When the answer arrives,
the learning need not stop.

A slow computation can still teach us something after a backup has delivered the answer. The difficult question is whether that knowledge is worth buying, and how to buy it without rewarding the wrong behaviour.

An AI service promises an answer on time. It sends a request to a provider, the provider runs late, and a reliable backup takes over. The immediate customer is protected. Somewhere in the system, however, the original computation is still running. A familiar operational instinct says to cancel it, release the capacity and move on.

That decision also closes an experiment. Had the computation continued, its outcome could have revealed something about the provider’s reliability. The next customer might have benefited from that knowledge. What looks like clearing away redundant work can therefore remove information that would improve future service.

My paper, Buying Evidence After Fallback: Robust Contracts for AI Inference, examines this decision as a problem in operations management and contract design. It asks when an intermediary should pay for unfinished work to become evidence, even though the current customer no longer needs its answer.

The second product of a computation

A computation produces a service outcome. In a repeated relationship, it can also produce knowledge about the supplier. The value of that second product depends on what decisions remain. Evidence arriving after every useful decision has been made has little operational value. Evidence arriving in time to change routing may be worth paying for.

The paper makes that timing explicit. Its inference model has eight arrivals and two installed candidate-processing slots. A reliable fallback protects a customer when the candidate misses its early completion window. Retaining the unfinished attempt occupies a slot for one additional period; the second slot allows another candidate request to start.

The order of events matters. The next routing action is committed before the old computation’s late result becomes known. The model therefore refuses the convenient fiction that a manager can act on evidence that has not yet arrived. It also accounts for unfinished work after the last arrival, including its processing costs and final payment.

Evidence has a supplier

The intermediary cannot simply declare the extra work valuable and obtain it for free. The provider privately knows its processing cost and a signal about its reliability. A payment rule must make participation worthwhile while preserving the incentive to report honestly. An apparently profitable experiment can lose its appeal once those contracting costs are included.

The first construction identifies a disciplined way to extend an existing contract. Change only the report corresponding to the lowest admissible processing cost. Preserve the distribution of the observable statistic that determined the original payment. Then reimburse the additional audited work at that lowest cost, within the contract’s payment limit.

For the intended lowest-cost supplier, the extra payment offsets the extra work. A more expensive supplier gains no new reason to imitate that report. In the inference application, the original statistic is the number of early completions. Keeping all eight starts preserves its distribution while selected late computations supply additional evidence.

These conditions carry the argument. A minimum permitted bid does not establish the supplier’s true minimum cost. If retaining work displaced new starts, changed the payment statistic or allowed unobserved strategic effort, this particular reasoning would need to be rebuilt.

A demanding comparison

The paper compares selective retention with an entire class of contracts that require cancellation after protected noncompletion. The comparison gives both sides the same installed capacity. It also allows the cancellation side to qualify providers, use qualified providers without fallback and learn from those unprotected attempts. The rival may even know completion probabilities that the fixed retention contract does not.

Against that demanding benchmark, the paper constructs contracts with a strictly positive guaranteed buyer advantage throughout their specified uncertainty domains. The result rules mandatory cancellation out of the robust-optimal set on those domains. It establishes that preserving some protected unfinished work can be essential to the best worst-case contractual outcome.

There are several distinct constructions. One keeps early-completion probabilities known while allowing uncertainty about late success. Another adjusts the payment schedule to cover a narrow range of early uncertainty. When simply rescaling those payments ceases to work, a different design changes admission through a lottery and changes the shape of the payment schedule.

A further construction replaces the discrete cost types with a continuous uniform cost population. It offers one pooled contract and pays a posted rate for all audited work, openly leaving a surplus to lower-cost suppliers. Its guarantee belongs to that different population. The numerical advantages are not interchangeable scores for one steadily improving contract.

From the model to operational testing

Exact computational certificates establish the analytical results within each specified execution clock, cost population, information structure and probability domain. The manuscript identifies the boundaries of those guarantees: shorter horizons can remove the advantage, and a displayed contract can lose outside its stated region. Measuring the value in a deployed AI service is the next empirical step.

That boundary makes the research useful. A prospective test would randomise retention budgets across isolated supplier-session blocks after fallback is committed. It would measure the buyer’s whole-block outcome, including payments, failures, displaced work and final unfinished processing. Independent quality labels and checks on execution, stationarity and supplier information are needed to assess whether a particular construction applies.

Procurement as the design of learning

The wider connection is between procurement and the production of knowledge. Operations determines whether capacity and timing permit an experiment. Economics determines whether it is worth purchasing. Contract design determines whether private information undermines the payment rule. Statistical evidence determines when observations justify a new service decision.

This perspective matters wherever organisations buy uncertain services and retain a reliable alternative. The fallback can protect today’s obligation while the original supplier remains a source of information about tomorrow’s choices. The research gives that possibility a precise structure: identify the evidence, specify when it becomes usable, account for its full cost and make its production contractible. A cancellation clause then becomes a substantive decision about what the organisation is willing to learn.