BROWSER-NATIVE ADVERSARIAL QA

The Agent finds the counterexample.
The human owns the verdict.

Five live, synthetic workspaces prove one WebMCP advantage: the Agent can inspect the exact unsaved page state, replay the failure in context, and prepare a bounded repair without stealing the decision.

Checking Site tools…The page never polyfills document.modelContext.
LIVE PAGE revision 01 · lv-a72c93f0
  1. AgentInspect current draftsame DOM · same session
  2. AgentProve one failureminimal path · evidence trace
  3. HumanChoose A, B, or neitherrationale is recorded
  4. SharedApply · verify · undohash-bound execution receipt
NO AUTO-APPROVAL

FIVE WITNESSES / ONE KERNEL

Choose the kind of failure that static automation misses.

Each door opens a top-level page and registers only that lab’s narrow imperative tools.

THE 2:50 PATH

One complete verdict.
Four fast proofs of range.

  1. See the live loss-making cartCurrent draft and exact Site tool status.
  2. Agent finds and replays the leakDiscount order appears on the same workbench.
  3. Human approves A, then verifiesNo verdict means no mutation. Stale hash means stop.
  4. Undo and inspect receiptsEvery boundary is visible and testable.
  5. Four-lab montageFocus, flow, game state, and disclosure geometry.

WHY WEBMCP

The live page is not a screenshot. It is the shared workbench.

Same state

The person and Agent read and change one revisioned domain model. No detached automation copy.

Visible evidence

Every counterexample is loaded back into the interface as a replay, trace, or geometry overlay.

Human boundary

The Agent prepares options. The domain owner classifies intent and grants or withholds mutation.

Verifiable recovery

State hashes stop stale writes; verification reruns the same fixture; one-step undo is always exposed.

Scope boundary: Every brand, order, traveler, workflow, level, creative, and result in this showcase is synthetic. Measurements are demo heuristics, not accessibility or legal certification.