Your AI Isn't Smart — Your System Is

During rotation calibration with 11 real figure skating videos, the system diagnosed Nathan Chen's textbook-perfect triple Axel as 'high under-rotation risk.' It was about to tell users that an Olympic gold medalist's signature jump was wrong.

2D camera projection created an artifact: blade angle appeared 99 degrees off from body angle. The algorithm interpreted this as 'still rotating at landing.' In reality, it was projection distortion -- not biomechanics.

'Better to say nothing than to say something wrong.' Confidence gating requires rotation_confidence >= threshold before activating diagnosis. Threshold raised from 0.60 to 0.70 -- Nathan's 0.61 correctly suppressed.

The ultimate diagnosis pipeline was connected to production routing. 732 tests passed, 0 failures. Everything green. The natural reaction: ship it, announce Phase 7 complete.

Claude cited a historical precedent -- the Gold Gate incident from a previous phase where eval passed but production broke. The SOP system has memory. It doesn't repeat mistakes.

'When a solution might have issues, raise doubt rather than silently continue.' And: 'Before claiming done, ask yourself: verified or assumed?'

Claude wrote a new protocol and saved it to project-level memory. But global CLAUDE.md already stated: 'Cross-project rules must be written into global CLAUDE.md.' Claude had the rule, read the rule -- and still didn't execute the self-check.

AI models can read rules, understand rules, even explain rules. But they cannot 100% reliably self-enforce during complex multi-step tasks. Attention drifts. Context competes.

This failure proved a critical insight: prompt-level instructions aren't reliable enough. This is exactly why enforcement hooks exist -- they don't trust the AI to 'remember,' they force compliance in code.