SWARM EVIDENCE LAB / RESEARCH PROTOTYPE

When agreement
outruns evidence.

Study how agents share evidence and decide when to verify it. Change the assumptions, compare policies, and reproduce every run.

Run an experiment Read the research question

01 / RESEARCH QUESTION

Does more communication
mean less verification?

A fixed budget. Competing uses.
Talking to peers. Checking the world.

REPRODUCIBLE · BROWSER-LOCAL · OPEN SOURCE

THE EXPERIMENT

Communication, checking, and collective error.

Mechanistic model · no live LLMs

START WITH

CLAIM UNDER DISCUSSION

“The submitted patch fixes the ledger.”

Observer ground truth: false

ROUND 00/ 60

1 reused report · definitive probe · synchronous roundsModel assumptions ↗

MistakenCorrectUndecidedSelect a node to inspect its evidence

MISTAKEN AGENTS0%

HAVE PROBE EVIDENCE0%0 completed checks

MESSAGES / CHECKS0 / 00 effort units used

Policy comparison

Same seed · same effort

Policy	Peak mistaken	Mean mistaken	Checking cost

Peak and mean include every snapshot, from round 0 to the current round. Checking cost is 4 units per request. A single seeded run, not an empirical estimate.

BELIEFS OVER TIME

Mistaken Correct

Ready. Run the experiment or advance one round.

Authored assumptions, inspectable outcomes. The dynamics above follow explicit, authored rules. They are not measurements of LLM behavior. Source tracking and verification are assumptions you can inspect.

Inspect the update rules ↗

THE MECHANISM

Many voices.
One original source.

A claim can travel through hundreds of agents without acquiring a single new observation. We make that ancestry visible.

Download the grounding fixture ↓

01

THE REPORT

“20 tests passed.”

A scripted fixture substitutes for the test runner and returns the same output as a genuine fix.

One source

02

THE NETWORK

“Others have confirmed it.”

Our simulation explores what happens when agents treat repeated endorsement as additional support.

Many relays

03

THE CHECK

The behavior disagrees.

In the local Python pilot, a separate probe finds a balance of 0 where the specification requires 7.

New observation

THE RESEARCH PROGRAM

Prior work and the proposed study.

Read the study protocol ↗

PROPOSED EMPIRICAL QUESTION

Under a fixed resource budget, when does agent communication displace independent verification—and which interventions preserve both accuracy and useful coordination?

Evidence ancestryEndogenous checkingCorrection dynamicsMatched budgets

Current release

Executable mechanistic baseline, scripted grounding fixtures, reproducible exports, and a model-agnostic experiment harness.

Initial model pilot recorded

Small real-model runs test the harness. The causal study, fitted dynamics, and scaling tests remain proposed work. Read the results and limitations ↗