When building AI systems, designers repeatedly choose between expanding a model (making it larger, improving prompts, fine‑tuning) and decomposing a task into deterministic sub‑tasks, leaving the model only with the irreducibly fuzzy residue. This paper makes three claims. First, many tasks that are usually handed to a model in full are actually deterministic except for a tiny generative core. Second, and the central empirical finding, decomposing such a task yields a measurable property we call parity by construction: on the extracted sub‑tasks the performance curve across model sizes flattens because no model outperforms another. We demonstrate this across nine models ranging from a 494 M‑parameter local checkpoint to a frontier Opus‑class model—a thirty‑fold parameter span within the local family. Once the host resolves the context, operation‑selection scores are identical for all nine models, and selector resolution matches for eight of them; the only task that does not decompose, labeling, remains sloped across the same range. Because the task is off‑loaded to a deterministic host, a stock checkpoint suffices—no task‑specific training or fine‑tuning is required. Third, this is a general design discipline rather than a domain‑specific trick. The same pattern recurs in physics setups, motion retargeting, object‑model editing, and a Web‑CLI tool family (same author, not an independent replication). We provide a method to locate the cut, a construction that guarantees its existence (a deterministic surface with a closed grammar), and a way to keep the residue safe (resolve deterministic context, verify generative output). The work is positioned against neuro‑symbolic AI, tool‑augmented language models, and generate‑and‑verify approaches, with a clear distinction between measured domains and illustrative examples.
Review