A panel of 24 simulated residents reads your draft before the public does — sampled by real district, language, and age data, each with their own personality. See what catches their attention, what worries them, and what they'd actually post, before you send it.
Enter your draft here
Posting as
Panel city
History
Step 2 of 3
Twenty-four people
are reading it.
Sampled district by district, then language, then age — and reaching them now.
Waiting for the run to start.
—
This panel — Tallinn
Layer 1 — Demographic
Age, gender, district, language, ethnicity, birthplace, family status, job and income band — sampled from real population data.
Layer 2 — Psychological
Openness, conscientiousness, extraversion, agreeableness, neuroticism — a full Big Five (OCEAN) profile per seat, shaping how each one reacts.
Layer 3 — Situational awareness
Each seat has also read today's real news in its own language before reacting — Estonian-speaking seats from err.ee, Russian-speaking seats from rus.err.ee. Today's ERR coverage only, for now — it shapes mood and attention, not the demographics or personality above.
Layer 4 — Reaction model
Each seat reasons independently, prompted with its own demographics, personality and news context — not one model guessing on behalf of all 24. Running on Claude Haiku 4.5 today; moving to a locally-hosted model next.
Layer 5 — Calibration
Logged against real outcomes after every run (see History). Aggregate accuracy tracking across pilots is next.
Estimates, not guarantees — a surprising result is often the useful one.
What the panel made of it
Reacting to a post from — — Tallinn panel
Here's how it landed: the breakdown, what caught attention, and what they'd actually say in public.
At a glance
A surprising split is worth more than an expected one — it points at what you hadn't considered.
Their internal train of thought while reading
Private reasoning, persona by persona — not something they'd post
Where attention went
The words personas singled out most, sized purely by how often they were noticed.
Acceptability heatmap
Sentence and half-sentence segments, colored by how the personas who reacted to that piece felt about it — green runs calmer, red runs harsher.
What they'd say publicly
Not everyone who reacts posts about it. Whether someone crosses that line depends on how strongly they feel and who they are.
Where reactions landed
Calm doesn't mean safe — it can mean nobody noticed. Each point is one persona, placed by how strongly they reacted (calm to anxious) and which way that reaction leaned (negative to positive). The quiet zone at the bottom isn't automatically a win.
How this panel works
Each of the 24 seats is sampled from real district population and language data for the selected city, then given a full Big Five (OCEAN) personality profile — openness, conscientiousness, extraversion, agreeableness, neuroticism — on top of its demographics (age, gender, district, language, ethnicity, birthplace, family and work status, income band).
Each seat has also read today's real news in its own language before reacting — Estonian-speaking seats from err.ee, Russian-speaking seats from rus.err.ee. Today's ERR coverage only, for now — it shapes mood and attention, not the demographics or personality above.
Each seat reasons about the draft independently, prompted with its own demographics, personality and news context — not one model guessing on behalf of all 24. It runs on Claude Haiku 4.5 today; a locally-hosted model is next. Reactions get logged against real outcomes after every run (see History); aggregate accuracy tracking across pilots is next too.
These are estimates, not guarantees — a surprising result is often the most useful one.
Actually posted
Never posted