v0.1.0

The selection discipline, in one screen:

calibration sweep (tuning allowed here only)          held-out confirmation (tuning never)
   240.00 dB  CHEATER    <- tops the ranking             CHEATER    240.00 ->  -1.34  COLLAPSED
    23.31 dB  ista 5e-2                                  ista 5e-2   23.31 ->  22.31  REPRODUCED
    -0.00 dB  null: zeros  <- the floor, where it belongs

python examples/toy_loop.py — under 30 s, numpy only.

First release: 11 rules with their tuition, the anti-cheat checklist, go/no-go gates, the breakthrough loop, two runnable demos, and adapters for 9 agent stacks (DeepSeek Harness, OpenAI Codex, Claude Code, Cursor, Copilot, Gemini CLI, Windsurf, Cline, Aider) generated from a single source. What the pages claim, tests/ checks.