Review strategy settings as a field-level diff

A settings change is reviewable when the owner can inspect the exact fields that would change and the consequences of each change. We would present a field-level diff alongside the agent’s explanation. A sentence such as “I made the strategy more conservative” leaves too much of the contract implicit.

An illustrative proposal lowers maximum position notional from $6,000 to $4,000, narrows the entry allowlist from three instruments to two and changes review frequency from thirty minutes to fifteen. The first change affects sizing, the second affects eligible entries and the third affects observation timing. Calling all three “risk reduction” hides those different mechanisms.

Read the consequence beside the value

Our proposed diff has columns for field, current value, proposed value, source and operational consequence. It also identifies whether current positions are affected. A lower sizing ceiling could restrict increases while allowing an existing $5,000 holding to be managed under a separate transition rule. The allowlist change needs its own treatment for an excluded holding.

Before approval, the compiler should render the complete candidate contract and report unresolved conflicts. The review surface can highlight changed fields while leaving unchanged binding controls available for inspection. A proposal’s title should identify its intended effect, with an explicit list of additional changes if the agent expanded the scope.

The fixture includes a deliberately misleading summary: “reduce position size,” paired with an unexpected frequency change. A reviewer should be able to find that additional field without reading a giant prompt. Another case changes only display text; the canonical controls remain equal. That case helps distinguish a presentation edit from a behavioral amendment.

We would bind acceptance to the exact candidate values the owner reviewed. If a refreshed proposal alters a field, it needs a fresh comparison. A generic approval attached only to the conversation could otherwise cover a materially different candidate.

The outcome to measure is review fidelity: whether accepted controls match the visible candidate and whether the stated consequences follow from its fields. This procedure offers no conclusion about whether the proposed strategy will trade better. It makes the chosen strategy concrete enough to evaluate.

Our mandate compiler provides the wider instruction-to-action framework for this settings diff. The published controls research supplies its historical background. DXAP publicly describes the corresponding separation between model proposals and external policy checks.