Merge groups · geometry order · mixed result
Prospective admission is mixed: some policies separate, not a cross-model rule.
Admission here means choosing which rank-one edits to keep in a merge group before the merge is applied. Geometry-ordered admission is compared with magnitude-only and random policies. The pre-registered outcome is mixed.
The object
A policy ranks candidate edits and keeps a prefix. Geometry order uses |I_cos|; magnitude order uses residual-norm products; random draws a matched-size prefix.
The question is whether geometry order reduces downstream collateral relative to those controls on a frozen reference cell — not whether a universal admission rule exists.
What is measured
- Three of four pre-registered predicates hold on the reference cell (3/3 seeds)
- Geometry ESR and retention inside the random-draw interval does not hold (1/3 seeds)
- Pooled downstream geometry-drop mean ≈ 5.02 (descriptive over the three seed reports)
- No abstract cross-model admission is licensed
Protocol
- Fix the reference cell (Llama-3.2-1B, L12, g = 5) and the policy menu.
- Draw geometry, magnitude, and random prefixes; merge; score ESR, retention, and logit drop.
- Score the four pre-registered predicates. Mixed means at least one required predicate fails.
Measurements
Mixed
Three predicates hold; geometry ESR and retention versus random does not. The result is Mixed / not admitted. Do not narrate an intermediate log line as admission success, and do not present this as a deployable admission rule.
Companion
These cells sit next to the edit-federation map. They do not upgrade the map into a deployable admission rule.
Quarantine
Cloud tables labeled QUARANTINED are missing raw draws and are not live truth.