read-eval

Use this skill when a solver eval or historical solver log needs to inform the next optimizer update.

1. Extract score, score_metric, score_components, and quest_qoi_v_by_demo from logs/evaluate/eval_summary.json or score.yaml.
2. Read the live target first: higher score means lower mean_d1_d10.
3. Inspect mean_d1_d4 and the full d1..d10 curve before declaring a change promising.
4. Note whether the candidate improved the full example-count curve broadly or only at a single example count such as d10.
5. When writing guidance, separate visible evidence from speculation. Cite the specific solver example and the specific score component you used.
6. Prefer short, reusable lessons over one-off stories tied to a single lucky run.

Primary references:

• guidance/docs/problem.md
• guidance/docs/directions.md
• guidance/docs/mutation_surface.md
• guidance/docs/literature_notes.md