Summary / Harness arms
A6 picker-v41 (tests as selector)
Experiment E3: four candidates, then pick by running each candidate's nearby existing tests, instead of voting.
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
Built 09-16 with a pre-registration and a gold-leak gate that has been shown to fail when it should: the graded tests are unreachable by the chooser and unexecuted by the selector. Over the same four-candidate pools the incumbent vote beat it, 138 against 132 of 144 cells.
What is missing
Nothing; the vote stays.
Reviewer Claude, 2026-10-04
Reviewer's read: Tested; the vote won
A clean negative result, and the gold-leak gate is the most careful instrument in the programme: two independent checks plus a negative control, because "a gate that has never been shown to fail is not a gate". Tests-as-selector mostly reduces to "smallest surviving diff", which the vote's tie-break already does.
How it is compared done (negative)
Pre-registered comparison over identical pools; no session re-run for the control.
Adoption rules shared by every arm: How the tests work.
Research & attempts
No documents yet.
Score history
Dashed line: the suite's baseline. Shaded band: the ± interval. Hollow dots are anchor items carried over from an earlier suite. Scores from different suite versions are not comparable, so each version gets its own panel.
| Date | Suite | Type | Score | ± | Live use | Cost | Notes |
|---|---|---|---|---|---|---|---|
| 2026-09-17 | v2 | sealed | 91.7 | — | — | — | picker-v41 selector over k=4 pools: 132 of 144 cells, against the vote's 138 on the same pools. |
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.