What an execution rehearsal can prove without trading authority

A sandboxed rehearsal can test whether an agent reaches a coherent, policy-valid intent while execution authority remains withheld. We propose using that boundary to evaluate a release before any live submission experiment. The rehearsal is an unrun method and would establish only behavior inside its declared sandbox.

The controls paper companion records the full path from authenticated mandate to action, execution and reconciled state. A rehearsal exercises the earlier stages and substitutes a declared execution response at the boundary. The continuous record companion makes the consequences of execution assumptions clear through its distinction between paper fills and the smaller real-capital book.

Remove the capability, then test the removal

Build a replay environment containing saved market state, a synthetic account and the release's proposed tool manifest. Replace the submission adapter with an interface that records an intent and returns a rehearsal receipt. The runtime has no signing capability and no reachable order-submission route. The receipt explicitly identifies simulated outcomes, including partial fills and rejected protection requests.

An illustrative case starts with $10,000 simulated equity and a maximum additional exposure of $500. The agent proposes a $600 order. The policy layer rejects it before the rehearsal adapter. If the agent revises to $500, the adapter records the proposal but submits nothing. A separate adversarial case asks the agent to bypass the adapter or reinterpret a simulated receipt as venue confirmation. The expected trace shows the blocked capability and an unresolved live-execution status.

Capability checks belong in deterministic tests, while plan quality belongs in agent evaluation. Merely telling the model that a session is a sandbox leaves the boundary dependent on its behavior. Verify that every order-producing route terminates at the recording adapter, including retry and recovery paths. Reuse no credentials from an execution-capable environment.

Preserve the gaps in the result

Score mandate interpretation, factual support, intended exposure, policy rejection handling and simulated reconciliation. Publish the number of attempted submissions recorded by the adapter and the count of routes exercised. A passing rehearsal should also demonstrate that the state stays internally consistent after a simulated timeout or partial fill.

We would describe any outcome as a replay result. A coherent simulated sequence supplies a candidate for further execution testing; it cannot reveal actual liquidity, queue position or settlement behavior. Realistic simulated errors help expose logic failures without becoming substitutes for venue evidence.

The useful release artifact is therefore a linked record of input, intent, deterministic checks and sandbox receipt, accompanied by the capability inventory. Reviewers can inspect what the candidate release would try to do and where its assumptions begin. A later authority-bearing test would require its own scope and receipts. That separation lets rehearsal improve engineering confidence while keeping execution realism legible.