How it works · Windrose Atlas
The instrument

One paragraph in. A compass out.

No survey to design, no panel to recruit. You describe an idea in plain language; a simulated population responds; you leave with a direction and the exact next test. Here is each step, and the shape of what you get back.

STEP 01

Describe your idea in plain language.

One paragraph is enough. Add up to three variants to compare, or let the instrument propose two contrasts for you. Pick an audience. No brief, no scale design.

STEP 01·B · OPTIONAL, AND STRONGER

Have data? Bring it.

Drop in a CSV or spreadsheet. The wizard maps your columns, clusters your real customers into personas, and marks which dimension of the answer your data can actually verify. Where it can, the result reads REAL instead of SIM-RANK. Grounded mode is the same instrument, one rung higher.

STEP 02

A population responds, not an average.

1,000 personas, parameterised by documented cultural frames, react through a cognition model. Order of options is counterbalanced 50/50, and the run resamples so we can measure how firmly the ranking holds.

STEP 03

You get a compass, not a verdict.

Which variant is stronger, for whom, and why, with an evidence label stating how far to trust it, and the three hypotheses worth testing in the real world next. A run is a queued job, not a chat reply: the population responds, the ensemble resamples, and the result lands in minutes with its label attached.

Anatomy of a result

Every result reads in the same order, every time.

102 · Simulation, directional
2RANKING
01
"Office espresso, delivered"STRONGEST
02
3"Taste your city's roasters"
53 HYPOTHESES TO TEST
01 · Founders miss the ritual, not the caffeine.
02 · A flat price beats variety.
03 · Frame the first delivery as a "first cup".
6
SaveShareExport

The order is fixed on purpose, on the landing, in the app, in exports and in a shared link. You always meet the evidence before the conclusion.

  1. 1Stage header - which rung this stands on, by number and colour.
  2. 2Ranking - variants strongest first, the winner tagged.
  3. 3Distributions - a curve per variant, never a lone number.
  4. 4Evidence label - always attached, black-and-white by rule.
  5. 5Hypotheses - the bridge to the real world.
  6. 6Actions - save, share, export, run again.
Every screen, four states

Including the one most tools hide: "we don't know."

STATE · EMPTY
No runs yet.

An invitation, not a shrug. Your first one takes a few minutes.

New run
STATE · RUNNING
SAMPLINGsamples: 412 / 1,000

Order of options is counterbalanced 50/50 to remove position bias.

STATE · RESULT
A ranked, labelled read.

The full anatomy above: ranking, distributions, evidence label, hypotheses, actions, in that order.

STATE · WE DON'T KNOW
We don't know, and here's why.

When a scenario falls outside our backtested classes, we say so. The label reads BACKTEST: OUT OF CLASS, we point you to what you can do instead, and you are not charged. This state is shareable, like any result. It is our proof of honesty in practice.

Under the hood · architecture

The whole instrument, from the outside in.

The whole instrument - three engines, an auditable file spine, the 8-stage loop. Every stage writes a file the next stage reads, so a run is resumable, auditable, and never a black box.

THE 8-STAGE LOOP
s1 ingest → s2 personas → s3 design → s4 simulates5 grounds6 calibrates7 models8 explain
stages talk through files - resumable, auditable

See the full architecture →

Minutes in, a compass out.

Your first run is free, no card. See the method behind it on the Science page.