The Harness Is Not the Model: How Far Scaffolding Takes a Weak LLM(github.com)

A practitioner-led synthesis by Manazir Ali. Formalizes the bet that a swappable weak model wrapped in a strong harness (deterministic retrieval, curated tools, verification loops) can deliver frontier behavior, decomposes it into five testable components, and grades each against roughly 30 published sources. The harness closes knowledge, format reliability, and narrow verifiable tasks, and is worth roughly one model generation from a naive baseline. It does not close multi-turn coherence, long-horizon autonomy, or novel hard reasoning, where same-harness model spreads run 1.5 to 4x on agentic coding and 50 to 100x on METR time horizons. The engineering answer is tiered routing with frontier escalation.