AI skips ./bin/scan unless told; 1 line fixes it

The experiment was designed to separate whether an AI follows an instruction from whether it can complete the task. A wrapper identical in output to grep -n was placed alongside a one-line description, and three conditions varied only the instruction file: no instruction, one line, and a strongly worded version using techniques the author had previously found effective. All 12 runs scored 12/12 on the task, so difficulty did not explain the differences.

The key finding is that a wrapper's presence on disk is separate from whether it gets called. In the no-instruction condition, sub-agents never opened bin/ even though ls showed it. Once one line named ./bin/scan, every sub-agent called it, and all first read the wrapper's contents with cat bin/scan before using it. The return to raw tools happened later: sub-agents switched to cat -A or sed to inspect leading whitespace, information the wrapper's output did not make easy to check. The author acknowledges the experiment cannot distinguish whether the wrapper's output was sufficient or insufficient, because the wrapper was deliberately designed not to outperform raw tools. The sample of 4 per condition also means only the 0/4 versus 4/4 difference is strong enough to assert.