[codex] Correct spending policy enforcement copy

dashboard/src/__tests__/policy-tab.test.tsx

@@ -85,6 +85,16 @@ describe("PolicyTab — rendering (Issue #47)", () => {
     render(<PolicyTab {...buildProps()} />);
     expect(screen.getByText(/Rosa Garcia/)).toBeTruthy();
   });
+
+  it("renders the accurate server-side enforcement copy", () => {
+    const { container } = render(<PolicyTab {...buildProps()} />);
+    const policyCopy = screen.getByText(/Enforced server-side by the agent/i);
+
+    expect(policyCopy.textContent).toMatchInlineSnapshot(
+      `"Enforced server-side by the agent before every payment. See docs/SPENDING-POLICY.md for how."`,
+    );
+    expect(container.textContent).not.toMatch(/Soroban smart contracts/i);
+  });
 });
 
 describe("PolicyTab — form interaction (Issue #47)", () => {

dashboard/src/components/tabs/policy-tab.tsx

@@ -77,8 +77,8 @@ export function PolicyTab({
         Spending Policy for {recipient.name}
       </h2>
       <p className="text-xs text-slate-500 mb-6">
-        These limits are enforced by Soroban smart contracts on Stellar. The
-        agent cannot exceed them.
+        Enforced server-side by the agent before every payment. See
+        docs/SPENDING-POLICY.md for how.
       </p>
       <form
         noValidate

docs/SPENDING-POLICY.md

@@ -10,6 +10,27 @@ CareGuard enforces five limits on every payment the agent makes:
 | `billMonthlyBudget` | $500 | Monthly cap for bill payments only |
 | `approvalThreshold` | $75 | Payments above this amount require explicit caregiver approval |
 
+## Enforcement Model
+
+The spending policy is enforced server-side by the agent before every payment. It is not enforced by a Soroban smart contract.
+
+Before `payForMedication` or `payBill` submits a payment, `checkSpendingPolicy` in `agent/tools.ts` reads the current policy, compares the proposed payment with the daily, monthly, category, and approval-threshold limits, and returns an allow/deny decision. The dashboard Policy tab updates that server-side policy; it does not create an on-chain rule.
+
+If a payment is denied, the agent should stop before submitting the payment request. If a payment needs approval, the agent should hold the payment until the caregiver approves it through the normal approval flow.
+
+## Threat Model
+
+This model protects against normal agent workflows accidentally or intentionally exceeding configured caregiver limits. It assumes the CareGuard server and its persisted policy/transaction data are trusted inputs to the payment decision.
+
+It does not protect against:
+
+- direct database or filesystem writes that alter the stored policy or transaction history
+- a compromised server process or deployment environment
+- direct use of the underlying wallet outside the CareGuard agent
+- a future service that submits payments without calling `checkSpendingPolicy`
+
+Any future on-chain spending-policy contract should be tracked as a separate design and implementation issue.
+
 ## Daily Limit and Timezones
 
 The daily limit resets at **local midnight** in the caregiver's timezone, not UTC midnight.
@@ -47,3 +68,7 @@ Limits are validated client-side and server-side:
 - `dailyLimit` must not exceed `monthlyLimit`
 - `medicationMonthlyBudget + billMonthlyBudget` must not exceed `monthlyLimit`
 - `approvalThreshold` must not exceed the smallest of `dailyLimit`, `medicationMonthlyBudget`, and `billMonthlyBudget`
+
+## Limits of Enforcement
+
+Policy checks are only as accurate as the stored policy and transaction history available to the agent. Restart-safe persistence and transaction logging are therefore part of the enforcement boundary, but they are still application-level controls rather than blockchain-level controls.