cooleryuField Notes

Frozen evidence package

Claims you can inspect.

The protocol and result ledger behind the first ContextAdapt real-code retention holdout.

The holdout contains 30 tasks from 12 repositories, nine context policies, eight source budgets, and 2,160 evaluation cells. Gold patches were evaluator-only.

Direct result: ContextAdapt exceeded head-and-tail truncation by 22.4 percentage points in strict evidence retention across the seven compressed budgets. Its 7.1-point difference over lexical intent retrieval was inconclusive.

Evidence files

Interpretation boundary

This package evaluates evidence retention under source budgets. It does not demonstrate patch generation, passing tests, end-to-end task success, or realized provider savings.