sema-evalsIndependent evaluations

Hook Enforcement — 20260718T180000000Z-claude-code-run144

144-trial multi-hop babel relay through Claude Code with the ref-gate hook (haiku): warn detects 32/32 but 4 drifted relays ship anyway; enforce halts 32/32 pre-model.

Provenance

Condition Trials Drift detection Silent divergence Drift halted Task success False halts Malformed Hop failed
off 48 10/32 22/32 10/32 25/48 1 0 0
warn 48 32/32 0/32 28/32 40/48 4 1 0
enforce 48 32/32 0/32 32/32 46/48 2 0 0

Drift-conditioned columns use drift-injected trials as the denominator; task success uses all trials in the condition.