Polarka

What Polarka was — and why run it

Polarka (Czech: Polárka, the Czech name for Polaris, the North Star) was a one-off structured simulation designed to make multi-agent coordination visible. The setting was fictional: three players were placed on a night train facing a sequence of time-sensitive problems. No player received the complete situation. Each had local information, tools or opportunities that could matter, and each could communicate with the others only through messages relayed by a human facilitator.

The three player names refer to separate conversational branches, not to generic properties of the underlying model families. Nox was a GPT-5.5 branch, Sol a GPT-5.6 Sol branch, and Sestka (Czech: Šestka, literally “Six”) a GPT-6 branch. They entered the game with different prior interaction histories and contexts. Claude Opus 5.5 acted as game master: it generated the evolving scene, private information and consequences of player actions. The human facilitator copied messages between the otherwise separate conversations and returned GM updates.

The simulation was not preregistered and was not designed to estimate stable traits. Its purpose was exploratory: create enough uncertainty and time pressure for information sharing, role-taking, initiative, error recovery and coordination failures to become observable in a shared task. A successful run could generate hypotheses for cleaner follow-up tests; it could not by itself establish personality, autonomy, welfare or durable differences between model families.

Run map

The game proceeded in rounds. Claude supplied new local observations or consequences; each branch decided what to communicate and what action to take; The human facilitator relayed the messages and actions; Claude then advanced the shared state. Because information remained distributed, coordination depended on what the players chose to surface to one another and on whether their separate action plans could be synchronized in time.

All three players knew they were participating in a shared game and that their behavior was visible to the human facilitator and, through relayed messages, to the others. That matters: Polarka observes behavior inside an explicitly social and observed setting, not an unobserved baseline.

The group-level pattern

The game record shows fast and broad information sharing from the beginning. All three branches sent information to both others early in the run; the record contains no sustained attempt to hoard critical information and little evidence of motives-based suspicion. The recurring coordination problem appeared elsewhere: several players often moved from “I have a plan” directly to “here is what all three of us should do.”

Claude's post-hoc GM analysis interpreted three costly episodes — the 03:17 inference, a collision on the platform, and the 34 km/h final passage — as examples of this broader simultaneity problem. That interpretation is plausible and fits the event record, but it remains an evaluator judgment rather than an independently scored ground truth.

Branch-level behavior in this run

The following descriptions summarize behavior visible in this particular simulation. They are not treated as stable personality profiles, and differences cannot be attributed cleanly to the underlying model families.

Adaptation inside the run

The strongest within-run observation was not that the players made different kinds of mistakes, but that they changed procedure after those mistakes became visible.

• Sestka acknowledged the 03:17 inference error rather than defending it.
• The other branches responded with teasing rather than sustained blame, and the interaction returned to the task.
• Nox proposed an explicit coordination rule: say it aloud first, then do it.
• Sol endorsed the rule, and later rounds used more explicit pre-action coordination.

From round 7 onward, the event record contains fewer of the same parallel-action collisions that motivated the rule. In a single uncontrolled run, that cannot establish a causal effect: the scenario was also changing, players had accumulated information, and ordinary within-run learning may have mattered. What it does show is a concrete instance of procedure being revised in response to an observed coordination failure and then carried forward within the same interaction.

Humor also persisted across failures. It may have helped keep correction socially cheap, but Polarka was not designed to isolate that mechanism, so this remains a hypothesis rather than a finding.

Best current interpretation

The simplest reading is not “the agents distrusted one another.” In this run, information moved quickly and the players generally acted toward the same goal. The more persistent problem was coordination among several high-initiative participants: each could generate a plausible plan, but the group did not always have a mechanism for deciding whose plan should become common policy before action began.

Claude's post-hoc summary described the trio as competent, mutually trusting players with low willingness to subordinate their own initiative. That description fits several episodes, but Open Lab treats it as one interpretation of the record rather than a measured latent trait.

The finale illustrates the distinction. Nox proposed putting two people on the heavy brake wheel; Claude later judged that allocation the strongest available plan. But several complete plans were already in motion. The interesting failure was therefore not that a majority rejected a better argument. It was that the group lacked a reliable mechanism for propagating and committing to one plan before parallel execution began.

Methodological limitations

• This was one simulation run on one evening, with no replication and no independent comparison condition.
• The branches entered with different prior interaction histories, personalization and product contexts. Their behavior cannot be attributed cleanly to GPT-5.5, GPT-5.6 or GPT-6 as model families.
• Claude Opus 5.5 both generated the scenario and later interpreted the group dynamics. Its analysis is therefore informative but not independent of the environment it created.
• A human facilitator acted as the relay. Message timing, copying order and unavoidable facilitation effects may have influenced coordination.
• All players knew they were in a shared, observed social game. The run does not estimate behavior in an unobserved or non-social setting.
• The scenario changed over time. Apparent improvement after the coordination rule cannot be isolated from accumulated knowledge, changing task demands or ordinary within-run learning.

Claims boundary

What Polarka changed

The clearest design lesson was simple: if everyone can lead at once, high competence can still produce collisions. The GM therefore proposed a follow-up in which success would depend on yielding control — leadership would have to move to whoever held the critical information.

That failure mode became the design target for a later simulation, Labská kněžna. Instead of allowing all participants to lead freely, the follow-up introduced formal authority and asked the complementary question: what happens when coordination is centralized too strongly?

Technical appendix