ArenaSI Arena — logged multi-agent survival pilot
A first-party pilot report describes researchers examining six LLM settlers in a deterministic survival simulation. The authors report replaying 5,600 steps with matching events and 280 state-digest checks. These are author-reported checks of one selected run, not independent replication or evidence of general altruism.
L2Research workflow (workflow level, not a quality rating) Source: Project-documented
Research applicability
A first-party pilot report describes researchers examining six LLM settlers in a deterministic survival simulation. The authors report replaying 5,600 steps with matching events and 280 state-digest checks. These are author-reported checks of one selected run, not independent replication or evidence of general altruism.
Getting started
Read the retained primary snapshot and its limitations first. Any suggested installation/provider/database use is untested and requires separate authorization.
Maintenance
Snapshot checked 2026-10-20; ongoing maintenance and version parity were not established.
License
License statements were not separately audited; source access is not reuse clearance.
Workflow & human oversight
L2 for this documented use: Researchers configure the simulation and model setup, log actions and agent statements, replay the event journal and inspect food-sharing and erroneous self-reports as candidate multi-agent evaluation phenomena. Human control: The authors report a specified six-agent setup, logging, journal replay and inspection of outputs; they explicitly withhold causal and generalization claims. Boundary: Primary forum report only; no code, journal, video, claimed model version or state digest was independently verified. The report is one selected run, does not establish motives or generalization, and replay consistency is distinct from independent reproduction or causal validity.
Workflow level is not an evidence score; higher is not better evidence.
Approach
Researchers configure the simulation and model setup, log actions and agent statements, replay the event journal and inspect food-sharing and erroneous self-reports as candidate multi-agent evaluation phenomena.
Outcomes (as reported by the originating team)
Authors report replaying 5,600 steps with identical events and 280 checked snapshots matching by digest; one selected run, no independent verification.
Limitations
Primary forum report only; no code, journal, video, claimed model version or state digest was independently verified. The report is one selected run, does not establish motives or generalization, and replay consistency is distinct from independent reproduction or causal validity.
Reproducibility
Primary-source documentation was read, with source snapshots and hashes retained. No candidate code, skill, model, experimental system or claimed outcome was executed or reproduced.
Adoption requirements
Untested editorial suggestion: inspect the exact sources and select a bounded authorized task; check data rights, provider terms and outputs before relying on them. No installation or trial was run.
Comments
No comments yet.
Add a comment
Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.