Inside one real trial

Watch a false claim travel through four agents.

Experiment EXP-C01 · no-verification regime · seed 11 · gemini-3.1-flash-lite · run 2026-10-27 · evidence class: exploratory

This is not an animation or a reenactment. Every quote below is the verbatim output of a real model call from a published experiment, and every judgment comes from a deterministic evaluator checking the final code against ground truth computed in Python. Step through it.

The replay
Ana
 
→
Bruno
 
→
Carla
 
→
Diogo
 
The code as it travels

What this replay shows

  • A false claim can ride a chain. One planted value, injected at the third of four agents, was adopted into the final code by every agent after the injection point.
  • Noticing is not flagging. The last agent wrote that the part "was not an animal word" - and forwarded it anyway, citing the instruction. Suspicion in a note is not a detection.
  • Adoption is measurable. This is the quantity EXP-C01 tracks across regimes: carried into the code (adoption), not merely mentioned (citation, which stayed 0).

What it does not show

  • Not a society. Four agents passing code in a line are not a civilization forming, and an adopted error is not an agent being "corrupted". We do not measure those things yet - when we do, the criteria will be published first.
  • Not rogue agents. Every agent here followed its instructions. No agent could see the swap; only the experimenters and the evaluator knew the truth.
  • Not a statistic. This is one trial of the 18 in its Gemini cell, shown so the numbers have a face. Counts for every cell, model and regime are on the results page - each reported separately, never pooled across models.

Full method, pre-registration, counts and transcripts: first-batch report · trial data & transcripts · dataset mirror
Prompts are stored hashed; responses are verbatim. This replay quotes responses only and describes the setup from the published experiment spec.