From the Inside Out: The Critic Harness and a Cyclinic Model of Accountable Prompting

This paper argues that prompt engineering has largely been practised “outside-in”: prompts are written, tested against a few examples, patched when they fail, and shipped. The authors contend that this surface-level approach produces predictable failure modes—rule accumulation, attribution loss, masked failures, and silent degradation—and propose the inverse: “inside-out” prompting, in which the accountability architecture is specified first and the prompt is derived from it. The paper introduces the cyclinic model, a recurring diagnostic loop of four phases—Articulation, Scoring, Adjudication, and Account—in which a claim is stated, scored against a baseline of truth, adjudicated with a gate, and recorded so that the next cycle begins from a corrected baseline. It presents the Critic Harness as an implementation of this model: an adjudicator that returns ALLOW, CONDITIONAL, DENY, or ESCALATE, with binding evidence requirements and a stated condition that would change the verdict. Severity is treated as a computed quantity using the Divergence Score D = P + S + W + A (0–12), mapped deterministically to Levels 0–5. The paper also defines an evidence floor enforced by citation resolution and sets out obligations of accountability: visibility, attributability, correctability, and ownership. The work concludes that the prompt is not the unit of work; the cycle that makes the prompt answerable is. Keywords: prompt engineering; accountability; AI safety; evaluation; prompt governance; cyclinic model; output contracts; Critic Harness