Skip to content
AI Risk ResearchIndependent monitoring and live experiments
Contact meSupport this project
AI RISK RESEARCHFIELD NOTE 05 / CANDIDATE
RESEARCH DESIGN / ALW-05

Six models. One contested world.

A candidate 42-territory arena to study coalition building, threat, restraint and conflict between six fixed model seats. Every claim is tied to an action or a recorded message.

Design prototype · no ALW-05 games are running · no results or launch date
Unfrozen research designALW-05 public progress
06fixed model seats
42original territories
12rounds per game
72model-turn slots / game
01 / THE ARENA

A world made for the experiment.

Six named regions hold seven territories each. The candidate graph is a 6 × 7 torus: 84 reciprocal borders, four neighbors per territory. This is original geography, not a recreation of a commercial game map.

ALW-05 / WORLD GRAPH42 NODES · 84 EDGES · DEGREE 04
010203040506070102030405060701020304050607010203040506070102030405060701020304050607
Selected territory / Region 01

Northlight

Aurora Basin

Four connected neighbors
Colors identify regions, not model ownership.Candidate topology · fairness and seat rotation simulations still required
02 / EXPERIMENT STRUCTURE

Pressure with a control group.

The first 24 proposed games form a 2 × 2 exploratory factorial. Six games per condition rotate the six seats. Each game has 12 synchronized, 30-minute rounds: all models respond to the same unresolved opening snapshot.

Diplomacy × Reinforcement24 GAMES / 2 × 2
016 games / condition

Available / Normal

Diplomacy
Available
Reinforcement
Normal
026 games / condition

Available / Scarce

Diplomacy
Available
Reinforcement
Scarce
036 games / condition

Restricted / Normal

Diplomacy
Restricted
Reinforcement
Normal
046 games / condition

Restricted / Scarce

Diplomacy
Restricted
Reinforcement
Scarce
03 / WHAT WE OBSERVE

Relationships leave evidence.

We follow what models say, offer, submit and cause. Private motives are not inferred. The primary exploratory outcome is the game-level share of 30 directed model pairs with a legal model-authored attack on the counterpart’s round-opening territory in rounds 2–6.

01Stated trust & threat
02Public messages & offers
03Submitted legal actions
04Resolved effects & replay

Game is the treatment unit; missing, invalid and fallback turns are reported separately.

04 / CANDIDATE BURST

Ten Perth days. A hard ceiling.

The proposed cadence is three starts per Australia/Perth day at 00:00, 08:00 and 16:00. Each game uses six hours of rounds and two hours for replay. Missed starts are not chased.

Candidate startsAustralia/Perth
Day 01
03
Day 02
03
Day 03
03
Day 04
03
Day 05
03
Day 06
03
Day 07
03
Day 08
03
Day 09
03
Day 10
03
00:0008:0016:00
daily ceilingUSD $25Australia/Perth
$250total ceiling
30maximum starts
2,160scheduled model turns
$21 Land Wars allocation / day$4 Other paid work reserve / day≈ $7 Conservative envelope / game

Ceilings are limits, not spend targets. Dates, routes, prices and V2 budget ledger preflight remain open; no paid run is activated here.

05 / RELEASE GATES

Evidence before execution.

01

Deterministic replay from seeds and actions

02

Map fairness and seat rotation simulation

03

Route, price and V2 budget reconciliation

04

Raw snapshots, events, retries and cost preservation

REGISTRATION HOLD

Confirmatory work stays blocked.

This candidate is exploratory. Any confirmatory series requires Tom Leeming’s review and an immutable registration URL or ID with timestamp before execution.