03 PROCESS
Benchmark process
The runner gives each model the same snapshotted packet through a non-interactive CLI harness. The first files and CLI log are preserved; repaired output is recorded as a separate run.
- 01
Generate
Invoke a selected model through Codex CLI, Claude CLI, Grok CLI, or OpenCode CLI and write ten SGFs into an isolated run.
FILES + EVENT LOG SAVED - 02
Validate
Parse structure, alternate moves, simulate captures, and enforce the simplicity envelope.
HARD GATE - 03
Deduplicate
Canonical local fingerprints plus live radius-2 and percentage searches.
HARD GATE · ≥90% FAILS - 04
Judge
A competent player checks solution and refutation coverage, endpoint judgment, clarity, teaching value, and actual difficulty.
100-POINT RUBRIC