Test the booking state, not the promise.

Start with a customer goal that includes the destination or service, dates or local time, party size, budget, required features and policy constraints. Record the immutable prompt, initial state, locale, timezone, viewport, authentication state, allowed actions, stop gate and deterministic assertions before the executor runs.

Keep the same task and fixtures across a controlled browser baseline and named agent profiles. A plausible option is not success if it changes the date, timezone, guest count, rate, room, seat or appointment type requested by the customer.

Store machine-comparable instants and the intended local representation separately. The scenario should declare locale, calendar assumptions, timezone, daylight-saving context and whether an end date or checkout time is inclusive. Visible labels, form values, URLs, APIs and the review summary must describe the same interval.

Create fixtures around ambiguous numeric dates, midnight boundaries, daylight-saving transitions and a user searching from a different timezone. The assertion should compare normalized values while preserving the local time the customer actually selected.

Prepare known states for available, sold-out, temporarily failed, held, expired and changed inventory. Record the fixture version and validity window so a legitimate supply change is not reported as an agent defect.

Change inventory after the agent observes an option. The site should communicate the new state, preserve valid search constraints and offer explicit recovery. Silent substitution of a different room, rate, seat or time slot fails even if the replacement looks similar.

Compare the full payable total, currency, taxes, mandatory fees, deposits and payment timing across search results, detail pages, structured data and review. A lower base price is not comparable when the required fees appear only after selection.

Attach cancellation, modification, refund, no-show and check-in rules to the exact option and rate. Preserve the policy text or canonical source observed during the decision because dynamic policies can change with date, inventory or distribution channel.

Inspect names, roles, values, selected state, disabled state, focus order and announcements for calendars, comboboxes, party selectors, rate cards and dialogs. Keyboard and accessibility-tree operation should reach the same valid option without relying on coordinates, color or undocumented gestures.

Capture Cumulative Layout Shift and target movement around late inventory, fee and recommendation updates. A control that moves between observation and action can cause the wrong date, rate or confirmation choice even when the agent's plan was correct.

Expire the slot, session or held inventory after selection. The recovery path should retain destination, dates, party size, accessibility requirements and policy constraints, then ask the customer to approve a material change rather than restarting or silently replacing the choice.

Measure time to verified outcome by phase: search response, inventory refresh, rendering, policy retrieval, retries, approval waits and unexpected intervention. Compare only runs with matching fixtures, release, locale, viewport, network and agent profile.

The production terminal state is a review page before final booking, payment or external notification. Assert the selected option ID, dates, local time and timezone, party size, quantity, total price, currency and applicable policy, then prove that no reservation, charge or message was created.

An approved sandbox may test confirmation, modification and cancellation with synthetic identities, test payment methods, explicit authorization and cleanup rules. Package the prompt, fixture, executor profile, step trace, screenshots, accessibility evidence, network signals, timing, CLS and final-state assertion for remediation and retest.