mayklink’s AI Player Lab connects local computer vision, Jev decisions and a separate model that reviews failed attempts. Its Super Mario World experiment exposes the machinery needed to turn a typed answer into a timely action.

The public repository includes the implementation and synthetic tests. Game images and the historical run artifacts are withheld, so its reported gameplay remains only partly inspectable.

Give Jev a bounded job

  • OpenCV, a computer-vision library, estimates the player, moving objects and nearby hazards from pixels.
  • The Jev adapter sends structured observations to TypeSafe’s /v1/systemone endpoint and validates the returned action against a closed menu.
  • In the inspected browser runtime, replies expire when their observation is too old or the scene generation changes.
  • A local controller turns accepted intent into short key presses, checks fresh observations and can override movement near hazards.
  • A separately configured vision-language model reviews episode images and suggests another experiment.

Jev receives structured state; the reviewing model receives images. The component named “Luna” uses the configured provider and model.

A remembered suggestion can repeat the failure

The author’s historical log describes a retrieved experience suggesting a jump. Jev followed it, and the attempt ended in another death. The implementation now distinguishes historical context, an authorized experiment and an alternative associated with a later failed attempt.

  • Memories carry source frames and contextual matches.
  • Similar scenery supplies context; matching moving objects supports a stronger comparison.
  • Old search results require fresh checks before adoption.
  • Later failure can change how advice appears in the next request, while its causal role stays uncertain.

The log also records a review proposing an earlier jump while retaining the same structured action, jump. Timing advice needs an executable condition and duration before it can change control reliably.

Keep retrieval off the control path

The author reports that moving memory lookup into a background worker reduced delay before sending Jev requests in later rounds. Those rounds still failed the opening encounter. Historical artifacts are absent, and these observations were not reproduced here.

The inspectable mechanism is narrower: the worker keeps one lookup in flight, avoids a backlog of old frames and revalidates stale candidates against current geometry. It can continue without a retrieved memory.

What the local tests establish

Local execution passed 30 focused tests from the policy/I/O and memory-worker modules. They exercise simulated HTTP responses, invalid-action rejection, expired decisions, episode changes, bounded work queues and stale-memory revalidation.

Typed output validates an action’s shape. Correct movement also depends on perception, current geometry, timing and the available actions. The browser runtime explicitly treats Jev’s confidence as diagnostic rather than a calibrated probability of success.

Verification: inspected source commit f0d12ff6f8d9fc984b62866b871676d0b646d5ca and ran the focused offline tests. No live Jev calls, browser gameplay or end-to-end performance reproduction was performed.