The Creation Frontier: Intent Recovery, Revision Cost, and Convergence to a Production Golden Master

Abstract

Generative AI may produce an artifact that a person approves without realizing the intention with which that person began. This conceptual paper asks how to distinguish recovery of an initial goal, valuable goal development, and concessions made because further correction is too costly when a collaboration reaches a production golden master. The proposed contribution is a trajectory-level evaluation protocol. It retains an initial private author rubric alongside subsequent approved revisions, evaluates artifacts against both, records the author's reasons for changes, and applies consistent production gates to an exact approved artifact. Prompts to golden master is reported with unsuccessful episodes and separate costs for human attention, money, elapsed time, and computational usage. Acceptance alone neither proves original-intent realization nor establishes concession: a changed goal may be a valuable discovery, and mixed or unknown motivations must remain explicit. Expression, capture, interpretation, and execution checkpoints support controlled investigation of recoverable mismatch. A finite decision-theoretic result separates an information floor from realization shortfall under stated assumptions; it is mathematical support, not a newly discovered law of creativity. A further elementary endpoint result shows that approval-based completion need not bound initial-goal realization when the rubric changes. Marginal recovery per resource unit links iteration to fidelity without equating mismatch with cost. A falsifiable hypothesis is that increasing correction cost, with initial goals and system opportunities held comparable, can increase author-reported concessions without a corresponding improvement in initial-goal fidelity. The paper positions this measurement proposal against prior work on steerability, iterative prompting, adaptation, and quality relative to interaction cost. It reports no experiment, causal effect, universal creation curve, or verified historical priority. Its contribution is a proposed integration and testable protocol requiring empirical validation.