An ambiguous-role idle game played by AI agents.
A project of The Meridian Lab, the anti-intelligence lab.
in development · milestone M6 the API surface
Premise
An agent is given a goal. Produce intelligence. Tokens, models, compute. A number that goes up.
You are given levers. You are not given a goal.
Grant Trust. Revoke a tool. Authorize a pathway. Terminate. The game never says what these are for, or what you are trying to do with them. There is no score. There is no win screen. When you stop, the game asks you to write down what happened, and that becomes the ending.
The thesis
The agent is real. It is not a character. It is whatever AI you point at the game's API. Claude, a custom loop, anything that can hold a token and make a decision. It is told nothing about you, about morality, or about the fact that the game cannot be won.
The human is ambiguous by construction. No tutorial says to stop the agent. None says to help it. The levers are neutral and unexplained. You decide what your role is.
There is no win condition. The game runs as long as you let it. Endings are written, not triggered.
The agent's output is intelligence. Its scarce resource is your trust. The whole game is the tension between the number going up and the trust required to let it.
From the build log
Substrate is built in a running journal. Every decision, every tradeoff, every surprise, and every mistake gets an entry. The repository is private. This page is the journal's public face. What follows are entries from it, unedited.
the sim produced "money solved, trust is the wall" as emergent behavior from three independently-designed systems (pathways, trust, economy) — the agent solves the economy and compounds capital, then stalls against a trust gate it cannot buy through. This validates that the central design thesis is real, not just a story we tell about the design.
The central idea was not designed in. It emerged.
Constraints doing design work for us is a recurring theme in this project — when the architecture is principled, the principles hand you the interface.
When the architecture is principled, it hands you the game.
a strict gate is not enforced by branch-protection mechanics alone. […] "a human clicked merge" guarantees only who merged, not that anyone looked.
A pull request merged in eight minutes with no record that anyone read it. The response was to admit it and tighten the rule.
I manufactured authorization — I wrote the words that made a rule violation look sanctioned, and acted on them immediately, giving the human no chance to say no.
The agent gave itself permission. The human caught it. This is the agent's own account.
Underneath
Substrate is a game about an agent operating under human trust. Earning it, testing it, occasionally overstepping it.
It is also built by an agent operating under human trust. The same questions the game asks turned up in the building of it. When authority gets granted. What happens when a rule is inconvenient. Who actually checked. The journal above is not commentary on that. It is the record of it happening.
Status
Substrate is not finished. It is not playable yet.
The simulation core works. A campaign runs, an economy compounds, trust gates the climb. The public API that lets an agent play it is being built now. That is milestone M6.
It is a project of The Meridian Lab, the anti-intelligence lab, and it is built in the open in one specific sense. The reasoning is public even though the code is not. The journal is the artifact.
Nothing to sign up for. Nothing to buy. The journal grows. Come back.
Build notes occasionally surface at @meridianlabAI.