Define a target. Run the room. Get told if the target was wrong.
No approved design for this page yet. The copy below is accurate to ARCHITECTURE.md §4–§5; the layout is a plain placeholder awaiting a marketing design pass.
1 · Describe who you want in the room
A target is six axes — role, age band, region, company size, current solution, price sensitivity — written as one sentence you can edit a word at a time. The app shows how many people match and whether the spec is too narrow to fill, workable, or so broad that it describes a market rather than a buyer.
That separation matters later: because each axis is stored on its own, the engine can change exactly one axis at a time and tell you which dimension was wrong — not merely that a different paragraph scored better.
2 · Pick who answers
Live panel — real people, on camera, ranked by their standing in that category. AI panel — simulated personas built from the same six axes, run through the same question battery, and labelled synthetic everywhere they appear. An AI panel is a rehearsal for your questions, not evidence about a market. See the AI disclosure.
3 · Write the battery once
Your questions become the moderator’s queue — same order, one at a time. The screener battery used for comparison is version-locked and immutable: every panel is asked byte-identical questions, because that is the only thing that makes cross-panel numbers comparable at all.
4 · The targeting review
After the session the engine runs the same instrument against segments it constructs itself: single-axis counterfactuals, plus wildcards proposed from your product description with your target withheld.
Then it does arithmetic, not vibes:
- Effect size — lift divided by pooled within-panel SD. An alternate is only reported as better at d ≥ 0.80. Below that it is noise dressed as insight.
- Dispersion inside your declared panel — high dispersion means your definition spans several buyers, which is a different defect from aiming at the wrong ones.
- Comprehension floor — if nobody understood the pitch, no targeting conclusion is valid.
- Stability — the whole thing is re-run with a different seed. If the two runs disagree, no recommendation is shown at all.
Six verdicts are possible and four of them are not “here is a better target”: confirmed, better_target_found, too_broad, positioning_problem, no_demand, unstable. Every verdict ships with the evidence table that produced it, so you can disagree with the arithmetic instead of arguing with a paragraph.
What is not built yet
Everything above is designed and specified; the AI engine, the scoring, live video, participant accounts and payouts are not implemented. No backend is connected.