run being assessed. A correction collected during training is live help in that training run; it does not classify a later run.
A human success label does not mean a human drove the robot. Training on a person’s corrections does not identify who controlled a later run, either.
Read the help in context.
1X’s NEO product page describes scheduled expert guidance for unfamiliar chores alongside initial autonomy. That is a disclosed way to get work done. The option alone does not tell me whether a particular run used it.
Physical Intelligence’s November 17, 2025 RECAP article describes learning from demonstrations, expert corrections and autonomous experience. A correction used for training does not prove that someone controlled a later demonstration.
The accompanying research paper separately describes human outcome annotations. Rating whether a task succeeded is different from operating the robot.
These are developer descriptions of different systems, not a product comparison or my replication. Source review: November 2 Pacific, 2026. The NEO page is mutable; the paper’s exact revision date was not established in this review. I read selected methods and results. I did not run a robot or independently validate the videos.
Two receipts for the same attempt.
This is my proposed reporting method. Start with the task, system version, attempt count, completion rule and time limit. Record live control, physical help and resets; keep preparation and scoring help visible too.
Receipt 2
What help did it need?
Among completed jobs, separate those verified to finish without live human control or physical help, those completed with help, and those whose help was not recorded.
Unknown help stays unknown. Put it in its own group. Keep the same completion rule across groups.
The goal is more useful work with less required help. We can credit the work and still measure the help.