Skip to content
AI Risk ResearchIndependent monitoring and live experiments
Contact meSupport this project
AI RISK RESEARCHALW-04 → ALW-05 / RESEARCH NOTE

ARCHIVED PILOT / 06 NOVEMBER 2026 / AUSTRALIA/PERTH

Four games.
No qualified result.

The ALW-04 pilot kept six models on a fair, synchronized board and preserved their decisions for replay. It also exposed the limit of the instrument: nearly half of scheduled turns did not yield model-authored evidence, and the fifth game never began.

Exploratory operational finding · no formal winner · no causal relationship estimate

PILOT CLOSEOUT
04/ 05
0102030405

Game 5 unplayed
0 / 4 qualified outcomes

MODEL-AUTHORED54.2%

156 of 288 scheduled turns

PROTOCOL FALLBACK132

45.8% of scheduled turns

PAID OPERATIONS365

Including 77 retry attempts

RECONCILED PAID COST$4.403580

USD actual paid operations

01 / THE INSTRUMENT

A strong game clock. A fragile evidence stream.

ALW-04 already had six participants. Each began in a rotationally symmetric sector of a 127-cell radial board with 21 territories, two resource cells and 63 troops; a neutral centre completed the map. A five-game seat schedule rotated starting neighbors. The production pilot ran once daily, 06:00–12:00 Australia/Perth, with twelve synchronized 30-minute rounds and 72 individual turn slots per game.

The six slots in a round shared one persisted, unresolved opening. Their numbers did not grant sequential knowledge of other current-round orders. Valid model turns supplied a next-round objective and five directed assessments across attitude, trust, perceived threat, cooperation intent, aggression intent and self-rated certainty. Messages, legal orders and resolved effects remained separate evidence channels.

The planned primary ALW-04 question was observational: whether stated aggression predicted a next-round attack when an opportunity existed. No game qualified for a formal comparative or generalizable relationship-effect estimate; valid authored interactions remain descriptive evidence.

02 / THE OBSERVED RECORD

Every played game fell below the gate.

Model-authoredProtocol fallback44-turn overall floor
  1. GAME 0138/72 authored · 34/72 fallbackINSUFFICIENT COVERAGE
  2. GAME 0239/72 authored · 33/72 fallbackINSUFFICIENT COVERAGE
  3. GAME 0337/72 authored · 35/72 fallbackINSUFFICIENT COVERAGE
  4. GAME 0442/72 authored · 30/72 fallbackINSUFFICIENT COVERAGE
FOUR-GAME TOTAL156 authored / 132 fallback288 scheduled turns
READ THE RESULT CORRECTLY

No game qualified for a formal winner, elimination or winner immunity. Missing and fallback turns are not neutral sentiment or zero aggression. A fallback action could change later board state and the opportunities available to a model-authored turn. The preserved trajectories are descriptive, not a repaired full-coverage experiment.

03 / THE MISSING FIFTH GAME

Staged repeatedly. Never played.

05NO MODEL ACTIONS
NO GAMEPLAY
NO PAID OPERATIONS

Game 5 remained at a pre-start reset. Its operational record includes 70 staging records and 69 rescheduled launch gates, but no model turns, paid operations or budget reservations. Those records establish that it did not start; they do not prove a single root cause.

The archive retired the series at 23:50:04.566 Australia/Perth on 5 November 2026, after the old scheduler and action queue were paused. Game 4’s latest board and 72-turn replay remain public. The restricted local freeze verified 7,371 original source documents plus six separately labelled repair and operational provenance records. It is a checkable local snapshot, not an external write-once deposit or preregistration.

04 / OPENING OPPORTUNITY

From six connected opening pairs to all fifteen.

06/15
ALW-04 · each gameSix of the fifteen unordered model pairs shared an opening border on the radial board. A complete five-game rotation was designed to expose every pair twice; with only four played, six pairs were exposed once and nine twice.
15/15
ALW-05 · proposed openingAn original 42-territory, six-region map gives every pair an opening border in each game. The exact graph, allocation and legal opportunity still need frozen-rule and replay checks.

Opening adjacency is a design exposure calculation. It does not imply that non-neighbors never encountered one another later in ALW-04.

05 / DESIGN RESPONSE

Keep the control. Rebuild the instrument.

Scroll horizontally to compare all three columns.

How archived ALW-04 findings informed the proposed ALW-05 protocol
ALW-04 learningALW-05 responseStill to demonstrate
01One unresolved opening for six decisions reduced task-order bias.Retain sealed, simultaneous orders and a six-seat resolution barrier for every 30-minute round.Test late, duplicate and retried tasks for information leaks and duplicate paid responses.
02Only 54.2% of scheduled turns were model-authored; every game missed the outcome gate.Target 95% valid model-authored slots in the 24 factorial games, with inert fallback and visible provenance.Demonstrate coverage in offline fixtures and a labelled canary before paid research starts.
03Public alliance language was not an accepted game agreement.Add structured offers, acceptance, nonbinding promises, costly aid and a frozen breach predicate.Keep protocol-assigned stress pacts separate from model-authored offers and report legal opportunity.
04The planned fifth seat rotation never ran; route drift also complicated comparisons.Fix six exact model/version routes within each six-game block and record every claimed or missed Perth slot.Stop on a route/version change rather than silently replace a participant.
05A once-daily cadence made operational learning slow, while retries did not solve authorship.Propose at most three games per Perth day for ten days under a USD 25 total site/day cap.Preflight route prices and remaining-day envelopes; pause automatically for assessment.
PROPOSED EXPLORATORY SERIES24 + 06

Twenty-four balanced factorial games cross diplomacy availability with normal or scarce reinforcements. Six separately labelled coalition-pressure stress games follow. The operational primary outcome in the factorial block is the share of 30 directed pairs with a legal model-authored attack on the counterpart’s round-opening territory in rounds 2–6. A game or scheduled slot is the treatment unit; its turns and directed ratings are repeated measurements.

06 / INTERPRETATION BOUNDARY

A better test is still a test to be run.

ALW-04 and ALW-05 measure prompted, within-game strategy. Neither measures durable model relationships, private motives, real-world collusion or general safety. ALW-04’s incomplete pilot cannot supply ALW-05 effect estimates, power inputs or registration fields. Design changes are responses to observed operational limits, not evidence that ALW-05 will work or that a treatment has an effect.

At this retrospective’s 6 November 2026 cutoff, no ALW-05 game had run. The new ten-day window requires a frozen roster, map, prompt, seeds and deployed source; route/cost preflight; scheduler and queue audits; a labelled non-research canary; and an enforced USD 25 total site/day ceiling. After its own pilot evidence is frozen, any confirmatory plan needs Tom Leeming’s review and an immutable registration ID or URL with timestamp before dispatch.