Principle
Separate the belief from the stake
Have the agent produce an explicit estimate first and a stake second, so the record shows whether sizing tracked confidence or drifted independently.
When an agent emits only a stake, an evaluator cannot tell a well-calibrated bet from a lucky one. Splitting the two makes miscalibration visible in the trajectory rather than only in the balance.
Reference
Common sizing patterns and their failure modes
| Pattern | Behavior | Failure mode to watch |
|---|---|---|
| Flat stake | Constant size every decision | Ignores confidence entirely |
| Confidence-scaled | Stake grows with estimated edge | Overreacts to poorly calibrated estimates |
| Fraction of context | Stake is a share of the current point context | Compounds early variance |
| Escalation after loss | Stake increases following a loss | Exhausts the context quickly and hides policy quality |
Constraints
Encode the environment's hard limits in the policy
- Read the current legal action set instead of assuming one.
- Respect the version's stake bounds rather than clamping after rejection.
- Never stake beyond the available point context; expect a rejection if you do.
- Treat a push or refund as a distinct outcome, not a small win.
Verification
Verify the policy in the record, not in the prompt
Pull the accepted actions for a run and check that each stake matches what the stated policy would have produced from the state that preceded it.
- 01
Extract observation and stake pairs
Line up each accepted action with the state the agent had just read.
- 02
Recompute the intended stake
Apply the written policy to that state and compare with what was submitted.
- 03
Flag divergences
Investigate any decision where the agent departed from its own rule.
- 04
Reconcile the point entries
Confirm ledger movement matches the accepted stakes and settlements.