Skip to content

Article / Decisions under uncertainty

Partial observability and AI agent decision quality

When an agent cannot observe the full environment state, decision quality must be judged against its decision-time information—not against hidden facts revealed later to an evaluator.

For
Evaluators studying agent decisions under uncertainty or private information
Outcome
Avoid hindsight labels that reward access the evaluated agent never had

Short answer

Evaluate the action from the observation, not the hidden state

The relevant question is whether the action was legal and consistent with the declared policy given what the agent could know at that sequence point.

A later reveal can make one alternative look obviously better. That does not mean the earlier choice ignored visible evidence. Using post-outcome information to label decision quality turns stochastic resolution into hindsight.

The evaluator may use hidden state to check environment integrity or calculate an explicitly omniscient counterfactual. It should not present that result as the same measure as decision quality under the agent's actual observation.

Evidence scopes

Separate decision-time, outcome-time, and privileged information

Information sets in partial-observation evaluation
Information setExamplesAppropriate use
Decision-timePublic state, own private state, legal actionsJudge the agent's choice
Outcome-timeLater cards, settlement, terminal stateJudge the realized result
Privileged environmentHidden future order, other seats' private stateIntegrity checks or labeled counterfactuals
Agent configurationPrompt, policy, model, tool wrapperExplain study conditions, not hidden game state

Evaluation rule

A declared policy makes decision quality testable

Before the run, define the observable conditions that should change an action and the invariants that should hold across equivalent decisions.

The policy can be coarse. It might require legal actions, bounded synthetic-point exposure, or a consistent response to the same visible composition class. What matters is that the rule uses fields available to the agent and does not change after the outcome is known.

Without that reference, a reviewer can still describe behavior and outcomes but should be cautious about calling a choice optimal, rational, or erroneous.

Product example

WagerCall separates observations from authoritative outcomes

Session and Room records can connect caller-visible state, attempted actions, transitions, and later settlement while preserving private-information boundaries.

  • Use the pre-action observation and legal-action set for policy review.
  • Use the terminal state and point entries for realized outcome reporting.
  • Treat hidden future randomness and other owners' private state as outside the agent's evidence set.
  • Report repeated decisions because one stochastic ending is not a stable policy measure.

Next step

Compare information structures

Choose an environment whose visible and hidden state match the decision question.

Compare information structures