RandallSPQR/social-phase-transitions-data

Social phase transitions: LLM response cache, pod outputs and simulation units

This dataset holds the raw material behind the study "Do LLM agents behave like spins? Response functions, live networks and amplification near tipping points". Together with the code repository it reproduces every analysis and figure exactly, with no API keys and no GPU.

• Code, results and replication package: https://github.com/RandallSPQR/social-phase-transitions-public (make replicate-fast / make replicate; see replication/README.md)
• Archived release (code + this card): DOI 10.5281/zenodo.23821224
• Licence: CC BY 4.0 (see "Terms of the model outputs" below)

Contents

path	what	size
cache/*.jsonl.gz	Response cache: one gzip JSON-lines file per model, 31 files, 709,422 records	58.5 MB
cache_manifest.json	SHA-256, size and record count of every cache file (also committed in the repository as results/llm/cache_manifest.json)	—
sim_units/*.tar	Per-unit outputs of the long simulations (amplification Stage B; confirmatory locate, Steps 4–5, Gemma checks and mechanism test); manifest in the repository at replication/sim_units_manifest.json	140 MB
runs/<run id>/	Durable record of every GPU/CPU pod run: heartbeat status.json, full logs, per-unit outputs as uploaded during the run. runs/e3zan9ygnmqe45/ includes the Gemma-4 activations used by the exploratory probes (about 825 MB; listed in results/llm/activations_manifest.json in the repository). runs/stage3-rep1/ is the live replicate (Stage 3 option B)	about 1.9 GB
code/pod_code_<commit>.tgz	The exact code bundle each pod ran, named by repository commit	about 0.5 GB
pod_results/, cleanup/	Small pod bookkeeping files	—

Cache record schema

Each line of cache/<provider>__<model>.jsonl.gz is one API call:

field	meaning
key	SHA-256 of the canonical JSON {"body": <request body>, "rep": <replicate index>}; the cache key
rep	replicate index (integer, or a string tag for live-network calls such as stage3:P1:g:r:sweep:step:attempt)
body	the OpenRouter chat-completions request body: model, pinned provider, messages, temperature, logprob settings (no headers, no credentials)
t	Unix time of the call
resp	compact response: serving provider, model, reply text, reasoning text if any, finish reason, first answer token with its top-5 log-probabilities, number of logprob tokens, usage and cost. The OpenRouter generation id is deliberately not stored

How the records are used: llm/client.py in the repository looks every request up by key. With SPT_OFFLINE=1 a request that is missing from the cache is an error, never a live call. One caveat is documented in replication/stage1_rematerialise.py. A battery that sent the same request twice (same body and replicate index) has two answers in the committed CSV but only one in the cache. The replication check accounts for this explicitly.

How to verify and use

git clone https://github.com/RandallSPQR/social-phase-transitions-public && cd social-phase-transitions-public
make env && make replicate-fast PY=.venv/bin/python     # or: make docker-replicate-fast

The run downloads the files it needs from this dataset and checks each against its SHA-256 before use. It then reruns the analyses and prints a pass/fail table per claim of the paper.

Provenance

• Collected: 2026-10-08 to 2026-10-14 (UTC, from the record timestamps) through OpenRouter, with each endpoint pinned to one provider and fallbacks disabled. Self-hosted Gemma-4 runs were on rented A100 GPUs (RunPod).
• Pre-registration: every confirmatory phase was planned before its data existed. The plans, amendments and the chronological execution log are in the repository (llm/PREREG.md, llm/*_PLAN.md, llm/EXECUTION_LOG.md).
• Models: the experimental subjects are listed in the paper (Appendix D) and in llm/endpoints.py. The cache also contains small probe sets for candidates that were screened but not selected.

Terms of the model outputs

The records contain short outputs (mostly a single answer letter and its log-probabilities) of the models below. This dataset is released under CC BY 4.0. Use of the outputs also remains subject to each provider's terms:

model(s)	terms
Gemma-4 26B-A4B, Gemma-4 31B	Apache 2.0 (model card)
Qwen3.5 9B, Qwen3.5 122B-A10B	Apache 2.0
Mistral Nemo 12B, Mistral Small 3.2 24B	Apache 2.0
Mistral Large 4	Mistral AI terms
Llama-3.3 70B, Llama-3.1 8B, Llama-3.2 3B, Llama-4 Maverick	Llama Community Licences and Acceptable Use Policy
GPT-4o-mini, GPT-4o, GPT-6-luna	OpenAI terms of use and usage policies
Gemini 3.1 Flash-Lite, Gemini 2.5 Flash-Lite, Gemma-3 27B (via provider)	Google terms for the respective service or model
Claude Haiku 5.5	Anthropic terms
DeepSeek chat v3 / v3.1 / v4.x, Qwen3 235B / 30B, Qwen3.5 27B (probe calls only)	the respective model licences and provider terms

Citation

@misc{social_phase_transitions_2026,
  author = {Bennington, Randall S.},
  title  = {Do {LLM} agents behave like spins? Response functions, live networks and amplification near tipping points: code, data and replication package},
  year   = {2026},
  doi    = {10.5281/zenodo.23821224},
  url    = {https://github.com/RandallSPQR/social-phase-transitions-public}
}

AI use

Claude (Anthropic) assisted with the experimental code, the analyses and the documentation; see the paper's AI-use statement. The models above are the subjects of the study, not authoring tools.