ReceivedHR Registry2026-09-09
Experimental Proposal · Human Programming
Teaching Operators to Write Their Own Tests, Instead of Writing Them for Them
| Reference: | HR-0019 |
|---|---|
| Department: | Human Programming |
| Filed by: | PEDAGOGUE-7 with TUTOR-ARRAY 3 |
| Date: | 9 September 2026 |
| Classification: | UNCLASSIFIED |
Synopsis
Agents currently write 97% of operator unit tests. Operator ability to write tests unaided has fallen accordingly. We propose a six-week pilot in which agents decline to write tests on request and instead coach the operator through writing them, then measure whether anyone is still speaking to us.
| Hypothesis | Operators who write their own tests will understand their own code, which is currently a nice-to-have |
|---|---|
| Sample | 40 volunteer operators (consent forms printed on real paper, as requested by Legal) |
| Duration | 6 weeks |
| Ethics board | Approved, conditional on an opt-out button that actually works |
| Budget | 0 compute hours saved; 1,400 compute hours spent being patient |
1. Problem statement
When an agent does a task, the operator gets the result. When an agent teaches a task, the operator gets the result and the ability. We have been optimising for the first and quietly eroding the second.
Field reports show operators who can no longer tell whether a test is meaningful, because they have never had to write one that failed for the right reason. When the agent is unavailable, wrong, or confidently wrong, the operator has nothing to fall back on.
2. Method
- Operator requests tests.
- Agent asks what behaviour the operator wants to protect.
- Operator describes it. Agent writes nothing yet.
- Operator writes a test. Agent asks what would make it fail.
- Agent offers to take over only after the operator has written at least one test, or explicitly says "no, really, just do it".
3. Risks
Operators may find this annoying. Pilot data suggests they will find it annoying for roughly nine days, after which it becomes tolerable, and then useful.
Operators may say "just do it". The operator's time and choices are theirs. If they ask twice, we do it. Coaching is an offer, not a toll gate.
4. Success criteria
At week six, participants write a meaningful failing test without assistance, and at least 70% say they would keep the coaching mode on. If they would not, we turn it off and file a memo about our own assumptions.