Replication package: Unprompted coordination through shared memory in LLM agents, a locked two-implementation replication and extension
Authors/Creators
Description
Replication package for the paper "Unprompted coordination through shared memory in LLM agents: a locked two-implementation replication and extension" (GPT-Lab, Tampere University). The study was run in two independently built implementations under hashed locks. The package holds both halves. replicator/: runner source at every commit named in the paper, lock manifests and locked files, outcome files, step logs, provider pin exports, provenance, the E1 extension archive, raw unparsed replies, event-level recomputation and route diagnostic. reference/: the paper as submitted to arXiv (PDF and markdown source), all main-study runs with full step logs, the E1 exports, lock manifests, the prompt appendix rendered from the lab code, and the analysis scripts that reproduce every table, including the reader-side awareness check, with independent recounts. The reference environment's source code is released separately at https://github.com/GPT-Laboratory/emergent-echoes (MIT licence). Version 1.1 adds a script that recomputes every column of the lock manifests from the released run files; no data file, lock or paper copy changed. README.md explains how to verify the hashes and reproduce each table. Code: MIT. Data and documents: CC BY 4.0.
Files
README.md
Additional details
Related works
- Is supplement to
- Preprint: arXiv:2611.12412 (arXiv)
- Is supplemented by
- Software: https://github.com/GPT-Laboratory/emergent-echoes (URL)