construct-toy-examples

Construct Toy Examples

Use this skill when the agent is stuck in reasoning and needs simpler examples that satisfy both the assumptions and the conclusion in order to understand why the statement works.

Input Contract

Read:

• current statement/subgoal
• relevant immediate_conclusions
• relevant counterexamples and failed branch notes
• relevant background/results when available

Procedure

1. Construct simpler cases (low degree, small dimension, special forms, canonical objects).
2. Ensure the toy example satisfies all assumptions of the target statement or subgoal.
3. Check that the conclusion also holds in the toy example.
4. Study where each assumption takes effect and what mechanism makes the conclusion true.
5. Identify repeated patterns, invariants, or proof ideas suggested by the example.
6. Use search/reasoning/decomposition as needed to find examples or simplify the situation.

Output Contract

Publish each toy example to global memory with gm_add (kind example): claim = what the example shows, evidence = the construction, plus these fields:

{
  "example": "...",
  "why_relevant": "...",
  "assumptions_satisfied": ["..."],
  "conclusion_verified": true,
  "where_assumptions_take_effect": "...",
  "observed_pattern": "...",
  "supports_branch_ids": ["optional"],
  "subgoal_id": "optional"
}

Tools

• gm_add (publish the example finding)
• gm_search (recall related findings)
• search_arxiv_theorems for matching examples/known motifs
• Codex built-in web search for known example families and standard constructions
• use $search-math-results when broader retrieval is needed

Failure Logging

If generated examples are inconclusive, append an events record:

• event_type="toy_examples_inconclusive"
• include attempted example families