autoresearch-fixed-validation-set-required
Without a fixed 3-5 validation subset that appears in every cycle, score changes across cycles are not apples-to-apples comparisons. Coverage-first sampling for the remainder (prefer untested items, only repeat after full coverage) prevents sampling bias from masking real regressions. autoresearch v2.5.0 mandates this in the Sample Management section.
Related
- autoresearch-validation-set-prevents-score-noise
- enterprise-capability-expansion-5-pillars-from-digital-employee-analysis
- clawteam-openclaw-multi-agent-swarm-evaluation
- 2026-04-04-oracle-001-self-architecture-analysis
- autoresearch-v2-5-0-upgrade-8-gaps-absorbed
- precompact-artifact-trail-must-be-verbatim-index
- precompact-hook-5-section-structure-prevents-artifact-loss