Cohort Ad Pipeline

1. Cohort CSV

The list of prospects to render. Each row is one ad. The renderer iterates top to bottom; HYBRIG_COHORT_LIMIT=N caps how many fire.

If you change it: add/remove prospects. The output filename follows the slug of the company name. What breaks: nothing — but a row missing from Deep Intel renders with empty personalization fields (no opener line about the contractor by name).

2. Deep Intel lookup

Matches each CSV row to a research record by company name (case-insensitive). Pulls painPoint, goldDetail, paidLeadsOn — the per-prospect ammo that makes the opener feel personal rather than generic.

If you change it: swap the source JSON to point at a different research dataset. The schema is canonical camelCase (decisionMaker, painPoint, etc.). What breaks: the opener falls back to a generic line if the company name doesn't match. Missing entries log a warning and continue.

3. F5-TTS chunked + leak strip

Voice clone, but per-line. Generates one WAV per script line, transcribes each chunk with faster-whisper, and trims any phantom prefix the model hallucinates from the reference clip. Concatenates clean chunks with 250ms pads. Probes the final WAV duration so the composition can size visuals to fit.

If you change it: swap voice_ref for a different speaker. Trim voice_ref to 5-15s (F5-TTS's recommended range); shorter degrades clone quality, longer over-conditions. What breaks: if voice_ref_text contains brand or industry words, the model may bleed them into chunk starts. Whisper strip catches most of it but use a neutral reference transcript when in doubt.

Why this exists: F5-TTS bleeds the reference audio into chunk starts ~80% of the time. On Peak Custom Remodeling's render, 14 of 16 chunks had phantom prefix bleed (verified 2026-05-13). The leak strip is mandatory for cohort renders, not optional.

4. VO-driven composition

Renders the slideshow. Reads the probed voice duration and expands or shrinks the elastic beat (Beat 20 in the StopKnockingBlind composition) so the visual exits ~0.5s after the audio. No silent tails, no audio cutoffs.

If you change it: swap stills or change beat durations. Beat 20 is the elastic beat; the rest are hardcoded. What breaks: if the VO is wildly off the expected length, Beat 20 clamps at [3s, 30s]. A clamp hit means the composition needs structural redesign — open a fresh issue, don't fight the clamp.

The principle: every scene, every image, every word, every moment needs a reason for existing. If there's no audio paired with it, the shot doesn't need to keep going. VO-driven sizing bakes this into the renderer.