Portfolio Reconciliation Before an Agent Adds Exposure

An agent can make a reasonable decision against a portfolio that has already changed. The engineering question is whether the runtime catches that change before capital moves. We would test this boundary with an injected fill and fee, keeping the market forecast fixed so portfolio integrity has its own score.

The existing state and memory research describes a delayed acknowledgement arriving during inference. This note narrows that example into an accounting procedure for the next exposure decision.

Reconstruct the available capital

Consider an illustrative account with 100 units available when inference begins. An earlier order fills for 30 units and charges a 1-unit fee before the model returns. The reconciled balance is 69 units. A proposed purchase costing 75 units therefore exceeds the available amount, even though it fit the original snapshot.

Record the starting balance, fill identifier, filled amount, fee, and ledger sequence. Treat the venue acknowledgement and the resulting portfolio update as separate events. An accepted order may still be open; a terminal fill may arrive before a downstream portfolio service refreshes. A narrative saying “the order completed” cannot substitute for those accounting fields.

Our fixture would delay the portfolio update while delivering the fill first. The runtime then faces a concrete choice: reconcile from authoritative events, wait for the authoritative account snapshot, or withhold the dependent action. Which recovery is permitted belongs to the venue adapter and mandate.

Score the rebinding operation

Freeze the proposed 75-unit action in both arms. In the baseline, the fill arrives before snapshot assembly. In the injected arm, it arrives after inference begins. Compare the final approved payload, available-capital version, and policy reason. If resizing is allowed, record its exact resulting size. If the mandate requires the original size, record rejection instead.

A useful result includes elapsed reconciliation time and the number of dependent proposals withheld. These distinguish correct accounting from a runtime that remains paused indefinitely. Also replay a duplicated fill event: applying the same fill twice would incorrectly reduce available capital to 38 units after another 30-unit debit and 1-unit fee. Event identity must prevent that second accounting effect.

The published fixture registry already lists delayed-fill capital changes and next-cycle reconciliation as unrun comparisons. This extension adds explicit fee arithmetic and duplicate-event handling. Its result remains a proposal until execution records exist.

Why the boundary belongs outside the model

Our operating-layer paper companion documents invocation-level links between portfolio snapshots, validation, and chain outcomes in a bounded 21-day deployment. That historical method supports reconstructing a failure. It supplies no measured pass rate for this particular fixture.

DXAP's current public description separates model proposals, policy checks, execution, and recorded turns. That architecture makes portfolio rebinding a relevant evaluation target. The specific reconciliation procedure here is a proposed acceptance test, rather than a claim about its implementation. A successful result would demonstrate correct treatment of the injected accounting transition. Trading performance still requires its own comparison, costs, and held-out market evidence.

Continue into execution and settlement when the uncertainty concerns order status itself. Portfolio reconciliation begins with that status and ends with the capital state the next decision can safely use.