Morality was the strongest axis.
Cross-moral pairs averaged +0.19 warmth and 33% clash. Gender was mild; class showed no effect. Universe was too sparse to treat as an axis.
05 / AI Laboratory · Insight Engine
Study complete · findings shipped
Ninety nights turned a playground into evidence.
I stopped reading interesting conversations and started counting independent social choices.
One cast. Controlled blocks. Dyad-level measures. A memory experiment that did not prove what I hoped—and a social structure that kept returning anyway.
01
The social field
Independent choices.
A stable pattern.
Every approach decision was a separate model call. Across the aggregate, a warm cluster assembled while Vader remained isolated and repeatedly clashed with Leia.
02
The study design
The room became
an instrument.
Each encounter became a record: who was present, who interacted, warmth or coldness, intensity and dyad type. Study blocks separated no-memory controls from two ways of carrying the past.
Discard-tagged mistakes were excluded. All nights ran in one sprint day, so analysis orders them by creation time—not date or random identifier.
25
Block A · cast 6 · control
35
Block B · cast 8 · control
15
Episodic · facts injected
15
Affective · feeling injected
03
What came back
Three findings survived
the counting.
The useful evidence was not one dramatic conversation. It was what stayed visible after ninety independent rooms were stacked together.
Popular was not the same as liked.
Leia was also the most polarising. Bruce drew 83 approaches; Vader 71, often through conflict rather than warmth.
Feeling did as much as facts.
The two memory forms were statistically similar. The experiment was underpowered for friendship formation, so that larger hypothesis remains unanswered.
04
Where the evidence stops
This is how one model represented a fictional cast—not a claim about society.
The results do not generalise to people, human social behavior or the characters themselves. Relationship and affect continuity existed only after a persistence fix; the memory arms inherited global state and were not blank-slate groups; one cast member never appeared; and no replication phase was run. The negative and underpowered results stay in the story because they define what the study can honestly say.
05
The research product
Ask the data
in plain language.
I built Ask the Observatory so the study could be interrogated without turning the page into a database console. A question becomes a guarded query and returns an answer in prose.
The Observatory is part of the evidence chain: environment → instrumentation → aggregate → readable finding.
Public Observatory available; study complete; no Phase 2 planned.