P1 Discovery — v3

1 · Captions were out of sync. They were timed as clip-duration ÷ caption-count, so drift was arithmetic, not bad luck. Now every caption is placed on word-level timings from forced alignment (254 words, alignment loss 0.018), chunked to real subtitle limits — 15 characters per second, 0.45–4s, ≤26 characters a line.

2 · No build-up. Pacing now varies deliberately. Setup holds long; the agitation beat is cut into three tightening shots inside one narration line; the pivot is nearly still — a held breath; the reveal is a hard cut with a long hold, and the music drops out 1.6s before it and swells after. Camera moves vary (push, pull, pan, hold) instead of one identical zoom.

3 · Flat voice. The whole script is now one continuous take with direction written into the text. v3 rejects both of its continuity mechanisms, so twelve separate generations could never have sounded like one person — that was structural, not a settings problem.

4 · "I don't see what you're saying." The narrowing is now a staged event: 500 pins, three criteria chips, three culling waves where rejects exit on one shared vector while the counter ticks 500 → 105 → 64 → 37, and the survivors bloom gold without moving. Every number is real, computed from the register today.

5 · The builder section didn't connect. Every joint in the script was a "therefore" or a "but" except one — that seam. It's now caused by the beat before it: you have thirty-seven, but you cannot visit thirty-seven and on paper they're identical, therefore the same register is worth reading twice. It also has its own demonstration now: four promise bars extending in red with +14, +22, +9, +31 months.