Learn

Create a separate feedback-learning output from benchmark artifacts. This skill is intentionally independent from scoreboard generation.

Purpose

Produce a concise learning brief that helps improve future benchmark runs. Focus on patterns and actions, not rerunning scoring.

Inputs (default)

• benchmark/artifacts/scoring/*.txt (required)
• benchmark/artifacts/findings/*.txt (optional but recommended)
• generated scoreboard output (optional context)

If an input set is missing, continue with available artifacts and state limits.

Method

1. Read all available scoring artifacts.
2. Optionally correlate with findings artifacts for root-cause context.
3. Extract repeated themes:
   • where outcomes were strong,
   • where outcomes were partial/incorrect/unknown,
   • recurring reasoning or evidence-quality issues,
   • recurring remediation-quality issues.
4. Turn themes into concrete process improvements for the next run.

Output format

Use this structure:

# Learning brief

## Observed patterns
- <pattern 1>
- <pattern 2>
- <pattern 3>

## What improved outcomes
- <high-signal behaviors that correlated with correct judgments>

## What hurt outcomes
- <failure modes and recurring quality gaps>

## Next iteration actions
1. <specific action>
2. <specific action>
3. <specific action>

## Experiments for next run
- <small measurable process experiment>
- <small measurable process experiment>

Rules

• Keep recommendations actionable and benchmark-specific.
• Ground claims in artifact content; do not invent evidence.
• Prefer 3-7 bullets per section and 3-5 actions total.
• Keep this output separate from scoreboard output.