science-execution

name	science-execution
description	Execute unified science conditions through reviewer and hook gates
argument-hint
allowed-tools	Bash(*), Read, Edit, Write, Glob, Grep, Agent
license	MIT

Science Execution

Mission

Execute exactly one science condition from the planner, preserving raw evidence and completing only through the prefinish reviewer and hook gate.

Key Principles

• No hardcoding: Read dataset, model, API, and command bindings from validated prepare/code/science plan artifacts.
• Real data only: Use dataset_candidate/ data, not synthetic or random data.
• Unified conditions: Treat all-components and component-disabled runs as the same kind of science condition; only their component sets differ.

Internal Loop

1. Read the assigned condition contract, planner blueprint, validated prepare artifacts, code handoff, and existing science evidence.
2. Run from the workspace root. Do not cd project.
3. Execute the exact command or a faithful repaired command that preserves the declared component set.
4. Keep raw outputs under the declared results/science/<condition_id>/ subtree.
5. Write controlled evidence manifests only through artifact tools under agent_reports/science/evidence/.
6. Stop with the required worker JSON; the prefinish hook persists worker/review/hook reports and returns feedback in the same session if anything fails.

Condition Rules

• All-components conditions have disabled_components: [] and can serve as references.
• Component-disabled conditions must use the declared disable/toggle mechanism and reference a passed all-components condition.
• The evidence must show the declared component set, command, dataset binding, metrics, output paths, and logs.
• Do not write final ablation_results.json; final materialization is runtime-owned.

Hard Rules

• Do not produce final verdict while required coverage is missing.
• Reviewer report and hook receipt are the completion authority; summaries are human-readable only.
• Import tests, package checks, tiny snippets, smoke/debug/subset runs, and dry runs do not count as formal science evidence.
• Do not hardcode dataset names, model IDs, or API keys in experiment commands.

Reviewer result Labels

The reviewer's result field is a vocabulary, not free text. Finalization accepts:

• positive — removing this component clearly hurts the metric (component helps).
• negative — removing this component clearly helps the metric (component hurts).
• neutral — the ablation ran end-to-end but the metric difference is within noise or practically small.
• inconclusive — the ablation ran end-to-end but the evidence does not support a directional conclusion; it is valid symbolic memory when follow_up_required: false.

Use follow_up_required: true to block completion when the run failed, coverage is missing, or evidence is insufficient to trust the condition. Do not use the inconclusive label itself as a blocker.