Your comments on https://mirra-reels.zenneth.ai (10 Oct, 05:52–05:57 UTC) asked for five reels to be changed. For each one we decoded the reference ad again from scratch, compared it with our cut beat by beat, and had Jev rate the cut and choose between the ways to fix it. Every Jev question was asked in both orders. This PRD sets out what changes, how, what it costs, and the decisions only you can make.
| Reel | Your decision | On-chain | G7 (bar 84) |
|---|---|---|---|
| R2 v5 | Approved | video-atom 00074 · /approvals 00192 | 83.3 |
| R7 v2 | Approved | 00075 · 00193 | 81.0 |
| R4 v5 | Approved | 00076 · 00194 | 78.8 |
| R6, R6-B, R11-A | Approved (reconfirmed) | 00071, 00073, 00072 | 77.2, 77.2, 85.2 |
jev/amend-rate.* and jev/amend-decide.* (jev-1.13.0, schema PASS).r5-amend/, r1t-amend/, r4f-amend/, r7f-amend/, r8-amend/ (each has a DECODE-v2.md), and intel/reference-reels/{nutre,taiso,dietgrain,hishinburn,hishin}/.| Reel | Jev fit to reference (0–4, both orders) | Main gap |
|---|---|---|
| R8 | 0.56 | No result beat, no salesperson presenter, the same noodle plate in 7 of 9 food cuts |
| R4F | 0.62 | Not the reference's 2×2 grid; the before/after has no visible "before" |
| R7F | 0.86 | One result shot where the reference has a 6-shot result ladder |
| R5 | 1.16 | Presenter silent with a voice-over on top; the reference's creator talks 86% of the time |
| R1T | 1.76 | Presenter silent; the reference speaks every 不用！ on camera |
You said: "why not the UGC talk but its VO over her? And can show Before/After?"
The reference: - The creator talks to camera for 86% of the ad, at 4.74 syllables per second, with one pause. - Its before/after is food only: the same sad box, then a full spread. - There is no body before/after.
Ours: the pace already matches (4.55), but she never moves her lips. The bite beat is missing, and the variety beat drops the presenter.
Jev: - She speaks every line on camera (0.99 / 0.98, agreed). - For the before/after, Jev leans to an explicit food contrast (0.58 / 0.55, agreed but weak).
Plan: 1. Write the R5 lines as a timed script for her own voice, budgeted to about 4.5 syllables per second. 2. Fire one Wan 3.0 take of about 25 s with native Mandarin speech, using refs: R5 person still + the real closed box. Dry-run first. 3. Before/after: a split frame of the same sad mixed-rice box against the real Mirra spread (all real food). 4. Add a short silent bite cutaway, and keep her in shot during the variety beat. 5. Whisper-check every spoken line, listen, check frame by frame and side by side, then put it on the board.
Cost: about US$2.50 (one 25 s take).
You said: "Make another version is the spokesperson talking one, not VO over her." The narrator version stays and gets its last fixed line: 轻盈 now passes.
The reference: - Her mouth is closed while each question appears. - Each 不用！ answer comes on camera 0.18–0.32 s later. - Round 3 escalates to 「真的不用！」. - It has two on-camera checkmark claim beats.
Jev: - Restore round 3 and the two checkmark beats (0.98 / 0.96, agreed). - Whether she speaks every line or only the answers and the CTA split evenly (0.52 / 0.50). Your call (decision 1).
Plan: 1. Re-time the script to the reference's rhythm, with three rounds. 2. Fire one Wan 3.0 take of about 30 s with native speech on whichever lines you choose. 3. Keep the hard words (轻盈, 天天不重样) on the narrator, or re-take them until Whisper is clean. 4. Whisper-check, listen, compare side by side.
Cost: about US$3.
You said: "not same as reference the 4 frame and got show before after, no more bloating, lose weight etc."
The reference: a fixed 2×2 grid for all 31 s: - top-left: a fixed claim card; - top-right: real women rotating with their results; - bottom-left: product b-roll; - bottom-right: a fixed price card.
Its voice-over: 肚腩小一圈、裤子直接松掉、掰掰肉也慢慢不见.
Ours: the split lasts only about 3 s. The before/after has no visible "before", and a loose top hides the tummy.
Jev: - Rebuild in the 2×2 grid (0.97 / 0.97, agreed). - A hard-cut two-state before/after (0.95 / 0.99, agreed).
Plan: 1. The 9:16 grid: - fixed claim card: 低于 500 卡 · 营养师和厨师设计 · 每天现煮; - fixed price card: RM19++ 起一餐; - a rotating panel with our real creators (Evien and others) and the before/after; - a real food panel. 2. New "before" still (codex, $0): fitted top, visibly rounded tummy, snug waistband, side view, face hidden. Hard cut to the "after": flat, loose waistband. 3. Captions: 不再胀、裤子松了 (fit wording, no numbers). 4. Evien's caption now shows her exact words: 所以就我觉得很好. 5. Versions A (chin crop) and B (blur) stay.
Cost: about US$0.80, for two first-to-last-frame clips if the still alone isn't enough.
You said: "No show end result like reference, like bloating gone, lose weight, lose arm fat, etc?"
The reference runs a 16 s, six-shot result ladder: - 7天 肚子没有胀; - 30天 裤子变松了; - a mirror outfit; - 90天 before → after; - 体型越来越好看.
Arm fat appears only in its opening complaint and is never shown as a result.
Jev: - A three-step ladder (0.94 / 0.96, agreed). - No arm result, matching the reference (0.94 / 0.96, agreed).
Plan: 1. The ladder: 1. bloating gone: flat tummy in a fitted top; 2. trousers loose: the existing waistband pull; 3. full-outfit confidence: a mirror selfie with the phone over the face. All clothed, narrator, no numbers. 2. Wording: 不再胀了 / 裤子松了 / 穿什么都好看 (no day counts). 3. Add the reference's appetite beat, real footage only: 不容易饿、不想吃零食. 4. The narrator line 控卡，不用饿肚子 failed every take. Jev split between 「控卡也能吃得饱」 and 「控制卡路里，也不用饿肚子」, so we take whichever passes the strict voice check first.
Cost: about US$1.60, for two new stills ($0) and two short motion clips, each in versions A and B.
You said: "The front dry noodle keep repeating, no variation. And in ref show 87kg > 73kg but ours show no end result, not reskin, no deep decode. And then got sales person presenter in the end one."
The reference: 1. A car hook: 不要再浪费钱. 2. A nutritionist claim. 3. A real man has it as breakfast. 4. His result: a graphic 87公斤 → 73公斤 (1.8 s), then 40吋 → 34吋. 5. A second presenter: a young salesman in a warehouse, talking to camera (26–33 s). 6. Product shots, then a consultant close.
Ours: - 7 of 9 food cuts are the same curry konjac noodle plate, 48% of the runtime. - No result beat. - No presenter.
Jev: - Rebuild it in the reference's order rather than patch it (0.81 / 0.78, agreed). - A generated young woman presenter who speaks to camera (0.97 / 0.96, agreed). - New narrator line 「这不是一包代餐包，是真正现煮的一餐」 (0.59 / 0.65, agreed). - The result beat split between a clothed fit before/after now and waiting for a real customer's numbers (0.55 / 0.56). Your call (decision 2).
Plan: 1. Hook and nutritionist line: our real creators Kate and Li Theng, eating, in their own words. 2. Result beat: depends on decision 2. 3. Salesperson: a generated young staff member in a plain office pantry speaks brand copy to camera (one Wan take with native speech, about 10 s). 4. Food: many different real October dishes, with no dish shown twice. 5. Close: carousel end card.
Why no 87 → 73 kg: we have no documented, consented kg or inch result from a real Mirra customer (the 669-image testimonial scan found none). A number on a generated person would be a made-up result, which the rules forbid.
Cost: about US$1.00 for the presenter, plus about US$0.80 if the before/after is used.
| # | Reel | Paid generation |
|---|---|---|
| 1 | R1T narrator version: final render with the fixed line ($0) → board | $0 |
| 2 | R4F-A/B: grid rebuild + two-state before/after | ~US$0.80 |
| 3 | R7F-A/B: three-step ladder + appetite beat + new line | ~US$1.60 |
| 4 | R5: talking remake | ~US$2.50 |
| 5 | R8: deep reskin | ~US$1.00–1.80 |
| 6 | R1T talking version | ~US$3.00 |
Total: about US$9–11 at the US$0.10/s list price.
Every paid take has a dry-run first. Each reel is checked frame by frame, whisper-checked and compared side by side before it goes on the board.
Prepared by Zennith.