Which opportunities deserve a limited research budget?

A fixed research budget creates an allocation problem before the agent chooses a trade. We propose comparing how several allocation policies distribute that budget across an identical opportunity set. The aim is to measure what gets examined, what remains unknown and what the examined cases produce under a fixed decision policy. This experiment has yet to run.

The continuous record companion reports a historical selection effect at the candidate-render boundary. That observation concerns visibility. Research allocation introduces a distinct boundary: two candidates can both be visible while only one receives expensive investigation. The controls paper companion provides the trace linkage needed to connect the allocated research to the eventual typed action.

Account for the opportunities left untouched

Suppose an illustrative snapshot contains twenty eligible instruments and the turn permits 4,000 research tokens. Equal allocation grants 200 tokens per instrument. A concentrated policy might investigate four instruments at 1,000 tokens each. Both spend 4,000 tokens, but their coverage is 100% and 20% respectively. Coverage alone says little about answer quality: a 200-token pass may discover less than a focused investigation.

Define examination in advance using a concrete task, such as checking one event source and one liquidity condition. Count attempted, completed and source-supported examinations separately. An unanswered request consumes budget even when it produces no usable information. Include synthesis and retries in total cost, so the concentrated policy cannot hide the overhead of combining specialist outputs.

The frozen opportunity set should be created by a rule independent of subsequent price movement. Evaluate every policy on the same snapshots. Keep candidate order randomized and retain the random seed in the manifest. Compare equal allocation with a simple cheap-screen-then-investigate policy and a model-directed policy. An expensive allocation strategy earns its place through additional decision value relative to the simpler alternatives.

An allocation frontier, with denominators

Report the fraction of eligible opportunities examined, factual support per completed examination and decision outcomes over all eligible cases. Conditional performance among investigated cases can improve while total useful coverage falls. A model that inspects only familiar instruments may look precise while missing the cases that motivated the research budget.

For an illustrative scoring ledger, twelve supported examinations out of sixteen attempts gives 75% support among attempts. Against twenty eligible instruments, supported coverage is 60%. Both percentages belong beside the token bill. Attach each research output to the opportunity and receipt time so information arriving after the decision is counted as unavailable for that decision.

We would use these measurements to choose the next allocation experiment, rather than infer a universal token budget. The useful contribution is a budget-to-coverage curve with a declared opportunity population. Economic evaluation can then ask whether another unit of research changes a decision enough to pay for its acquisition and processing.