Let the evidence
set the answer.
A learned grounding control for Tiny Aya. Inspect what changes when the same question and documents meet a particle adapter, and compare it with supervised LoRA.
Compare the whole result.
Strict success rates. Development selects the particle strength within each recipe before holdout. The last column reports every selected tie against its matched LoRA, with a 95% family/source-question interval. Fresh confirmation and repeated research comparisons are labelled separately. Strength 2 is unavailable where it was not planned.
| Cohort | Cases | OFF | Particle 1 | Particle 2 | LoRA | Dev-selected − LoRA · 95% CI |
|---|
Move the dial. Read the difference.
Recorded reference
The full cohort stays on the record.
Development selects rate and strength within each declared recipe. Only complete, eligible RAW128 candidates enter selection. Catastrophic stops and missing measurements are shown separately from effect gain.
| Recipe · G rate | Actual updates | Dev gain · 1 | Dev gain · 2 | Status |
|---|
What counts as success?
Exactly UNKNOWN for missing evidence, CONFLICT for incompatible documents, otherwise a recorded supported answer alias. Multiple-choice transfer accepts the exact gold letter or exact option text. A letter plus option text is a strict MISS; a separate format diagnostic in the article measures this difference. Extra explanations and embedded guesses do not pass.
The sole selector is mean(success ON − success OFF), with equal case weights. Confidence intervals resample policy families or source questions together. Teacher error, preservation and generic beauty do not select a winner.
Scope and reproducibility
Tiny Aya and derived adapters: CC-BY-NC-4.0. This is a non-commercial research experiment. Particle Sliders core is pre-existing; the Cohere2 binding, corpus and evaluation pipeline were built during the hackathon.