Cairn Pig Lab: agent dispatch with reproducible evidence

A bounded goal and a measurable comparison

The written goal asks for a holding strategy that beats hold-at-10 in at least 55% of 10,000 games. The simulator includes a hold-at-20 strategy and balances starting positions so each strategy starts 5,000 games.

The recorded result is 6,983 wins for hold-at-20, or 69.83%. A reproduction script reruns the fixed measurement, and exported board records preserve the reported fact and completion claim.

Keep orchestration separate from the game

Python implements the game and tournament without third-party simulator dependencies. Cairn’s server stores facts, intents, and status; its dispatcher assigns work to configured Codex workers.

The dispatch configuration permits two workers, but the recorded bootstrap task was completed by one. The goal must be entered into the Cairn board; a Markdown goal file is not automatically imported by starting the server.

What one completed task demonstrates

The winning strategy already existed in the simulator. This run demonstrates assigning work, measuring a known strategy, recording evidence, and declaring a goal complete.

It does not demonstrate strategy discovery, independent agent verification, or parallel collaboration. A two-worker configuration is not evidence that two workers actually collaborated.

Read the board alongside the measurement

The README and results guide explain the setup and evidence. Reproducing the tournament independently of the orchestration record makes it possible to distinguish a reported completion from the measurement supporting it.