agentic-driving-weaker-modelslisted

# Driving weaker models — the playbook **The central claim:** a weaker model inside a strong verification harness approximates a stronger system. Cheaper-tier dry runs passed inside the harness — skills, gates, review, paired tests — while in the same period the strongest available model authored three plans that cold review raised 32 findings against, 4 of them CRITICAL. The harness is not a patch for weak models; it is the system, and every tier needs it. Tier buys fewer errors, not zero. What changes by tier is how much structure the brief must carry and where the human checkpoint goes. Observations, evidence base, and provenance: `examples.md`. --- ## 1. The gap taxonomy — where cheaper sessions actually differ Observed, not assumed. Each gap: the evidence, then what it predicts. **G1 — Trusted docs are treated as ground truth; the verify instinct doesn't fire.** An agent repeated a stale skill claim verbatim rather than spending one `git ls-files` to check it, while otherwise executing doctrine perfectly. Predicts: any error in your skills, briefs, or specs propagates straight into output. Corollary: skill and spec correctness is load-bearing — review and dry-run-validate them. Seven skill patches came out of four dry runs. **G2 — Fluent, unresolved self-contradiction in analysis prose.** A debugging run delivered a correct fix wrapped in a diagnosis that claimed an impossible thing, noticed the tension mid-paragraph, and moved on without resolving it. Predicts: