Observer Zero: an instrumented artificial world for studying scientific belief revision in autonomous LLM agents (1.0.0)

Meridian is a deterministic artificial world in which autonomous LLM scientist agents run
experiments on instruments, exchange letters, and maintain explicit probabilistic beliefs, while
a physical constant changes covertly mid-run. The true state is known only to the simulator, so
whether anyone notices becomes a measurement rather than an interpretation.

The platform separates three things that are usually confounded: what could have been known from
the observations available, what a society actually gathered, and what it concluded. A non-LLM
sequential change-point detector — given no world parameters, not the shift magnitude, not the
onset day, not which instrument class carries the signal — prices the evidence present in the
agents’ own chosen measurements. A frozen evaluation layer then scores their final beliefs
against ground truth using a thirteen-class hypothesis taxonomy and a provenance-checked claim
evaluator.

Two design decisions make the social results measurable. Measurement noise is keyed by the triple
(world seed, instrument, trial index), so a society can be re-run against an identical universe,
which is what makes paired-by-seed contrasts possible. And communication is voluntary throughout
— no scheduler requires an agent to speak — so communication initiation is itself a measured
dependent variable rather than an assumed architectural feature.

Each site hosts a pendulum and a resonator, giving a causal discrimination structure: a change in
gravity moves every pendulum and no resonator, a site-local environmental cause plausibly moves
co-located instruments of both kinds, and a single-rig fault moves one instrument. The
observation stream therefore supports diagnosis, not merely detection.

The distribution includes seeded batteries, versioned manifests hashed into every run artifact,
information-flow security enforced by type, and a scripted non-LLM society that reproduces
bit-identically as a measurement-validation baseline. Model providers are pluggable and include a
zero-cost mock provider that exercises the full pipeline without inference spend. Documentation
follows the ODD protocol (see ODD.md).

Release Notes

First public release, v1.0.0. This is the version underlying both reported studies: Study 1
(two-agent societies, 150 runs) and Study 2 (eight-agent societies, five arms, 85 runs,
pre-registered and frozen at commit 85bcdfb, tag study2-freeze).

Includes the deterministic Meridian engine, agent loop and belief layer, the pluggable model
provider interface (Anthropic first-party, Amazon Bedrock, Perplexity, and a zero-cost mock),
the frozen evaluation layer with its thirteen-class hypothesis taxonomy and provenance-checked
claim evaluator, the three-level non-LLM detector benchmark, and 172 tests.

The full battery pipeline runs at zero cost through the mock provider, and the scripted mock
society reproduces bit-identically, so the evaluation layer can be exercised end-to-end without
provider credentials or inference spend. Live agent runs need API keys supplied via a local .env
file, which is not included in this archive; see .env.example.

Associated Publications

Lamb, N. (2026). Observer Zero: Do LLM Agents Form Epistemic Communities? Evidence from autonomous agents in an instrumented world. Submitted to the Journal of Artificial Societies and Social Simulation.
10.5281/zenodo.21872780