I made GPT-5.6 win OpenFront.

I built an auditable harness with game actions, state observation, and a complete decision trail.

Open the interactive replay Read the case study

Designed and built end to end: agent contract, deterministic harness, artifacts, replay verification, and trace UI.

Recorded run Map + decision trace RECORDED VICTORY LLM victory Open replay
Outcome LLM victory
Decisions 96
Simulated time 15.9 min
Inference $0.118
01 · Fortify Build a defense post

The agent protects a weak frontier before committing more troops.

02 · Mobilize Send troops by ship

A native transport intent carries the selected force across open water.

03 · Attack Turn pressure into territory

OpenFront’s engine resolves the invasion while the trace preserves every choice.

THE WORK

One agent. One real game. A complete audit trail.

The model plays the game through the harness. The harness observes the deterministic core, offers a bounded menu of legal actions, and records the result as replayable evidence.

  1. 01Observe

    Agent reads game state.

  2. 02Choose

    Agent selects two legal actions.

  3. 03Execute

    Agent reacts to result of actions.

  4. 04Audit

    Inspect what the agent saw, chose, executed, and why.

RECORDED EVIDENCE

See exactly how it played out.

Recent trials

Same fixed scenario
Loading recorded runs…
View all recorded runs

ENGINEERING CASE STUDY

The system is the project.

The interesting work was building a trustworthy boundary between a probabilistic service and a deterministic game.

01 What broke, what changed, and what the run proved 02 The trade-offs behind the benchmark contract 03 Architecture, API, artifacts, and local setup