I ran an effort bake-off. Three reasoning-effort settings - high, xhigh, max (or equivalent). I was surprised that all three scored 79.2% recall on 24 seeded bugs. The only difference was time spent, which increased with effort. Bugs were across 3 repos, 5 blinded adjudicators, 209 graded findings, 0 false positives. The max run took 4.5x the wall-clock - roughly 54 minutes vs 12 - and bought one extra partial. Recall stayed at 79.2%. The Claude-family caught 6 of 6 blatant seeds - including a committed, live-looking API key that all three OpenAI-family arms looked at and dismissed as a placeholder - but only 4 of 7 subtle ones. The OpenAI-family ran the reverse - 7 of 7 subtle, as few as 3 of 6 blatant. Different families missed different bugs. The winning config is both families at lower effort. Run in union, they caught 23 of 24 seeds in about 25 combined minutes (~14min parallel). One model at max effort got 79.2% in 54 minutes. I also tested 39 escalations to a bigger model on stuck problems. The bigger model helped about 3 times in 8. So there's still nuance in when and how to escalate. Still testing what works best. In my daily setup now, I'm running implementation on Opus, high effort with cross-family audits and loops with guardrails to escalate to redesign. Anyone else have tests they can share, or setups? Or a place to look for community results?

Claude Code loads the first 40,000 characters of CLAUDE.md for every session and silently drops the rest. Make each one matter, and enforce the count. 1. Count: wc -c CLAUDE.md. (Mine was at 47,851 - and the dropped rules were about not trusting a green gate) 2. Relocate: reference material, keep every rule. 3. Block: the regression with a pre-commit hook to refuse any commit that exceeds the 40k. If you have a doc-only fast-path hook, make sure you exempt CLAUDE.md. Any other improvements to consider?

So far, I'm seeing the 70% problem as real. But AI can still help fill in the remaining 30% if you build the machinery: basic testing and confirmation, tracking to-do and testing lists, cross-family audits. The human-in-the middle still has to decide how it should work, test the edge cases and push the AI to finish properly.

24 hours of Fable 5 auditing the app I've spent 4 months building with Opus 4.5 and 4.6 - 553K LOC, ~4,500 tests. I spec every feature. I run adversarial reviews and blind agent reviews on every implementation round. It still found 10 P0s. The worst: my AI code reviews were approving code they couldn't actually read. And 30 test failures were invisible because the gate command never ran them. Code generation is solved. Verification isn't. I'm building for exactly that - verification that doesn't depend on trusting any single model.