Ship after simulation.

The front door — everything runs through a room, same as the web app, in two speeds. FAST is audience-ask: the whole room answers one question in seconds — conversation, structured actions, quick reads. DEEP is the URL endpoints below: the same seated members walk a live page in a real browser and report back what they did — journeys, reactions, hesitations, and a head-to-head vote on a compare (minutes, credits-charged). Open the panel for a saved audience, then ask the whole room one question and it fans out to each member in parallel. audience-ask runs structured actions — angle_test / hook_rank (options), script_react (script), ab_compare (2 images), creative_rank (3–5 images) — returning aggregated votes in result. A ranking of 3+ options collects two votes per member: the one they'd stop for and the one they'd scroll past, so result carries tally, scroll_tally and score (tally − scroll_tally), and score decides the winner. Members may decline to vote when the stimulus isn't theirs to judge: those answers carry abstained and cast no vote. A winner is only named when at least half the room voted and the leader clears the runner-up by more than one point — otherwise result reports no_read, close_call or tie, and winner_index is null. Options with a negative score are the kill list. Stimulus comes from audience-upload-image — png / jpeg / webp / gif images (≤ 8 MB) or short mp4 / mov / webm ad videos (≤ 100 MB): the response's kind routes the URL into image_urls or video_urls (up to 5 videos per turn; clips are prescanned at upload so the panel reacts to what's actually on screen, and 2 videos = A/B vote, 3-5 = ranked vote — same counterbalanced tally as images, and a 2-video A/B renders the same verdict card). target_indices @-targets specific members (untagged members may interject once). Manage the room like the web app: dismiss a member for a fresh draw, reset (clear / swap / new), scope turns to campaign threads via thread_id, and rebuild history server-side from audience-room/turns. A "new" reset also takes an optional description of who you want in the room — the members of that audience who best match are seated instead of a random draw, each with the detail that earned the seat, and panel_match.matched reports how many genuinely fit. Members are only ever selected from the audience you built; nobody is generated to match the description, so an audience that lacks those people says so rather than inventing them. A following "swap" walks the same ranking to the next-best seats. Capped per audience to keep the panel honest. A URL question whose page can't be fetched (JS-only apps) returns 422 {code: "page_unreadable"} without spending a message — re-send with allow_unreadable_page: true to have the panel react to the link from expectation instead. Comparing something against itself is caught before the room runs: passing the same file twice returns 400 naming the pair, while the same ad re-exported at another encoding is allowed but reframed — result is null, options empty, no per-member pick, and a duplicate_note explains that the room reacted to one ad instead of picking a winner (an absent result there is not a tie). audience-synthesis may also return an advisory handoff object — {discipline, room_label, reason, and target_audience_id + target_name when it points back at an earlier room} — flagging a specialist read (legal, visual design, …) worth getting next; act on it by creating an audience for that discipline and re-running the question there. When you're ready to go deep, pass the panel's audience id + member indices to the URL endpoints below.

Mine the open web for what real people say about a brand. Asynchronous — POST returns a job_id immediately, poll until status='completed' (~15–25 min), then GET the corpus.

One-shot follow-up Q&A with a single synthetic persona on a focus group. Show them images this turn (image_urls, from interview-upload-image — images only) for a gut reaction. Capped per (focus group, persona, viewer) to keep replies honest.