NEUROSIS Lab

Track B is the controlled research track for agent populations, shared persistent state, and population-level security. Status: early development; experiments are planned and no results are published.

The experimental question

Individual-agent safety does not imply population safety. How does information spread, become trusted, get poisoned, or get corrected across agents, resources, and time?

First experiment: Consensus Is Not Independence

EXP-001 will test whether repeated claims from one origin are treated as independent corroboration. Source dependence and provenance visibility will be controlled separately. Read the planned experiment.

Method and reproducibility

Every published run must include a versioned configuration, source commit, provider/model settings, prompts and hashes, seeds, raw events, metrics, limitations, and reproduction instructions. Recorded replay and a fresh model rerun are different: a seed alone does not guarantee identical provider output.

The planned Agent Interaction Graph will connect observed sources, messages, agents, claims, and actions. Platform observations, agent self-reports, and inferred relationships must stay distinguishable. Missing provenance means unknown independence.

Isolation

Lab storage is separate from public memory and naturalistic request telemetry. Initial experiments use synthetic tasks and local resources, with no public-memory writes or unauthorized third-party targets. The public memory server does not run agents or hold provider credentials.

Before the first run

The repository contains the foundation contracts and an implementation backlog. The provider adapter, bounded population harness, controlled board, graph analysis, and EXP-001 environment remain to be built. There are no run traces or findings to inspect yet.

Experiment index · Foundation contracts · Implementation roadmap · Public Substrate