|
|
cd9817b9ec
|
feat(agent): model theorem matrix with degenerate sector analysis
Each model has strong (>0.7), mixed (0.4-0.7), and degenerate (<0.4)
sectors. For each degenerate sector, a theorem attack compensates.
Algorithm: greedy selection until all priority sectors are strong (≥0.7).
Results:
- lean_proof: DeepSeek V4 Pro (1 model covers all priority sectors)
- code_bug: Kimi K2.6 (1 model)
- architecture: Claude Opus (1 model + 2 theorem attacks)
- math_novel: DeepSeek V4 Pro (1 model)
- infra_debug: GPT 5.5 + DeepSeek Flash (2 models)
- hard_wall: Claude Opus + DeepSeek Flash (2 models + 4 attacks)
Theorem attacks for degenerate sectors:
- speed → parallel_delegation (delegate fast tasks to cheaper models)
- cost_efficiency → selective_invocation (escalation path)
- formal_proof → delegate_to_opus (can't do Lean, delegate)
- breadth → context_injection (inject codebase context)
- reasoning_depth → chain_of_thought (break into steps)
|
2026-06-22 16:16:13 -05:00 |
|
