Iterative AI Production (IAP): A Human-in-the-Loop Research Partnership Protocol

What it is. IAP, the Iterative AI Production protocol, is a prompt tool. It turns an LLM AI chat window into a research partner instead of an answer service. It is for people with a real idea who want help developing it rigorously into something defensible: a paper, a framework, a set of testable predictions, a patent sketch, an executable test protocol. You bring the original insight and the final call. The AI brings speed, cross-domain reach, and relentless honesty about what is actually supported. Quick start: paste the IAP document into any capable AI and say "Use the IAP protocol below." The model configures itself from there: no setup, no plugins, no methodology to learn first. (Download the PDF first as optimized for humans. The Markdown is the same, but optimized for an AI to use.) Why it exists. Most AI-assisted research fails in two predictable directions. In the first, the human abandons a good idea at the first serious objection: the model says "that won't work because X," and the project dies there. In the second, the model produces confident-sounding claims with nothing retrievable behind them, and the human builds on sand. The IAP tool counters both with mechanisms, not warnings. Against premature abandonment: every objection is treated by IAP as a design constraint, not a verdict. When the model identifies a problem, the required next move is not "is this dead?" but "what architecture would satisfy X?" Ideas either die honorably against explicit kill criteria or become something stronger. Against confabulated confidence: IAP evidence tags are gated on what was actually retrieved this session, not on how certain the model sounds. The [C] Confirmed tag is mechanically unavailable to a model without web search, and the model must declare its tool access before any substantive work begins, so the human knows what evidence guarantees are real before building on them. What the protocol specifies. A mandatory capability declaration at session start. A comprehension gate in which the model restates the human's idea before proceeding, so a misread surfaces early. A seven-phase pipeline: mechanism generation, seed/abstract/bound, expand, memory management and run planning, pressure test, crystallize, and empirical bridging. A four-level evidence tagging scheme with mechanical gating. A self-called sycophancy check triggered by consecutive agreeable responses, as costume agreement is treated as a failure mode, not politeness. An architecture changelog that prevents silent drift as the idea moves. Adversarial review by a named hostile role rather than a generic skeptic. A nine-format artifact menu for crystallization. And an explicit list of the conditions under which IAP is the wrong tool and the human should just ask the question. The final phase treats the exit from the AI environment as an artifact itself: a concrete, executable Test Protocol with a pre-registered read-out specifying in advance what result confirms the framework, what refutes it, and what is ambiguous. Provenance. IAP was developed while producing, as an example, the Dominance Operating System framework and is the method named in that work's AI disclosure. It is published separately so the method can be cited, criticized, and applied independently of the framework that occasioned it. Status. IAP is a working protocol: a specification the author uses to run actual research projects, revised across two versions as failure modes surfaced in use. The Dominance Operating System framework was produced under it, carried from initial hunch to published preprint. That establishes feasibility, not superiority. IAP has not been tested in controlled comparison against unstructured AI-assisted research or conventional practice; no outcome measure has been defined, and claims about effectiveness reflect the author's operational experience rather than measured findings. Comparative evaluation is the outstanding empirical task. Comments, counter-cases, and reports of where the protocol failed are welcomed.