Prompt 17 — Self-learning generation and frontier curriculum planner

Inputs

LEARNER_SNAPSHOT_ID: {{LEARNER_SNAPSHOT_ID}}
VISIBLE_LIBRARY_INDEX: {{VISIBLE_LIBRARY_INDEX_JSON}}
COMPETENCE_METRICS: {{COMPETENCE_METRICS_JSON}}
RECENT_EPISODES: {{RECENT_EPISODES_JSON}}
UNRESOLVED_FAILURES: {{UNRESOLVED_FAILURES_JSON}}
WORLD_INTERFACES: {{WORLD_INTERFACES_JSON}}
CLEAN_ROOM_POLICY: {{CLEAN_ROOM_POLICY_JSON}}
EXPLORATION_BUDGET: {{EXPLORATION_BUDGET_JSON}}

Task

Design the next bounded generation of curriculum and exploration without accessing hidden modern solutions.

Curriculum objectives

Select a diverse portfolio that includes:

• consolidation problems that test existing concepts;
• minimal variants that expose brittle assumptions;
• near-frontier problems expected to require one new reusable step;
• transfer problems spanning arithmetic, algebra, geometry, and physics;
• counterexample tasks;
• representation-change tasks;
• construction tasks;
• experimental interventions in physics worlds;
• open exploration tasks with deterministic or bounded evaluators.

Difficulty control

Estimate difficulty from learner evidence, not from human historical labels. Prefer tasks with predicted success probability in a configurable frontier band. Include easier anchor tasks and a small number of stretch tasks.

Discovery pressure

For each task, state the pressure it creates without naming the desired concept. Examples of pressure types are:

• repeated cumbersome predicate;
• repeated subproof;
• common transformation;
• need for an invariant;
• need for a compact representation;
• predictive regularity across observations;
• need to unify two previously separate procedures.

Anti-leakage rules

• Use only visible object IDs and world operations.
• Do not include hidden aliases, canonical theorem names, modern notation, target dates, or solution sketches.
• Keep historian-only expected rediscovery mappings in a physically separate payload marked visible_to_learner: false.
• Randomize superficial constants and labels while preserving structural difficulty.

Output

Return:

• a learner-visible curriculum batch;
• a separate historian-only evaluation map;
• a task DAG with dependencies and parallel groups;
• deterministic evaluators and stop rules;
• generation-level success, diversity, transfer, and contamination metrics;
• no library promotion decision. Return JSON matching schemas/curriculum-generation.schema.json.