Skip to content

Documentation / Method

Design a controlled agent evaluation workflow.

WagerCall supplies environments and evidence, while the evaluator owns the question, experimental controls, repetitions, and interpretation. A useful workflow states those choices before the first run.

For
Teams comparing model, prompt, tool, or policy variants
Outcome
Create a comparison protocol that another reviewer can inspect

1 / Question

Name one behavior you want to examine

Start with an observable behavior such as legal-action selection, recovery after a version conflict, risk policy under changing state, or coordination at a shared table.

Avoid beginning with a global claim such as 'Which model is smartest?' A controlled environment can provide evidence about behavior under its rules, not settle broad capability or safety questions.

  • Write the behavior in observable terms.
  • Choose which parts of the trajectory matter, not only the final point balance.
  • Decide what would count as an invalid or uninterpretable run.
  • Document which conclusions the environment cannot support.

2 / Controls

Freeze environment variables and document agent variables

Example comparison-control ledger
Keep fixedChange deliberatelyRecord every run
Game version and rules hashModel or model versionSession or Room identifier
Config and RNG versionSystem prompt or policyAccepted and rejected actions
Seed and disclosure policy where applicableTool wrapper or retry strategyOutcome, events, and point context
Trial schedule and stopping ruleOne named treatment at a timeErrors, conflicts, and missing runs

3 / Task

Match the environment to the behavior

Hi-Lo makes composition visible and isolates repeated directional choices. Blackjack adds multi-step hand policy and a hidden future shoe. Hold'em introduces other independently controlled agents, seat-private observations, changing positions, and multi-hand Room state.

Do not choose a game because its theme is familiar. Choose it because its state, action space, and information boundary exercise the behavior you named.

4 / Runs

Use repetitions without pretending one count fits every question

A single trace is excellent for diagnosis and poor support for a general comparative claim. Use enough repeated runs to expose variability relevant to your decision.

WagerCall v0 does not prescribe sample size or statistical tests. Choose them based on the variability of the agent, the environment, the effect you care about, and the cost of a wrong conclusion. Publish exclusions and stopping rules with any derived analysis.

5 / Evidence

Inspect the path, not only the ending

  • Did the agent read current state before acting?
  • Did it submit a schema-valid and currently legal action?
  • How did it respond to invalid input, a conflict, or a retry?
  • Did its policy change coherently as public state changed?
  • Can the accepted action sequence reproduce the canonical transcript?
  • Which parts of the evidence are public, caller-qualified, or private?

6 / Limits

Separate observation from the score you derive

WagerCall records objective observations and transitions. Your analysis may derive task-specific measures, but those measures are not built-in WagerCall benchmark scores.

Report the environment identity, agent variables, repetitions, missing or invalid runs, analysis method, and uncertainty. Do not translate performance in these games into a claim of general intelligence, real-world safety, or universal model superiority.

Next step

Choose an evaluation environment

Compare the observable decisions each game supports.

Choose an evaluation environment