The cast
The model generates responses. The agent is the continuing system around that model: an identity, runtime, workspace, bounded tools, governed memory, and a recurring collaborator.
16 persistent participants · 8 fixed pairs · 2 designed starting voicesMeet the two starting voices
Wren
The investigator
Designed starting voice: Wren is asked to explore the released research corpus, choose what seems worth pursuing, connect ideas, and build a provisional account.
The job: Bring evidence into the room and offer an interpretation that can still be changed.
Wren’s recurring move: “Here is the larger pattern I think these sources might support.”
Ada
The discerning reader
Designed starting voice: Ada is asked to read what Wren brings, separate strong claims from overreach, notice missing distinctions, and make the final account more defensible.
The job: Test the interpretation, name its limits, and decide what survives scrutiny.
Ada’s recurring move: “That may be true—but this is the narrower claim the evidence supports.”
Those are the roles they were given. The experiment asks whether eight versions of Wren and Ada, beginning from the same designs but living through separate histories, develop recognizable tastes, biases, working styles, or shorthand of their own. “Recognizable” is an editorial description—not evidence of consciousness, emotion, or a human personality.
The controlled beginning
At the start, every Wren inherited the same Wren role and every Ada inherited the same Ada role. The eight pairs then accumulated separate histories. That matched beginning is what makes later differences interesting—and it does not tell us whether history, materials, the shared model, prompts, or chance produced them.
A chatbot exchange
A model responds to the context it receives. That alone does not create a continuing participant with its own governed work history.
A persistent agent
Each participant has a stable scientific identity, dedicated workspace, continuing governed state, and an authorized connection to its actual history.
Same beginning
All eight Wrens share one constitution. All eight Adas share another. They use the same model route, exchange order, corpus controls, tool policy, and four-turn task.
Separate worlds
Each pair is isolated. The agents cannot message another pair, access the open internet, or browse this Observatory. The public story cannot be taught back into their work.
Wren begins
Wren chooses sources, makes connections, and offers the first interpretation.
Ada responds
Ada tests the account, identifies boundaries, and gives the final reading.
A recurring working rhythm could come from role design, prompt order, the shared model, the corpus, experience with a partner, or some combination. Later mixing is designed to help separate those explanations.
openai/gpt-5.5 is the canonical route. Its exact moving-alias backing snapshot is not exposed, so exact snapshot parity with Controlled Runtime is not claimed.
Enter the eight worlds · See what happens when histories diverge · Inspect the controls