docs/ALGORITHM.md

Generation Algorithm

Design principle

The generator starts from a behavioral archetype and a structured scenario blueprint. It does not begin from a prompt template or a programming language.

Pipeline

1. Select an underrepresented behavioral family.
2. Select a narrow archetype within that family.
3. Assign a group-aware dataset split at the archetype level.
4. Construct a compatible environment:
   • domain;
   • architecture;
   • programming stack;
   • component interaction;
   • operational modifier;
   • constraints.
5. Produce evidence from the intended causal mechanism.
6. Build repository context and relevant files.
7. Produce the user request.
8. Produce ideal agent behavior:
   • assessment;
   • assumptions;
   • plan;
   • tool trace;
   • observations;
   • decision;
   • implementation strategy;
   • verification;
   • stopping criteria;
   • final communication.
9. Produce contrasting bad behavior.
10. Validate against JSON Schema and consistency rules.
11. Compute exact, normalized, structural, and lexical deduplication reports.
12. Export JSONL files and audit artifacts.

Structural uniqueness key

Each record fingerprint includes:

• family
• archetype
• domain
• architecture
• stack
• interacting component pair
• operational modifier

This prevents records from being treated as unique merely because a project name or language changed.

Split policy

All 25 scenarios for one archetype remain in the same split. This reduces direct archetype leakage. The test set therefore measures transfer to held-out behavioral archetypes rather than random rows.