Finding 10 · 2026-08-16 (later)
The five audit controls existed but were never called by either driver
Controls written into the codebase — run_cpu_controls, margin_over_baseline, n_for_robust_rho, similarity_split — should have been part of the evaluation pipeline that actually runs. They had zero callers. When finally invoked, they condemned the split the driver had been using.
What we expected
Controls written into the codebase (run_cpu_controls, margin_over_baseline, n_for_robust_rho, similarity_split) should be part of the evaluation pipeline that actually runs.
What happened
Five controls, zero callers from either driver. When the controls were finally invoked, the instrument condemned the split the driver used. Commit e6bc365.
Why it happened
Not found in source — the notebook records the fact of zero callers and the result of finally invoking them, but does not state why the controls were written and never wired in.
What changed
The finding is folded into the section’s broader framing — the controls were a demo, not an instrument — and stands alongside the harness’s other self-audit corrections from the same session (the inverted knockout control and the tie-ranking metrics bug) as part of a single-day pattern the notebook itself names explicitly.
“The controls were a demo, not an instrument.”
Bring us the number you are least sure about.
That is usually the one worth thirty minutes.