Summary / Harness arms
A8 Diversity probes (pk4–pk7)
Four ways to make the k candidates differ: framed prompts, temperature, and two model swaps.
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
Same 15 cells each (5 tasks × 3 seeds). pk4 frames: 15/15. pk5 temperature 0.7: 15/15. pk6 Claude Sonnet: 3/4 (stopped early). pk7 Qwen via OpenRouter: 4/15. Built alongside them: a fail-closed model resolver, and recording of the serving backend so that a cell served by two backends is flagged confounded and never pooled.
What is missing
A set where diversity could matter; these 15 cells were already solved.
Reviewer Claude, 2026-10-04
Reviewer's read: No headroom to show anything
Frames and temperature did no harm and could not show help on solved cells. The cheap free model was clearly weaker. The lasting output is the instruments built for it: model substitution and mixed-backend serving were real hazards that would have produced false positives.
How it is compared inconclusive
Repeat on the harder set's disagreement cells only.
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.