The Value of What the Model Knows: A working note on governed knowledge as a variable in the economics of AI tasks

Cordier’s per-task value equation has no knowledge variable. It derives the value of a token from task success, oversight, and the worst outcomes, and reads task success off a benchmark as a property of the chosen model, while the benchmark record shows that a part of it moves with what the model is given to act from. We restate the equation and write it with a knowledge state K, typed as a pair: the content served to the agent, and the governance metadata on each claim, its review status, permitted use, and stratum. Task success, detection, review cost, and token spend become functions of K, each tied to the coordinate it responds to. The share of acts the check refuses and the price of taking a refused act come from the companion note on the mandate, which prints the same display at a fixed knowledge state, and the upkeep of the knowledge enters as a stock cost, with a condition for when the stock pays for itself. Cordier’s equation is the case with no refusal and no upkeep. We state the set of knowledge states at which a model clears break-even, which can be empty, show that reusing one reviewed claim across many tasks lowers its upkeep and concentrates the tail in the same step, and place the classical value of information on both sides of the workflow: the firm keeps its valuation identity, the reviewer is tested against the optimal use of what he is shown, and the theory restricts no sign for the model’s response to more knowledge. A paper follows, with a factorial that separates actual review, the visibility of the review label, and enforcement. Publication status: This deposit is a preliminary working note, not a peer-reviewed article. It develops a conceptual and economic framework and proposes an experimental design. The proposed benefits of governed knowledge remain hypotheses to be tested; the note does not report results from the proposed experiments.