[r3dbars]

Summary

Q14 ran as two paired campaigns (Gemma 4 E2B, Qwen 3.5 9B; conservative vs tuned display stack) on the certified-v2 development partition, 3,600/3,600 evaluations each, reviewed and compared with a 10,000-iteration root-clustered bootstrap.

• Gemma: bad-when-shown 76.6% → 57.1%, improvement 0.19 [0.14, 0.25], P(>0) 100%; useful displays 73 → 199
• Qwen: 61.1% → 30.8%, improvement 0.30 [0.26, 0.35], P(>0) 100%; useful 373 → 747
• Guardrail (95% useful retention) held by a wide margin in both
• Registered hypothesis (Gemma gains at least as much as Qwen) rejected: intervals do not overlap. The tuned stack's direction is generator-independent; its magnitude is not.
• Every arm fails the bad-suggestion and latency hard gates under the 8-worker campaign runtime, so nothing promotes from this pair, as registered; the standard validation path with a Gemma control starts from here.

Docs only: Q14 result section, lab log entry, Learning Ledger entry (tuned-display-stack-helps-gemma-at-half-qwen-gain, counts re-pinned 47/40).

🤖 Generated with Claude Code