draft / research

To generate an evidence-backed research brief, the skill loads the latest stage plan and produces a document satisfying every Success criteria item. It walks context references, resolves merges, and snapshots relevant fields into the trail when decisions depend on context state.

Creating a JSONL dataset requires setting the learning_variant to dataset for the playbook research action. The skill processes the specified learning window, emits dataset.jsonl training rows, and writes a strict dataset.manifest.yaml file.

A calibration quadrant table is a variant-specific artifact that joins gate recommendations, operator answers, and verify outcomes. It is generated when the learning_variant is set to calibration, providing a matrix to analyze learning-loop performance.

Context snapshotting works by walking context references, resolving merges, and recording needed context fields into the trail. This ensures that when drafting decisions depend on context state, the exact state is preserved for auditability.

When encountering malformed windows or unreadable trail files, the skill handles them safely without crashing the workflow. It enforces observability of Success criteria and records standardized trail entries while managing these edge cases gracefully.

Yes, you can aggregate rejected review patterns by setting the learning_variant to rejections. The skill processes the action's trail context and produces a dedicated report analyzing the rejected patterns within the specified learning window.