Scoped Trial

Binding sources

• docs/FINAL_SYSTEM_VISION.md — scoped experiments feed continuous research data for L6 analysis; must stay traceable (baseline vs full vs scoped).
• AGENTS.md — default-off; paper-only; no collapsing modes into an ambiguous black box.

Purpose

Use this skill to manage the narrow paper-only rollout candidate path for scorer-confidence experiments.

The scoped trial exists to test whether a concentrated benefit can be isolated without changing the baseline or globally enabling the experiment.

Use When

Use this skill when:

• editing scoped trial config or scope parsing
• editing scope matching behavior
• comparing baseline vs full experiment vs scoped trial
• improving operator visibility around scoped trials

Do Not Use When

Do not use this skill for:

• baseline scorer changes
• full experiment governance-only summaries
• generic strategy or risk changes unrelated to scoped trial behavior

Safety Contract

• Scoped trial remains off by default.
• Baseline behavior remains untouched unless the scope explicitly matches and the trial is enabled.
• Scoped trial stays paper-only.
• Full experiment and scoped trial must not silently collapse into one ambiguous mode.
• No auto-enable, no auto-tuning, no default changes.

Primary Workflow

1. Read AGENTS.md.
2. Inspect scope flags, matching logic, and current comparison outputs.
3. Preserve baseline and full-experiment paths first.
4. Keep scope syntax explicit and operator-readable.
5. Verify logs and summaries show whether the scope matched.

Required Checks

• Scoped trial is clearly labeled in logs, journal, summaries, and panel views.
• Scope fields stay explicit, readable, and bounded.
• Baseline vs full vs scoped reporting stays comparable.
• Match and miss paths both remain explainable.

Expected Output

• Scoped Mode: what the scope can match.
• Comparison Layer: how baseline, full, and scoped views differ.
• Safety Check: default-off and paper-only confirmation.