Initial records come from the local preview; new trials run on the replacement server. Recovery record ↗
ZEBRAFISH NEURAL / BROWSER LAB
Change the scene.
Measure the response.
A browser, a visual target, and the fish’s controller. We move the target, remove the cue, or add a distraction. Every trial leaves a record you can replay.
Switchback
The target changes position.
Blackout
The visual cue disappears.
Decoy
A competing cue appears.
LIVE BROWSER
Waiting for a trial
The actual browser view will appear here.
No sample footage or substitute simulation.Encoded visual cue
32 × 16Screenshot luminance mapped to a bearing cue.
Motor response
Waiting for samples.
FROM OBSERVATION TO EVIDENCE
Replay the response.
Pick a saved trial. Recompute its recorded inputs, then test what changes when one signal is removed. Interrupted attempts are labeled separately.
One trial, inspectable from start to finish.
Recorded screenshot inputs, neural states, and cursor positions stay aligned in model time.
RECORDED TRIAL / —
—
The recorded trial has not been recomputed in this browser yet.
Single trial; fixed weights. Baseline includes the initial approach.
Recorded encoded cue
—Motor response comparison
Original trace: Motor L − Motor R.
Remove one signal.
Replay the same recorded inputs with an intervention.
Select an intervention to compare motor output.
Recorded inputs, replayed with one intervention. This does not predict a new browser trajectory.
WHAT IS BEING TESTED
A visual response, with a paper trail.
A real browser screenshot.
The adapter finds the brightness-weighted center of the screenshot’s bright regions and maps its bearing into visual input. The controller does not receive target coordinates or text instructions.
Same controller, different conditions.
A defined protocol changes the scene. We record the controller’s activity and movement before, during, and after the intervention. Target contacts are measurements, not scripted outcomes.
Recompute what was recorded.
Replay checks the saved inputs against the published controller. It verifies computational consistency, not where the original computation ran. This experiment does not update learned weights.