science-planning

name	science-planning
description	Build unified science condition plans from validated prepare/code evidence
argument-hint
allowed-tools	Bash(*), Read, Edit, Write, Glob, Grep, Agent
license	MIT

Science Planning

Mission

Translate validated handoff artifacts into one ordered science condition plan. A condition is fully defined by enabled_components and disabled_components; the all-components condition is the reference case where disabled_components is empty.

Key Principles

• No hardcoding: Read dataset, model, and API bindings from validated prepare handoff. Do not invent or hardcode values.
• Unified science: Use one science.plan under agent_reports/science/plan/latest.json.
• Real data only: All experiments must use dataset_candidate/ data.

Protocol

• Plan all raw outputs under results/science/<condition_id>/.
• Plan exactly 1 + len(idea.json.components) conditions.
• The first condition must be the only all-components reference: all canonical idea.json.components enabled, disabled_components: [], and no reference_condition_id.
• Every later condition must disable exactly one canonical component and reference the first all-components condition.
• Across component-disabled conditions, every canonical idea component must be disabled exactly once, with no omissions and no repeats.
• Every condition must declare dataset bindings, metrics, exact command, output directory, raw evidence paths, and component set.
• Only place conditions in parallel when they use disjoint output dirs and cannot overwrite each other.
• Keep plan, worker reports, reviewer reports, hook reports, evidence manifests, and summaries under agent_reports/science/.
• Preserve reviewer fields result, metric, value, confidence, analysis, and method_context for disabled-component conditions so final materialization can write ablation_results.json.

Hard Rules

• enabled_components plus disabled_components must cover idea.json.components exactly for every condition.
• Component names must match idea.json.components exactly.
• Commands must run from the workspace root and should call project entrypoints as python project/....
• Preflight, smoke, debug, subset, import-only, or dry-run evidence does not satisfy final science completion.
• Do not write final ablation_results.json; it is runtime-owned.