01
Run the trial
Complete a short obstacle course while the page records compact, local performance signals.
AIMAXIN Lab / Second Player
Challenge build in private validation
The player gets the final word.
Second Player is a WebMCP play lab where an agent can inspect a short trial and propose one narrow adjustment. Nothing changes until the player approves it—and no comparison can call it better until the player reports how it felt.
46 automated checksContract, persistence, lifecycle, and core behavior.
Local signals onlyThe current prototype uploads no play telemetry.
Explicit approvalThe apply capability appears only after the player says yes.
Keep or undoEvery applied tune stays narrow and reversible.
One bounded experiment
Play. Inspect. Approve. Play again.
The agent helps carry evidence across the loop. The human decides what may change and remains the only authority on whether the experience improved.
02
Inspect the evidence
The agent reads bounded measurements and their limits instead of inventing a diagnosis.
03
Approve one change
A narrow proposal becomes actionable only after the player gives exact, visible permission.
04
Report the result
After the second run, the player's own check-in unlocks comparison, keep, and undo.
The interface is the permission model
The tool surface changes when consent changes.
WebMCP is not a hidden shortcut around the game. Each capability appears only when the visible human workflow has earned it.
Observe and propose
The agent can inspect the trial and supported options, but it cannot quietly apply a tune.
inspectread evidencepropose
Apply one bounded tune
The apply capability becomes visible for the approved proposal—and no broader change.
apply approved tuneundo
Compare without overclaiming
The agent can summarize both runs, but the player's report constrains the verdict.
comparekeepundo
Evidence, not magic
A working proof with visible limits.
The prototype demonstrates a control pattern for human-agent collaboration. It does not diagnose a player, certify accessibility, or pretend that telemetry can speak for someone's lived experience.
- Sample runs stay visibly labeled and cannot masquerade as player evidence.
- A worse player report blocks any recommendation to keep the tune.
- Undo preserves the comparison record instead of erasing inconvenient evidence.
- The ordinary playable interface remains complete without an agent.
Built for the 2026 WebMCP Challenge
An agent can help carry the evidence. It cannot take the player's place.
The public challenge build, source, and demonstration will be linked here after the final validation pass. Until then, the page documents exactly what the current build proves.