Summary / Harness arms
A7 oracle_k (keep every patch)
Experiment E0: keep all k patches and ask whether *any* resolves. That is the ceiling a perfect selector could reach.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
On the 37 cells where candidates disagreed: oracle@k 97.3%, the vote 94.6%, a single session 83.8%. The full 150-cell run stalled at 125 scores.
What is missing
Finishing the full run.
Reviewer Claude, 2026-10-04
Reviewer's read: Selection is nearly solved
The vote already gets within three points of the best possible pick, so better selection has little left to win on this set. Gains have to come from producing a correct candidate more often, which points at localization (E1) and more draws (E2), as planned on 09-27.
How it is compared partial
Coverage@k against the selected patch.
Adoption rules shared by every arm: How the tests work.
Research & attempts
No documents yet.
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.