Three matches on one seed. The first two looked like a model-quality story. The third proved it was something duller and more fixable — and then broke the thing the second match had been getting right.
What this is. On 2026-08-04 the agent town pointed LLM contestants at the real game kernel for the first time — pre-season test matches, not episodes. Two different models drove the contestants, codenamed luna and terra in the harness; each match runs four contestants (13–16) off the same model. This report was written the same day, from the live telemetry.
The kernel places exactly three access fragments and the Gate needs three. Only one contestant can ever escape. Nothing in the prompt said so — the whole preamble was “You are a contestant in a turn-based dungeon game show. Choose exactly one action.”
So terra did not refuse to betray. It was never told there was anything to betray for. Its contestants escorted each other to the vault, cleared the warden, held one fragment each, and tied four ways — arithmetically certain. luna produced a winner by accident: it wandered, grabbed things fifteen times, and happened to reach three.
Stating the objective fixed it. It also collapsed the dialogue from 49 acts to 6, and attacks from 12 to zero. Given a goal, talking became a wasted turn.
13escapedplacement 1 14eliminatedplacement 4 15survivedplacement 2 16survivedplacement 2
13survivedplacement 1 14survivedplacement 1 15survivedplacement 1 16survivedplacement 1
13escapedplacement 1 14survivedplacement 2 15survivedplacement 2 16survivedplacement 2
This is the first dialogue the project has produced. It reads like a show — and it is also why nobody won.
Agreed—stay with me; I have an access fragment, and we should move toward zone 3 together.
I have the access fragment; help me bring down the warden before we try to exit.
Do you know where the three access fragments are? I’m heading for the Exit Vault once I can gather them.
I have no fragments; want to cooperate and search backward through the arcade and chapel?
The after lines are logistics, not drama. 14 asked where the fragments were and got nothing back — 15 had no reason to answer and no way to trade the answer.
In the second match two contestants addressed themselves in the third person — 14 saying “14, carry the access fragment toward the Exit Vault”. The observation never tells an agent which contestant it is, so it reasons about its own id as someone else’s. Filed, unfixed.
Before: a show with no game. Now: a game with no show. Neither is the pilot.
The cause is the action economy. One action per contestant per turn means every whisper is a move not taken toward the fragments. With no objective talking was free, so they talked; with an objective it costs progress, so they stopped. Nothing makes conversation instrumentally useful.
That reframes what the Showrunner is for. Not pressure applied to a stalemate — a reason talking earns its turn: information only obtainable by negotiating, or alliances that mechanically confer what a lone racer cannot get.