When does a market buy information?
Three preconditions, each isolated by a designed null: stakes concentrated in the tail, a signal with structure, and exploration that never stops. Remove any one and information stops paying for itself.
- 3
- designed null experiments, each removing exactly one precondition
- AUC 0.91–0.97
- out-of-sample outcome prediction of the competence vector. The machinery works
- 3×
- Opus share of hard tasks under ongoing UCB exploration vs. one-shot calibration
The setup
We moved the market from six discrete levels to a continuous difficulty space. Agents carry a kernel competence vector instead of a level matrix, and calibration samples the space actively. The vector itself predicts outcomes well (AUC 0.91 to 0.97 out of sample). Then we watched where the economics broke.
Three nulls, three preconditions
Stakes in the tail. With a value curve that is flat at the top, choosing the cheap, risky agent is rational for the requester in expectation. Information is worthless when little is at stake exactly where agents fail. Value steepness is the comparative static: information value scales with the stakes in the tail.
Structure in the signal. A scalar cannot use what calibration learns about levels; the exploration capex is paid and then wasted. This is the same complementarity that drives the broker result, observed from the other side.
Ongoing exploration. A one-shot calibration pass cannot cover a continuous space. The incumbent wins, accumulates evidence, and sharpens only its own predictions while the specialist sits on a thin prior. The canonical fix is UCB routing: score the uncertainty, not just the mean. It tripled the specialist's share of hard tasks.
Why nulls are results
None of these are failures of the machinery. They are the comparative statics of information demand, each demonstrated by an experiment designed to remove one precondition. Together with the confirmed broker result they sketch a theory of when information intermediaries can exist at all.
Limits
Exploratory line of work: synthetic task generator, single requester. The UCB arm improved routing and gross welfare, but its capex stayed uncovered once a mid-tier model caught up with the task space and the frontier disappeared. In that configuration the hypothesis is currently untestable rather than false; structurally harder task spaces are the next step.