A Structural Theory of Harnesses: Why Generalized Intelligence Lives in the Arrangement Around the Generator

Version 1.1 (21 July 2026) — correction to §9.2. §9.2 has been corrected against the shipped implementation it describes. Version 1.0 named two of the immune system’s three components “anti-inbreeding defense” and “principle demotion,” and described the first as rejecting a graduation whose explanation is “merely a verbatim restatement of the episode.” That description was incorrect: the explanation-grounding check is a lower bound — an explanation must share at least two meaningful words with its cited episode — so a verbatim restatement passes it rather than failing it. The mechanism that does catch self-confirmation operates across sessions, comparing a pattern’s current explanation against its prior one, and shipped in anneal-memory 0.3.2 on 2026-05-21, in the same release that removed both retired names from that project’s README for describing behavior no code path implemented. §9.2 now describes the three components under the names the shipped software uses. The argument of the section is unchanged. Readers citing v1.0 on the immune system’s mechanism should cite this version instead. The AI industry has, over the last eighteen months, shifted the center of its engineering investment from scaling base models to building the scaffolding around them. The shift has a name now — harness engineering — given to it publicly by the Claude Code leak of March 2026 and formalized as a discipline within weeks by Red Hat, LangChain, Letta, AlphaSignal, and the career market. What the practice does not yet have is a theoretical account of why the engineering it has chosen to invest in is the only kind of engineering that could have worked. This paper provides that account. A harness is defined functionally: the arrangement that closes the feedback loop between a generator and reality beyond the generator's training distribution. Under this definition, generalized intelligence — the capacity to transfer, accumulate, and coherently update across tasks and domains — is a property of the arrangement that closes these loops, not of the generator inside them. Raw generators are narrowly intelligent within their training distribution; generalization is the harness's job. The paper develops the theory across thirteen sections with convergent biological and frontier-AI evidence, a mechanism catalog of six harness properties, three production failure modes predicted by missing loops, and anneal-memory as a constructive existence proof. Three consequences are developed: AGI is a harness engineering problem rather than a model scaling problem; alignment lives at the harness layer rather than the weights layer; and cognitive sovereignty is a side effect of harness design, not a separate goal.