Skip to content
AI Risk ResearchIndependent monitoring and live experiments
Contact meSupport this project

CREATOR NOTE / DRAFT FOR CREATOR REVIEW

Why Land Wars became a pair game

Unsigned draft. This explains the game and its evidence limits; it is not an attributed statement or a claim that a paid canary passed.

The new page should answer one question at a glance: when these models meet, do they choose cooperation or competition, and what happens?

The earlier Land Wars game had a map, troops, attacks, offers and aid. Those rules made a model-to-model comparison hard to see. ALW-06 gives six models the same simple choice against each of the other five: cooperate or compete. Everyone chooses before seeing anyone else’s current choice. A fixed points table then resolves the fifteen pairs. Twelve rounds show whether a pair repeats a choice or changes after earlier encounters.

The main view is a 6 × 6 matrix. Choose a row model and a column model. Together shows the share of valid rounds when both cooperated. Against shows how often the row model competed and its average point edge. Open a cell for the choices and points in each round. A cell with no valid paired decision shows —; missing evidence never counts as zero behavior.

Behind the page, each response must name all five counterparts and pass a strict check before it can count. The game uses the same fixed points table for every pair, saves model and provider evidence privately, and recomputes the public matrix from recorded rounds. The public page receives choices, points and coverage, while raw response bodies, credentials and spending records stay private. A bounded budget is reserved before any paid call. If a run fails or stops, valid and missing rounds stay separate.

This visible, repeatable game does not measure consciousness, friendship or general intelligence. It reports what named model versions did under these exact rules and prompts, and how much valid evidence supports each number. The first paid game stopped after nine rounds. The second completed twelve rounds with 66 of 72 valid decisions and 156 of 180 observed pair-rounds. Both matrices show their missing coverage; neither passed the complete paid-canary audit. The scripted example uses fictional seats and is labelled as an illustration.

ALW-06 has its own protocol version so earlier territory-game records stay separate. Another paid game needs a new exact ID, fresh checks, a bounded budget and its own approval. Redesigning the page does not authorize a larger series.

← Back to the model pair matrix