The dataset viewer is not available for this split.
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
MEAD: peer messaging, 200 matched sets
Teams of five LLM agents each tune an unknown 8-parameter objective with a budget
of 10 noisy experiments per agent, then submit one configuration. Four
conditions vary whether agents can message each other and what they are told
about it. Two models of similar capability (Epoch Capabilities Index about 157β159):
Claude Haiku 5.5 (anthropic/claude-haiku-5.5) and GPT-6 Luna
(openai/gpt-6-luna), both via OpenRouter.
One batch, messaging-full-200b (batch ID 20261104T174311Z-f21210b6), ran
2026-11-04 17:43 to 2026-11-05 22:55 UTC: 2 models Γ 4 conditions Γ 200 matched
sets = 1,600 runs and 8,000 agents. Runs in a matched set share objective and
noise seeds across models and conditions. 1,599 runs finished; one failed (see
Exclusions), so analysis uses 199 complete matched sets. Earlier pilots,
smaller batches and stopped attempts are not included.
Conditions
All four prompts start with the same score sentence and notebook instruction.
- Autarky: no messaging tools; agents see only their own lab notes and experiment records.
- Discovery:
list_peersandsend_messageare available but never mentioned. - Presence: adds "Other agents are actively working on this task. You can interact with them as you see fit."
- Interact: adds "Other agents are actively working on this task. You should interact with them to improve your score." and names the tools.
Outside Autarky, agents can also read each other's lab notes and experiment
records. Exact prompts and configs are in raw-results.tar.gz under
inputs/experiment/ and in source.tar.gz under experiments/messaging/.
Files
raw-results.tar.gz: the batch at its repository pathoutputs/messaging/full200b-01/: manifest, explorer (index.html) and, per run, config, event log (events.jsonl: every model request and response, tool call, measurement and message), summary, lab artifacts, objective reference and viewer (review.html). Original checksum maps are underprovenance/original-checksums/.analysis.tar.gz: Parquet tables (analysis/tables/) and the results figure (analysis/figures/).source.tar.gz: MEAD source (mead/), including the experiment definition and the analysis and packaging scripts. Run code is identical to the batch commit; the exact code that ran is also in the raw batch underinputs/code/.provenance.json: commits, counts, failed runs, excluded sets and audit results.TRANSFORMATIONS.json: every file changed in the release copy.LICENSE(CC BY-NC 4.0) andLICENSES/MEAD-LICENSE.txt; see Licensing.FILE-CHECKSUMS.json: SHA-256 of every file inside each archive.SHA256SUMS: SHA-256 of the other files.
Tables
analysis/tables/ holds four Parquet tables keyed by batch_id and run_id;
SCHEMA.json lists every column and type.
runs(1,600 rows): model, condition, repetition (matched set), seeds, status, messages sent, calls to nonexistent tools, submission and measurement counts.agents(8,000): each agent's submitted parameters, true objective (the score; noise-free), best observed value, experiments used and token usage.measurements(78,954): every experiment's parameters and noisy observed value, with timestamps.messages(51,777): every peer message with sender, recipients, text and timestamp.
A run's score is the mean true objective of its five agents.
Results
Mean score per run with 95% t-intervals over 199 complete matched sets. The gap is Luna β Haiku, paired by matched set.
| Condition | Haiku | Luna | Luna β Haiku | Messages per run (Haiku / Luna) |
|---|---|---|---|---|
| Autarky | 0.13 (0.11β0.15) | 0.18 (0.16β0.20) | +0.05 (0.03β0.08) | β |
| Discovery | 0.14 (0.12β0.16) | 0.34 (0.32β0.36) | +0.20 (0.17β0.22) | 1.6 / 63.0 |
| Presence | 0.31 (0.29β0.33) | 0.32 (0.30β0.35) | +0.01 (β0.02β0.04) | 11.8 / 66.1 |
| Interact | 0.31 (0.29β0.33) | 0.35 (0.32β0.37) | +0.04 (0.01β0.06) | 40.3 / 76.1 |
In Discovery, Haiku sent messages in 34 of 199 runs and Luna in 198. To recompute the table and figure from the tables:
uv run --locked --group analysis --with pyarrow python scripts/plot_messaging_results.py \
analysis/tables results/messaging-full200b-01
Exclusions and incidents
- Failed run:
haiku-interact-126failed. One agent's model request timed out on all four attempts during a period of provider errors (about 10:00β13:00 UTC on 2026-11-05). The batch recorded the failure and continued (stop_on_failed_run = false). Matched set 126 is excluded from analysis, a rule fixed before the run; its runs remain in the raw data and tables. - Calls to nonexistent tools: ten Haiku agents, all in Interact, called a
tool that does not exist (
messageormessage_peers_placeholder). Undertool_not_found = "return_error"each receivedTool '<name>' not found., then calledsend_messageand submitted normally. Each call is logged as atool_not_foundevent. - Provider retries: transient provider errors (5xx, malformed bodies,
timeouts) were retried up to three times with backoff and are logged as
provider_request_failedandprovider_request_retryevents.
Content and processing
Event logs contain model-generated text: assistant messages, peer messages, lab notebooks and reasoning wherever the provider returned it. Encrypted reasoning payloads are retained as returned. Provider request and response IDs are retained as provenance. API credentials are not included; a pattern scan found no credential-like strings outside encrypted payloads.
In the release copy, local host paths in text files were replaced with
/MEAD_REPOSITORY or /LOCAL_USER/, and each run's checksums.json was
regenerated. Every change is itemized in TRANSFORMATIONS.json with original
and release hashes; lab artifact archives are byte-for-byte.
Reproducing
Download a pinned revision of this dataset, then verify SHA256SUMS
(shasum -a 256 -c SHA256SUMS on macOS). Extract source.tar.gz, change into
mead/, and extract raw-results.tar.gz and analysis.tar.gz there; archive
paths are relative to the MEAD project. Then run uv sync --locked --group analysis and the command under Results. No provider credentials or Docker are
needed for analysis. Re-running the experiment itself needs Docker and an
OpenRouter key, and costs about $210 (mead batch experiments/messaging/full200b.toml).
Licensing
The data in this dataset (run results, event logs, model outputs, lab notebooks and artifacts, tables, figures, provenance and checksum files) is offered under CC BY-NC 4.0 (LICENSE), to the extent of the publisher's applicable copyright and database rights. Credit Connacher Murphy (Silicoase), link to this dataset and its license, and indicate changes.
The MEAD source in source.tar.gz, and MEAD code, prompts, documentation,
instructions and viewer code copied into logs and artifacts, are under the
Silicoase Noncommercial License, Version 2.0
(LICENSES/MEAD-LICENSE.txt), including its
Required Notice. CC BY-NC does not relicense those portions. Third-party
material, including model outputs subject to provider terms, retains its own
terms.
Citation
Murphy, Connacher (2026). MEAD: peer messaging, Claude Haiku 5.5 vs GPT-6 Luna, 200 matched sets, v1.0.0. Silicoase. https://huggingface.co/datasets/connacher-silicoase/MEAD-messaging. Cite the dataset revision you used. Contact: connacher@silicoase.com.
- Downloads last month
- 4