analyze

To extract config levers from trajectory files, read YAML frontmatter and body excerpts from task files under _meta_scratch/traj to identify recurring patterns, then map them to candidate hypotheses for HarnessConfig levers with supporting evidence and citations.

Analyzing passing and failing trajectory clusters involves reading per-task frontmatter and body excerpts to surface recurring patterns and failure modes, then generating candidate hypotheses mapped to specific levers with intent classifications like corrective or preservative-lock.

Generate candidate hypotheses by reading YAML frontmatter to capture Behaviour, Eval, and Judge signals, applying the Lens × Lever reflection axes, and mapping findings to Configuration, Control, Action, or Instruction levers with corrective or preservative intent.

Use retroactive check variants A/B/C during global analysis rounds to enforce a structured workflow that validates candidate hypotheses against trajectory evidence, ensuring generated config changes are backed by rigorous frontmatter-based signals and citations.

Yes, you can batch read multiple trajectory files from _meta_scratch/traj to surface recurring failure modes across tasks, and delegate batch reading operations to spawn_reflect_worker when processing large volumes of trajectory frontmatter and body excerpts.

Trajectory pattern mapping relies strictly on rigorous frontmatter-based signals from task files, meaning insufficient or malformed YAML frontmatter limits the ability to generate evidence-backed candidate hypotheses for HarnessConfig lever adjustments.