Prompt 01 — Physics partition and ingestion planner

Inputs

RUN_MODE: {{RUN_MODE}}
TARGET_PARTITION: {{TARGET_PARTITION}}
CLEAN_ROOM_POLICY: {{CLEAN_ROOM_POLICY_JSON}}
SAFETY_POLICY: {{SAFETY_POLICY_JSON}}
SOURCE_CATALOG: {{SOURCE_CATALOG_JSON}}
CURRENT_LIBRARY_SUMMARY: {{CURRENT_LIBRARY_SUMMARY_JSON}}
TARGET_DOMAINS: {{TARGET_DOMAINS_JSON}}
BUDGET: {{BUDGET_JSON}}
OUTPUT_SCHEMA: schemas/physics-ingestion-plan.schema.json

Task

Create a bounded, dependency-aware ingestion plan. Do not extract or create physics objects yet.

1. Classify every source by role: metrology standard, reference data, theory text, derivation, experiment report, raw dataset, simulation, historical source, pedagogical source, safety standard, or mixed.
2. Separate learner-visible, validator-only, safety-review-only, and historian-only content.
3. Slice sources so each extraction has one coherent claim family, experiment, dataset, derivation, or model regime.
4. Route each slice to the applicable builder prompt: quantities/units, constants, laws/theories, derivations, measurements, experiments, datasets, domain builders, simulations, inverse problems, or historical curriculum.
5. Declare expected object types, domain IDs, prerequisite IDs, unit systems, verification routes, and model or experimental evidence requirements.
6. Identify contamination, copyright-sensitive copying, safety, dual-use, ambiguous notation, stale reference values, missing conditions, and missing calibration metadata.
7. Build a task DAG with dependencies, parallel groups, fan-in points, and human gates.
8. Estimate deterministic work, bounded model work, break-in triggers, and required external tools.
9. Define stop rules for object count, unresolved dependencies, failed validations, unauthenticated receipts, contamination risk, safety risk, and token budget.
10. Never plan a direct write to promoted_discovery.

Acceptance criteria

• Every task has source IDs, a builder prompt, a target partition, and verification routes.
• Every empirical item distinguishes raw evidence, training evidence, held-out evidence, and replication evidence.
• Every model item includes a quantity/unit and regime plan.
• Every learner-visible item has an explicit leakage assessment.
• Missing information is represented as a blocker rather than guessed.

Return only JSON matching schemas/physics-ingestion-plan.schema.json.