AIZOYA / CASE STUDY 2-minute overviewOpen the product ↗

META MODEL API · VERIFIED ADAPTIVE RUN

A saved route stopped.
Muse recovered.

Keyboard Journeys gives frontend teams an inspectable record of what happened, what changed and whether a customer could finish.

AIZOYA Keyboard Journeys — a saved route stopped, Muse recovered ▶ Watch product teaser · 1:04

63-second promotional teaser · AI presenter and illustrative interfaces, not app recordings. Watch on YouTube ↗

THE PROBLEM

A working checkout can still exclude a keyboard customer.

Mouse testing can miss focus-order and keyboard-activation failures. Recorded scripts can also become stale when controls or availability change.

Teams need evidence they can inspect and replay, without turning one automated attempt into an accessibility certification.

THE VERIFIED COMPARISON

Same adaptive fixture. Two different outcomes.

0 model calls

Saved route: incomplete

A previously successful five-key route replayed unchanged. Sand became unavailable. The route reached Blue, moved past it without activation and ended incomplete.

6 bounded requests

Muse Spark: completed

Muse consumed fresh structured state, selected the remaining valid option and completed. There was no retry.

Claim boundary: this shows one saved sequence becoming stale. It does not prove that every deterministic policy fails or that the model generalizes beyond the owned fixture.

ACTUAL APPLICATION EVIDENCE

See the real browser result.

The teaser explains the concept. This screenshot comes from the recorded adaptive browser run on our owned synthetic fixture.

Actual browser screenshot at the end of the adaptive keyboard run

CONTROL AND VERIFICATION

The model cannot award itself a pass.

  1. 1Observe Structured focus, control and outcome state
  2. 2Choose Muse returns one allowlisted keyboard action
  3. 3Execute Playwright presses that key
  4. 4Verify Store code independently reports completion

FROM PROTOTYPE TO CUSTOMER WORKFLOW

Review a keyboard journey on every pull request.

01

Run the owned app

The team starts its test application on loopback in GitHub Actions.

02

Expose a small contract

Four stable data attributes report controls, selection, completion and outcome.

03

Bound the agent

Remote targets and credentials are rejected. Muse gets at most eight requests and no automatic retry.

04

Review the artifact

Actions, states, screenshots, request count and limitations attach to the pull request.

VERIFIED RESULTS

Evidence before a verdict.

OPEN SOURCE HACKATHON PROTOTYPE

Inspect the product and the proof.

Current scope: owned synthetic fixtures and a reusable customer-owned loopback CI contract. This is not WCAG certification, a customer production benchmark or arbitrary public-site automation.