About SpareBrain

Most “LLM as user” tools ask one model to play a persona and report back. The answer sounds reasonable — and collapses the variability of real people into a single confident voice. SpareBrain runs a cohort instead: simulated users whose underlying conditions are rolled by seeded dice outside the model, so the model embodies variance rather than inventing it.

Why cohorts, not personas

Each cohort member gets a substrate — mood, financial pressure, prior frustrations, expertise, attention, age, place — sampled per member by external randomness across six layers. Persona templates shape the probabilities, never the outcome: no combination is impossible, so the affluent-but-anxious and the skeptic-who-signs-up still occur, the way they do in real samples. The model’s job is to inhabit the hand it’s dealt.

This is measured, not assumed: against a conventional persona-only baseline on the same stimulus and model, nine of fifteen baseline responses opened with the identical sentence; the substrate cohort produced fifteen distinct openers and roughly three times the length variance.

How to read a study

Splits carry the signal.Every generalisation must name its numbers (“9 of 15”). A collation that smooths disagreement into consensus is rejected and regenerated automatically.
The dissenting voice is verbatim. It is checked character-for-character against a real cohort response — never a paraphrase.
Strong consensus replicates; mid-range counts wobble. In same-seed replication tests, findings held by 12+ of 15 members were stable, split magnitudes moved by ±2, and counts of spontaneously mentioned themes swung widely. Read those as “this objection exists,” not “this many hold it.”
Uncalibrated means uncalibrated. Until a use case has been validated against real research, every output is exploratory — a hypothesis generator, not a finding.

What it refuses to do — until calibrated

Purchase-intent prediction, willingness-to-pay, and conversion forecasting are structurally refused. Not because synthetic prediction is impossible — published work shows it can correlate with human panels in specific, carefully calibrated settings — but because those numbers don’t transfer: not across products, not across audiences, and never from an uncalibrated cohort. A number that looks like a forecast will be treated as one, so until a use case has earned trust against your real data, none is produced. The refusal runs deep: even inside legitimate studies, a deterministic check strips intent tallies (“11 of 15 would subscribe”) out of collations and demands the objections and reasons instead.

Reproducibility

Every study records its seed. The same seed re-deals the identical cohort — same substrates, ages, and locations — so duplicating a study and changing one thing (the copy, the model, the template) is a controlled comparison, not a vibe check. Randomness enters once, at study creation, and never inside the model.