The NOMI reel deck

# NOMI + NOMI KIN — Content Audit for the Instagram Reel Deck
**Date:** 2026-07-30 · **Scope:** every video asset in `~/Desktop/NOMI ` and `~/Desktop/NOMI-4k Asset Library`
**Method:** 182 files ffprobed · 82 measured for grade / motion / clarity · every room family reviewed on labelled contact sheets · product and claim truth pulled from `brain_knowledge`

---

## 0. Headline — four findings that change the plan

| # | Finding | Consequence |
|---|---|---|
| **F1** | **Every exterior clip is 16:9 landscape.** 7 of 11 are only 1280×720. There is *no* native 9:16 exterior footage in the library. | A centre-crop to 9:16 yields 405×720 — a 3.6× blow-up. Exteriors **cannot** be cropped. Solved with a letterbox band (§5). |
| **F2** | **Grade spread is 0.152 R/B — more than 2× the visible-pop threshold.** Interiors run warm (R/B 1.09–1.155) and bright (luma 172–196). Exteriors run cool (1.003–1.02) and dark (luma 84–156). | Any exterior→interior cut flashes and shifts temperature. **Every reel must be graded at cut time.** |
| **F3** | **The kitchen is the weakest family and it is defect-compromised.** Median detail 549 — lowest of any room. Its two best-looking clips (K08, K12) carry the known window defect; its two clean clips (K03, K04) are the *softest in the entire library* (detail 240 and 306). | The kitchen reel has to be carried by K01 (pantry pull-out, detail 1739, no window in frame). |
| **F4** | **The existing 30s and 60s TOUR reels are not reusable.** They contain E06/K08/K12 per the pipeline lessons, and measure detail 740 / 837 — *below* the library median of 920. | Rebuild from clips. Do not re-cut the finished tours. |

---

## 1. Inventory

| Location | Files | What it is | Usable for this deck |
|---|---|---|---|
| `masters-4k/coastal-video-9x16 /Approved/` | **54** | 4K 9:16 room + element clips, 5s, 24fps | ✅ **the primary pool** |
| `masters-4k/coastal-video-9x16 /Elements Videos /` | 8 | 6× 10.25s room reels + 30s/60s tours, 4K 9:16 | ⚠️ reels yes as texture, tours no (F4) |
| `masters-4k/coastal-video-9x16 /` (root) | 8 | in-flight masters incl. 2K tiers | ✅ selective |
| `masters-4k/coastal-video-9x16 /_2k-original/` | 81 | 720p/1080p/2K companions of the above | ➖ redundant at 2K delivery |
| `NOMI-4k Asset Library/external-flyovers/` | 6 | exterior flyovers, **1280×720 16:9** | ⚠️ letterbox only (F1) |
| `NOMI /films /` | 5 | vineyard exteriors, **1080p–1440p 16:9** | ⚠️ letterbox only (F1) |
| `NOMI /Talent Images /` | 6 | a woman in cream loungewear | ⛔ **off-model** — see §6 |
| `NOMI /Undefined /` | 4 | older H-series interiors | ⛔ superseded styling |
| `masters-4k/coastal-images-9x16/` | 60+ stills | 4K + 2K plates and elements | ✅ for end cards / stills posts |

Total moving-image runway in the approved pool: **54 × 5s ≈ 4m 30s** of graded 4K 9:16, plus 62s of room reels.

---

## 2. Measured technical profile, by room family

Medians. `noise` = mean |px − 1px blur| on a native 40% centre crop · `detail` = edge variance after a 3px median denoise at a fixed 900px box · `motion` = mean abs frame diff over 0.5s (≈10 reads as a real tracking shot) · `R/B`,`G/B` = channel ratios on a neutral mask.

| family | n | noise | detail | d/n | motion | R/B | luma | contrast |
|---|---|---|---|---|---|---|---|---|
| **living** | 16 | 2.19 | **1307** | 599 | 11.31 | 1.12 | 185.5 | 31.8 |
| **bed1-main** | 5 | 1.94 | 1007 | 519 | 9.54 | **1.13** | 184.3 | **43.3** |
| **kin-bed2** | 16 | 2.10 | 982 | 513 | 10.21 | 1.09 | 184.3 | 38.0 |
| ensuite | 2 | 1.31 | 962 | 758 | 4.54 | 1.10 | 186.5 | 26.5 |
| **bathroom** | 11 | **1.08** | 679 | **820** | 6.57 | 1.08 | **196.4** | 29.8 |
| **kitchen** | 10 | 1.19 | **549** | 687 | 8.61 | 1.10 | 172.8 | 44.6 |
| exterior | 10 | **3.98** | 1208 | **268** | **18.87** | 1.08 | **117.7** | **66.8** |
| room reels | 4 | 2.09 | 1254 | 574 | 22.48 | 1.09 | 185.4 | 40.9 |
| tours | 2 | 1.61 | 789 | 507 | 25.08 | 1.11 | 179.0 | 33.3 |

**Library medians:** noise 1.92 · detail 921 · motion 10.21 · **R/B 1.099** · G/B 1.026 · luma 184 · contrast 37.3 · sat 0.167

### Reading it
- **Living is the strongest family** — highest detail (1307) with mid noise. The slat wall, joinery lines and cane fronts give it fine repeating geometry in focus, which is exactly what the pipeline lessons identify as the driver of perceived quality.
- **Bathroom is the cleanest** (noise 1.08, d/n 820) but only mid-detail — it is *smooth luxury*, not *texture luxury*. Marble and brass read expensive without carrying much edge information. Cut it slow; it does not reward fast cutting.
- **Bathroom is also the brightest** (luma 196) — the closest family to clipping. Anything cut against it needs a lift or the bathroom will flash.
- **bed1-main carries the most contrast** (43.3) and the warmest grade (1.13) — the deeper slate bedding and warm cream read as the "primary suite". **KIN bed2 is visibly lighter and airier.** That contrast is a genuine editorial asset for the one-bedroom-vs-two story.
- **Exteriors are the technical weak link**: noise 3.98 (3.7× the bathroom), d/n 268 (worst), luma 118 (66 points below the interiors), and the highest motion (18.9). They are dark, grainy, cool and fast against bright, clean, warm and slow interiors.

---

## 3. The grade problem, precisely

| | R/B | luma |
|---|---|---|
| warmest interior — `0001-living-pan` | **1.155** | 171.9 |
| interior median | 1.099 | 184 |
| coolest exterior — `EXT2-toowoomba` | **1.003** | 116.3 |
| darkest — `vineyard-EXACT-scene-pan` | 1.022 | **84.2** |
| **spread** | **0.152** | **112** |

The pipeline lessons establish that an inter-shot R/B jump of ~0.07 is *plainly visible* as a temperature pop, and that the K08 pink fix Brad accepted was a 0.047 shift. A 0.152 spread is therefore two visible pops wide.

**Worst offenders (deviation from the 1.099 library median):**

| clip | R/B | dev | note |
|---|---|---|---|
| EXT2-toowoomba-flyover | 1.003 | −0.096 | coolest asset |
| EXT3-herveybay-flyover | 1.007 | −0.092 | |
| vineyard-EXACT-scene-pan | 1.022 | −0.077 | also darkest (luma 84) |
| `0002-living-tvwall-pan` | 1.036 | −0.063 | *the library's best clip* — and it is cool |
| `0001-living-pan` | 1.155 | +0.056 | warm ceiling, unfixed panel defect |
| `B01-bed1-main-bedside` | 1.148 | +0.049 | |
| `N10-kin-bed2-ottoman-rug` | 1.147 | +0.048 | |
| `K01-kitchen-pantry-pullout` | 1.136 | +0.037 | both variants |

**Decision: grade every reel at cut time to R/B 1.099 / G/B 1.026, attenuate-only** (gains normalised so the largest is 1.0, which mathematically cannot clip), applied inside the segment filter during the cut so the masters stay pristine. Exteriors additionally get a luma lift toward 150 so the cut to an interior is a ramp, not a flash.

---

## 4. Clip ranking — the heroes and the ones to keep off the front

### Top 12 by detail (the "looks expensive" clips)
| detail | noise | motion | R/B | clip |
|---|---|---|---|---|
| **3510** | 5.00 | 4.80 | 1.090 | `0009-kin-bed2-walkin-robe-display-panright` |
| **3460** | 3.83 | 5.36 | 1.093 | `0005-bedroom-cane-robe-vanity-panright` (4K) |
| **3015** | 4.11 | 11.69 | 1.128 | `E04-element-panright-dyn` (cane TV cabinet) |
| **2861** | 4.91 | 30.83 | 1.143 | `N04-kin-bed2-cane-wardrobe-door-pullback` |
| **2656** | 3.09 | 6.43 | 1.116 | `0002b-living-tvon-hero-panleft` |
| **2596** | 2.82 | 14.84 | 1.081 | `0002R-living-sofa-corner-pushin` |
| **2442** | 4.26 | 8.97 | 1.084 | `0010-kin-bed2-walkin-robe-angled-pullback` |
| 1932 | 4.05 | 16.28 | 1.099 | `0003-master-D3move-z095` |
| 1862 | 3.25 | 28.12 | 1.108 | `E01-E09-living-elements-reel-dynamic` |
| 1739 | 1.95 | 21.95 | 1.136 | `K01-kitchen-pantry-pullout-pushin` |
| 1690 | 1.92 | 18.17 | 1.036 | `0002-living-tvwall-pan` |
| 1550 | 1.78 | 31.29 | 1.136 | `K01-kitchen-pantry-pullout-pullback` |

**The cane and rattan shots dominate.** 0009, 0005, E04, N04, 0010 are all woven-cane subjects, and woven cane is the single densest repeating texture in the product. That is the visual signature to lead with.

### Bottom 10 by detail — clean but soft, never a hero shot
| detail | noise | clip | verdict |
|---|---|---|---|
| **240** | 0.19 | `K03-kitchen-cooktop-rangehood-pushin` | native render, pristine, almost no edge information |
| 285 | 0.28 | `K03-cooktop-rangehood-panright` | as above |
| 299 | 0.12 | `BA1-mirror-basin-pullback` | d/n 2486 — the cleanest clip in the library, and the softest |
| 306 | 0.36 | `K04-kitchen-zone-island-pushin` | |
| 307 | 0.14 | `BA1-mirror-basin-pushin` | |
| 331 | 0.34 | `0003-master-pullback` | hazy, washed out on the contact sheet |
| 379 | 0.96 | `K08-kitchen-window-bench-full-pushin` | ⛔ also window-defect |
| 429 | 0.78 | `N08-kin-bed2-curtain-light-panright` | intentionally blown-out backlight |
| 464 | 1.91 | `vineyard-drone-flyover-5s` | |
| 477 | 0.82 | `BA4-marble-niche-pushin` | |

This confirms the pipeline lesson exactly: **optimising for low noise gets you a soft image.** K03/K04/BA1 are the cleanest assets in the library *and* the least visually interesting. Use them as 1-second connective tissue, never as an opening or closing frame.

---

## 5. Aspect-ratio verdict on the exteriors (F1)

| clip | source | centre-crop to 9:16 | **letterbox at 1440 wide** |
|---|---|---|---|
| `vineyard-EXACT-scene-pan-10s` | 2560×1440 | 810×1440 → 1.78× up | **0.56× down — sharpens** ✅ |
| `vineyard-drone-flyover-10s` | 1920×1080 | 607×1080 → 2.37× up ⛔ | 0.75× down ✅ |
| `hf_…155026` | 1912×1080 | 607×1080 → 2.37× up ⛔ | 0.75× down ✅ |
| EXT1 / EXT1b / EXT1d / EXT1e (Byron) | 1280×720 | 405×720 → **3.56× up** ⛔ | 1.125× up — negligible ✅ |
| EXT2 (Toowoomba) / EXT3 (Hervey Bay) | 1280×720 | 405×720 → **3.56× up** ⛔ | 1.125× up ✅ |
| `NOMI-REAL-vineyard-barossa-goldenhour` | 1280×720 | as above ⛔ | ✅ |

**Cropping destroys every exterior. Letterboxing rescues all of them at or near native resolution.**

A 1440×810 band inside a 1440×2560 frame, on ink `#2A241C` with a fine gold hairline, is a deliberate editorial device — it reads as a film still rather than a compromise, and it buys space above and below for type. Adopted as the house treatment for all exterior footage.

**This also matters commercially:** EXT2 Toowoomba and EXT3 Hervey Bay are the only assets that speak to the regional-Queensland buyer, and both are 720p. Letterboxing is what keeps that audience addressable at all.

---

## 6. ⛔ The talent footage is off-model

`NOMI /Talent Images /` holds 6 clips of a woman in cream loungewear. Reviewed frame by frame, the interior around her is **not the coastal NOMI set**: navy bedding, ginger-jar lamps, a gold-framed glass coffee table, hydrangeas, panelled walls. It is a Hamptons-styled room.

Intercutting it with the coastal library would break continuity in bedding, lamp, table and wall treatment simultaneously. **Excluded from all ten reels.**

**Consequence, stated plainly: there are no people in this deck.** No human scale, no lifestyle proof, no faces — which for an interiors campaign is a genuine weakness, not a stylistic choice. It is recorded as the top Phase-2 gap (§8), not papered over.

---

## 7. Motion census — and why it constrains cut order

| band | motion | clips |
|---|---|---|
| very slow | 0–4 | 3 |
| slow pan | 4–6 | 11 |
| push-in | 6–9 | 19 |
| active | 9–14 | 20 |
| fast | 14+ | **29** |

The library skews fast, so energy is not the constraint — **pace whiplash is**. Bathroom sits at 6.6 and exteriors at 18.9; placing them adjacent is a 3× velocity jump inside one cut.

Rules adopted for every reel:
1. **Never two shots of the same move family back to back** (families: PAN / IN / OUT / OTHER) — carried over from the tour-reel lessons.
2. **One clip per subject.** The library holds push-in *and* pan variants of the same subject; using both reads as a stutter.
3. **Order motion monotonically within a phrase** — do not jump from 6 to 19 and back.
4. **Never adjacent:** `N09` ≈ `B01` (cane bedside + white ceramic lamp) and `N03` ≈ `N12` (same vanity and oval mirror) are near-duplicate subjects.

---

## 8. Known defects and hard exclusions

| Asset | Defect | Ruling |
|---|---|---|
| **K08, K12, E06** | Kitchen window rendered as a fixed pane with a thin frame and a tile column to its left. All three are wrong; K04 is the only correct window reference. | ⛔ **Excluded from all ten reels.** |
| **30s + 60s TOUR reels** | Contain E06/K08/K12; measured detail below library median. | ⛔ Rebuild, do not reuse. |
| **K01–K12 kitchen elements reel** | Inherits the same clips. | ⛔ Excluded. |
| **E01-E09 living elements reel** | Per the pipeline lessons E06 sits in this reel. | ⚠️ Usable only if the E06 segment is trimmed out; otherwise excluded. |
| **N06** | Filename says `fanslow` but the fan does not rotate (measured 25.6°/s against a 120°/s brief). | ⚠️ Treat as a static detail. Never cut it where motion is expected. |
| **`0001-living-pan`** | Unfixed sofa-wall panel-pitch defect; warmest clip in the library (R/B 1.155). | ⚠️ Short cutaways only, graded hard. |
| **`0003-master-pullback`** | Hazy and soft (detail 331). | ⚠️ Connective tissue only. |
| **Talent Images** | Off-model interior (§6). | ⛔ Excluded. |
| **`NOMI /Undefined /` H-series** | Superseded styling. | ⛔ Excluded. |

---

## 9. Product and claim truth (from `brain_knowledge`)

### Verified and safe to state
- NOMI: **40 m² internal, 21 m² covered outdoor**
- NOMI KIN XL: **~65 m² internal**
- Lead time: **12–16 weeks**
- Supply from **Mansfield, Brisbane**; delivery and set quoted separately
- KIN is the model with the **second bedroom** — base NOMI has one
- Kitchen: light natural oak, **completely handleless**; matte white overheads; white gloss narrow vertical finger-tile splashback; pale white stone benchtop with a thick mitred edge; **black glass induction** cooktop; slimline brushed stainless undermount rangehood; brushed **brass** gooseneck pull-down; **black** undermount sink; pale timber plank floor; floor-to-ceiling slatted timber pantry face
- Bathroom: large-format white marble, brushed brass tapware and frames, dusty-blue and white towels
- Market: Queensland — Toowoomba, Hervey Bay, Gold Coast, Cairns, Brisbane, Sunshine Coast, Rockhampton

### ⛔ Must not appear in any reel
| Claim | Why |
|---|---|
| **Any price** | Genuinely contested in the brain: NOMI $120k–132k / $129k / $132k; KIN $154k / $159k / $179k. Two live memories flag the KIN figure as still open. A wrong number on a public post is a real commercial risk. **Every CTA is DM or link-in-bio.** |
| **Wind rating / cyclonic region** | No customer-safe row states one. A fabricated version was caught by the rails. |
| **BAL / bushfire rating** | Entirely fabricated when a model attempted it. |
| **Building classification** (e.g. Class 1a) | Asserted, not sourced. |
| **"Council approved" / any suggestion Ruach lodges the DA** | Ruach does **not** manage council approval. The old Meta headline *"Fully Installed & Council Approved!"* violates this twice — it also contradicts supply-only. Do not echo it. |
| **"Fully installed"** | Contradicts supply-only from Mansfield. |
| The word **"shed"** | Explicit persona rule for the caregiving audience. |
| The word **"built"** | Hard-blocked in the brand guardian. Use *made, engineered, manufactured, assembled, finished*. |
| **Emoji, em-dashes** | Brand form rules, observed clean across 61 measured replies. |
| "trust me" / "I promise" / "guarantee" | Persona rule for the builder-burned audience. |

### Brand tokens (extracted from ruach.io, not invented)
Cormorant Garamond display and body · Cinzel letterspaced gold eyebrows · ink `#2A241C` · paper `#F7F3EA` · cream `#F2E8D5` · gold `#C9A961` · deep gold `#9A7B3E` · corner ticks · 2px radii

### Voice
`Danielle - Ruach Professional` — ElevenLabs professional clone, id `sEP3AdMfjaiGYc3FBMrC`. The canonical Ruach voice, already in the workspace. Used for every narrated reel so the deck matches the phone channel.

---

## 10. Gaps this deck cannot close

1. **No people.** No human scale or lifestyle proof anywhere (§6). Biggest single weakness.
2. **No native 9:16 exteriors.** Letterboxing is a good solve, not a fix (§5).
3. **No craft, factory or delivery footage.** The strongest trust story for a builder-burned buyer — steel frame, factory tolerances, a module on a truck — has zero footage.
4. **The kitchen is under-served.** One clean hero clip (K01) and two soft ones; the best-looking kitchen views are defective (§F3).
5. **No floor-plan motion graphics.** Four plans exist as PNGs; the 40 m² / 21 m² / 65 m² story is currently words, not picture.
6. **No exterior at any time of day but golden hour.** Every exterior is dusk or dawn. No daylight, no overcast, no evening-interior-glow-from-outside.
7. **Pricing is unresolved in the brain.** Blocks every price-led creative angle until Brad rules.

---

## 11. What the audit implies for the ten cuts

- **Lead with cane and woven texture.** It is measurably the densest material in the product and the top five detail clips are all cane.
- **Living is the hero room; bathroom is the luxury signal; kin-bed2 is the emotional room; kitchen is the spec room.** Cast each reel to a room family's actual strength rather than tour every room in every cut.
- **Exteriors go in a letterbox band, always.** That becomes a recognisable brand device across the deck.
- **Grade every cut to R/B 1.099 / G/B 1.026, attenuate-only.**
- **Cut the bathroom slow and the living room fast.** Their motion medians differ by 1.7× and their d/n by 1.4×.
- **No price, no ratings, no approval claims, no people, no emoji.** Every CTA is a DM or the link in bio.

# GAP HUNT — NOMI Reel Deck, adversarial review
**Date:** 2026-07-30 · **Target:** `briefs/reels.json` (md5 `df87c10548633e704e7b3783023c92a4`, mtime 09:08) + `audit/AUDIT.md`
**Baseline:** `masters-4k/HIGGSFIELD-PIPELINE-LESSONS.md` · 54 clips in `coastal-video-9x16 /Approved/` · 11 exterior sources · 156 shot instances checked

> ⚠️ `reels.json` was **edited live during this audit** (R09 `shot_frames` changed 96 → 80 and its `pace` prose was rewritten between 09:02 and 09:08). Every finding below is against the 09:08 hash above. Re-check R09 before acting.

---

## BLOCKERS

| Sev | Class | Reel+Shot | Finding | Evidence | Fix |
|---|---|---|---|---|---|
| **BLOCKER** | D | R09 (all) | **Frame arithmetic is broken.** `shot_frames` 80 × 10 shots = **800 frames = 33.33s**, against a declared `frames: 960` / `duration_s: 40`. The `pace` field claims "3.33s per shot, **twelve shots**" but the `shots` array holds **ten**; `proof` and `accepted_repeats` still say "ten shots" / "a 40s spec reel". Someone changed the tempo to fix a source overrun and never added the two shots. | `python3` over the on-disk JSON: `R09 declared=960 (40s) computed=800 (33.33s)  <-- MISMATCH`. All nine other reels compute exactly (R01 1440, R02 1080, R03 720, R04 720, R05 960, R06 840, R07 720, R08 720, R10 1440). | Add 2 shots (960 = 80×12). K12-slider-panleft and K08-slider-panright are available and are the 2nd/3rd best kitchen clips — see the K08/K12/E06 row below. Then align `pace`, `proof`, `accepted_repeats`. |
| **BLOCKER** | D + A | R01 s1/s24 · R02 s1 · R03 s1/s12 · R06 (whole) · R07 (all 6) · R09 s3/s8 | **The prescribed luma lifts clip 8–19 % of the frame to pure white, under a deck-level `"mode": "attenuate-only"` mandate that forbids them.** Four reels prescribe multiplicative lifts (R01/R03/R07 "exteriors get a luma lift toward 150"; R02 "needs both an attenuation and a lift"; R06 "grade on a RAMP… open at 150 finish at 195"; R09 "K01 is 45 points dark and will read as a different room unless lifted"). Every one of those clips already has highlights at 248–253. The lessons doc's whole reason for attenuate-only is that it *mathematically cannot clip*. | Measured gray p99.5 / % pixels ≥255 after the required linear gain (360px frame, `ss=2.0`):<br>`K01-pushin  mean 128.2  p99.5 250.0  → ×1.334 → 19.34 % clipped`<br>`EXT2-toowoomba mean 118.7 p99.5 248.0 → ×1.290 → 17.84 % clipped`<br>`0002-tvwall-pan mean 187.6 p99.5 252.0 → ×1.071 → 7.98 % clipped` | Do the luma moves as **gamma**, not gain — for K01, γ = ln(173/255)/ln(128/255) ≈ **0.56** raises the mean to 173 with the white point untouched (0 % clipped). Same for exteriors (γ ≈ 0.79 for 118→150). Keep the channel-ratio work attenuate-only as specified; state explicitly that the luma stage is gamma. |
| **BLOCKER** | A + C | R07 (whole reel) | **`proof: "The same home on five completely different pieces of land"` is false — there are four sites over six shots, and two of the five place labels name the same land.** s2 (`EXT1e-byron-deck-panmove`, labelled "THE DECK") and s4 (`EXT1b-byron-carport-glide`, labelled "BYRON HINTERLAND") are the *same* Byron plate — identical building, identical hilltop, identical headland/ocean backdrop, differing only in camera position. s5 (`vineyard-drone-flyover-10s`, "WINE COUNTRY") and s6 (`NOMI-REAL-vineyard-barossa-goldenhour`) are the same vineyard and are **adjacent**, which also breaks the reel's own alternating AERIAL/GROUND `rule_deviation` on the "the land carries the variety" argument. | Frame grabs at t=3.0s from all 6: `EXT1`, `EXT1b`, `EXT1d`, `EXT1e` are four camera positions on one plate (same jacaranda-free hilltop, same twin-headland horizon). `vineyard-EXACT`, `vineyard-drone`, `barossa-goldenhour` are three views of one vineyard. Distinct sites = Byron, Toowoomba, Hervey Bay, vineyard = **4**. | Cut to 4 shots (one per site) at 180f = 30s, or relabel honestly ("BYRON — from the air / from the deck"). Do not put two vineyard shots adjacent. Also see the QLD-market row. |
| **BLOCKER** | C | R02 s18–s22 | **KIN-only second-bedroom features are presented as base-NOMI inclusions.** R02 makes no model distinction anywhere, labels s15 "Main bedroom", then labels s18 "Walk-in robe, lit", s19 "Lit display niche", s20 "Window seat with storage", s21 "Arched cane robe doors", s22 "Backlit oval mirror" — all five shots are `room: kin-bed2`, i.e. the second bedroom **which only NOMI KIN has**. A viewer reads them as standard NOMI inclusions. | AUDIT §9 verified fact: *"KIN is the model with the second bedroom — base NOMI has one."* Lessons plate map: *"bed1-main is only 0003, 0004 and B01"*; 0009/0010/N-series = KIN bed2. The five clips at s18–22 are `0009`, `N03`, `N01`, `0010`, `N12` — every one kin-bed2. | Either add an on-screen qualifier from s18 ("KIN second bedroom") or drop s18–22 and replace with bathroom/living inclusions. This is the same mis-selling risk the deck's own `banned_copy` discipline exists to prevent. |
| **BLOCKER** | E | R03, R05, R09, R10 (all VO reels) | **No captions or subtitles anywhere in the deck.** Four reels are VO-carried and R03 states it outright — `type_strategy: "Almost none. One opening line, one closing line. The VO carries it."` Instagram plays reels muted by default. R03 also has 11 of 12 shots with no on-screen text at all, and R10 has 14 of 18. Muted, these reels carry almost no information. | `grep -ci` over `reels.json`: `subtitle` 0, `srt` 0, `caption` 1 (and that one is inside the `safe_area` note about IG chrome, not about subtitles). Shots with no text: `{R01:18, R03:11, R04:20, R05:12, R06:9, R08:6, R10:14}`. | Add a burned-in caption track spec for every `vo: true` reel (Cormorant/Cinzel per brand, inside the corrected safe area) plus an `.srt` sidecar. Non-negotiable for a VO-led reel. |

---

## MAJOR

| Sev | Class | Reel+Shot | Finding | Evidence | Fix |
|---|---|---|---|---|---|
| **MAJOR** | A + B | deck `excluded_clips`; R09 `constraint` | **The blanket K08/K12/E06 exclusion is stale — the three clips now sitting in `Approved/` are the CORRECTED assets, not the defective ones.** `K08-…-slider-panright`, `K12-…-slider-panleft` and `E06-…-slider-pullback` all show the fixed window: chunky white stepped frame, visible vertical meeting stile, and **no finger-tile column to the left** (frame runs out to the return). The defective versions were already retired — `_superseded/` holds `K08/K12/E06 …-OLD-WRONG-WINDOW-fixedpane-tileleft.mp4`. Only `K08-…-window-bench-**full**-pushin` still carries the defect. Consequence: AUDIT F3 ("its two best-looking clips K08, K12 carry the known window defect") is wrong, R09's `constraint` ("removes the three best-looking kitchen views… carried by K01") is wrong, and R09 repeats K01/K03/K10 for no reason. | Frame grabs at t=2.5s. `K08-full-pushin`: tile column left of window + slim frame + single pane = **defect confirmed**. `K08-slider-panright`, `K12-slider-panleft`, `E06-slider-pullback`: chunky gloss-white stepped frame, meeting stile visible, no tile left = **fixed**. Measured detail (`clip-analysis.tsv`): K12-slider **984.7**, K08-slider **681.2**, vs K10 581/517, K04 306, K03 285/240 — the two excluded sliders are the kitchen's 2nd and 3rd best clips. | Narrow `excluded_clips` to `K08-kitchen-window-bench-full-pushin` only. Re-admit K12-slider and K08-slider (note K08-slider's flat-noise is 1.842 — grainy, so cut it short) and E06-slider. This also supplies R09's two missing shots. |
| **MAJOR** | A + C | R05 s7 vs s10 · R01 s13 vs s15 · R06 s3–s5 | **B02 (bed1-main) and N13 (kin-bed2) are the same picture with a different throw — and they are placed 2–3 shots apart in the reels whose whole job is to distinguish the two bedrooms.** Verified: identical cream upholstered headboard, identical slate-blue waffle euro pillows, identical ribbed cream lumbar, identical jute rug, identical bed make. The *only* difference is the throw — B02 dark slate, N13 light blue. In R05 (the "One Bedroom or Two" hinge reel) B02 is s7 in the NOMI half and N13 is s10 in the KIN half, with one title card between. In R01, B02 is s13 and N13 is s15 labelled "Second bedroom — KIN" — the viewer sees the same bed macro relabelled as a different room. **R06 is worse:** it sandwiches B02 (bed1-main) between N11 (s3) and 0006 (s5), both kin-bed2, with no label — intercutting two bedrooms as one room, the exact failure the 2026-07-28 plate-map correction exists to prevent. | Frame grabs t=1.0s. B02: dark-slate fringed knit throw + fluted-ball lamp + framed art edge (matches 0003 exactly). N13: light-blue waffle throw (matches N17, the approved kin-bed2 anchor). Everything else pixel-consistent between them. | Replace B02 in R05/R01 with a shot that is unmistakably bed1-main — `0004-master-pushin` (frontal, two cane bedsides, VJ wall + framed art) or `0003-master-D3move`. Replace N13 in R05's KIN half with `N17-room-wide` or `N01-window-seat` (features the base NOMI does not have). In R06, either drop B02 or move it out of the kin-bed2 block. |
| **MAJOR** | A | R02 s2/s5, s9/s13, s11/s14 | **Three undeclared same-subject repeats in the deck's fastest reel.** AUDIT §7 rule 2 is "One clip per subject — using both push-in and pan variants reads as a stutter". R02 breaks it three times at a 1.5s cadence and, unlike R08 and R09, carries no `accepted_repeats` field. Worst case is **E02 at s9 and s13, which are visually the same frame** (a push-in and a pull-back of one plate at the same zoom) labelled "Oak coffee table" and "Sisal rug" — the viewer sees one picture twice with two different captions. Also K03 at s2/s5 (gap 3) and 0002 at s11/s14 (gap 3). R02's own `proof` claims "Thirty named inclusions" but delivers 27 distinct pictures. | Subject-prefix scan: `R02 K03 shots [2,5] gap 3 · R02 E02 shots [9,13] gap 4 · R02 0002 shots [11,14] gap 3`. Frame grabs of `E02-element-pushin-dyn` and `E02-coffee-table-pullback` at t=2.5s are near-identical framings of the same glass-topped oak table on the same jute rug. | Drop one E02 and one K03. Unused clips are available: `BA2-towel-shelf-pushin`, `BA1-mirror-basin-pullback`, `N05-display-niche-panleft`, `N10-ottoman-rug-panleft`, `0007`/`0008-bedroom-pan`, plus the re-admitted K12/K08 sliders. |
| **MAJOR** | C | R02 s30 | **Label does not match its picture: "21m² covered outdoor" sits on `EXT1b-byron-carport-glide`, which is the CARPORT — a car is parked under the roof extension.** The covered outdoor *living* space is `EXT1d-byron-deck-glide`, which shows the timber deck under the same roof with no car. | Frame grab `EXT1b` at t=5.0s: silver car under the pergola, gravel driveway, no deck visible. `EXT1d` at t=5.0s: timber deck, no car. | Swap s30 to `EXT1d-byron-deck-glide-9s.mp4`. |
| **MAJOR** | C | R02 s29 | **Label does not match its picture: "Full-size walk-in shower" sits on `BA9-toilet-shower-lowangle-pullback`, where the dominant subject is the wall-hung toilet and the shower is a small hinged-glass enclosure behind it.** "Full-size walk-in" is also a size claim with no measurement anywhere in the audit's verified facts. | Frame grab `BA9-…-pushin` at t=2.0s: wall-hung pan foreground-left, brass-framed glass shower *door* (hinged, not open walk-in) behind. AUDIT §9 verifies only "large-format white marble, brushed brass tapware and frames, dusty-blue and white towels" — no shower dimension. | Either relabel to something the frame proves ("Brass-framed glass screen") or move the label to `BA7-brass-frame-panleft`, which does show the screen full-height. Drop "full-size". |
| **MAJOR** | C | R06 s1 | **"5.40 AM" sits on a dusk frame.** `vineyard-EXACT-scene-pan-10s` is a warm low-sun shot with **festoon party lights lit** across the pergola and the interior glowing — it reads unambiguously as evening. The `note` even calls it "darkest asset in library, luma 84 — correct as the opening frame", which is a luma argument, not a time-of-day one. Opening a reel called "Morning in a NOMI" on a lit-up evening scene stamped 5.40 AM is a visible contradiction. | Frame grab at t=3.0s: sunset-coloured sky, mist in the vine rows, strung festoon bulbs illuminated, warm interior spill. AUDIT §10.6 already records "No exterior at any time of day but golden hour… no daylight". | Use `EXT2-toowoomba-flyover` (dawn mist, no artificial light) as the 5.40 AM opener, and move `vineyard-EXACT` out of any reel that stamps a morning time. |
| **MAJOR** | D | R06 s1 · R10 s1 | **`vineyard-EXACT-scene-pan-10s.mp4` is 30 fps — the only non-24fps source in the deck — and no frame-rate conversion is specified.** Delivery is `"fps": 24`. Cutting a 30fps source with `-ss`/`-t` into a 24fps concat produces either dropped-frame judder on a slow pan (the worst case for it) or timestamp drift through the concat demuxer. | `ffprobe`: `vineyard-EXACT-scene-pan-10s.mp4 2560,1440,30/1,300 dur=10.000`. Every other used source is `24/1`. | Pre-conform it once: `ffmpeg -i … -vf "fps=24,scale=1440:810" -crf 14` to a 24fps intermediate, then cut from that. Do not rely on ffmpeg's implicit fps filter inside the segment graph. |
| **MAJOR** | D | all reels | **Mixed audio streams across sources, with no `-an` in the brief.** Only 2 of the 30 referenced sources carry an AAC track — `RUA-NOM-AR-VID-0002-coastal-9x16-living-tvwall-pan-4K.mp4` (used in R01 s3, R02 s14, R05 s4, R06 s13, R10 s3) and `NOMI-REAL-vineyard-barossa-goldenhour.mp4` (R01 s24, R05 s16, R07 s6, R10 s18). The other 28 are video-only. Concatenating mixed-stream segments either hard-fails or silently drops/desyncs audio, and any surviving source audio will fight the score/VO. | `ffprobe -select_streams a` across `Approved/` (54 files) returns a codec for exactly one file; across the 11 exterior sources, for `barossa` and the unused `hf_…155026`. | Add `-an` to every segment cut and build the audio bed as a separate pass. State it in the brief so it is not rediscovered at build time. |
| **MAJOR** | D | deck `safe_area` · R07 `type_strategy` | **The declared safe area is too shallow at the bottom and has no left/right margin at all.** `bottom_px: 420` of 2560 = 16.4 %; Meta's reel guidance is ~20 % (≈512 px at this height) for caption + CTA + audio attribution. There is no allowance for the right-hand action rail (like/comment/share/profile ≈ 180 px at 1080-wide, ≈ 240 px at 1440). R07 puts "one place name per shot… in the ink field below the band" — with an 810px band centred in 2560 that field is y 1685–2560, of which the bottom 512 px is IG chrome, leaving a 363 px window the brief does not acknowledge. | `safe_area` in JSON: `{top_px: 220, bottom_px: 420}`, no left/right keys. Band geometry: (2560−810)/2 = 875 px ink field top and bottom; band occupies y 875–1685. | Set `bottom_px: 512`, add `left_px: 90` / `right_px: 260`. Then re-check R07's place names and every `label` in R02/R09 (30 and 10 lower-third labels respectively). |
| **MAJOR** | C | R08 (title) · R01 s20 · R05 s15 | **"Ensuite" is asserted but unverified, and "Same ensuite, both models" is not a verified fact at all.** AUDIT §9 verifies only *"Bathroom: large-format white marble, brushed brass tapware and frames, dusty-blue and white towels"* — nothing establishes that the room opens off a bedroom. The plate set cuts against it: `0011-hallway-cabinet-closed-**ensuite**` implies hallway access, and `0016-accessible-**bathroom**-grabrails` suggests one room serving the whole home. R05 s15's "Same ensuite, both models" additionally asserts parity between a 40 m² NOMI and a ~65 m² KIN XL with no source. | `coastal-images-9x16/` masters: `0011-…-hallway-cabinet-closed-ensuite-4K.png`, `0013/0014/0015-ensuite-…`, `0016-accessible-bathroom-grabrails-4K.png`. AUDIT §2 also splits "ensuite (n=2)" and "bathroom (n=11)" into separate families, so the BA-series the reels use is the *bathroom* family, not the ensuite family. | Ask Brad: is it an ensuite, a main bathroom, or both? Until answered, retitle R08 "The Bathroom" and drop "Same ensuite, both models". This is a spec claim on a public post — same class of risk as the price. |
| **MAJOR** | C | R05 (whole reel) | **"There are two. That is the whole decision." contradicts the range.** The brain records **KIN standard and KIN XL as separate models** (memory 2026-07-29: KIN standard $154,000 / XL $179,000, **8 SKUs not 6**), and AUDIT §9 itself lists "NOMI KIN XL: ~65 m² internal" as a *verified* fact. R05 asserts a two-model range, never mentions XL, and the verified 65 m² figure appears in **no reel in the deck**. | AUDIT §9 verified list includes the 65 m² XL. `grep -c '65' reels.json` on copy fields: the figure appears nowhere; R05's title card reads `40m² internal · 21m² covered` only. AUDIT §10.5 also flags the 40/21/65 story as "words, not picture". | Either rewrite R05 as "one bedroom or two" without claiming the range has two members, or add the XL. Do not publish a reel that mis-states how many models exist. |
| **MAJOR** | C | R10 VO · R06 `cta` | **Delivery-promise language.** R10's VO closes "Every decision resolved. Every detail held. **Every date kept.**" — that is a delivery guarantee in substance, and `banned_copy` prohibits "guarantee". R06's CTA "**Yours from twelve weeks**" states the optimistic end of a 12–16 week range as an offer. Both are exposed for a supply-only manufacturer whose delivery and set are quoted separately. | `banned_copy: [… "guarantee" …]`. AUDIT §9: lead time **12–16 weeks**; *"Supply from Mansfield, Brisbane; delivery and set quoted separately"*. | R10: "Every detail already decided." R06: "Twelve to sixteen weeks. Link in bio." |
| **MAJOR** | A | R01 s15 · R02 `cta` | **On-screen type violates the deck's own `banned_copy` em-dash rule.** R01 s15's *label* — the type that will be burned into the frame — reads `"Second bedroom — KIN"`. R02's `cta` reads `"Full inclusions list — link in bio."` | Regex scan of every `text`/`titlecard`/`cta`/`vo_script` value for `—`/`–`: two hits in deliverable copy (R01 s15 label, R02 cta). The other two hits are metadata annotations (`R03.copy_rules`, `R10.campaign_line`) and are fine. | `"Second bedroom · KIN"` (the deck already uses the middle dot in R05's title card) and `"Full inclusions list. Link in bio."` |
| **MAJOR** | C | R03 `vo_script` | **The VO does not fit the reel at the pace the brief demands.** 66 words in 30s requires **132 wpm**, against a brief that says "slow and level — 2.5s per shot, no acceleration anywhere", "unhurried", "Dignity is the whole brief". A natural ElevenLabs read of caregiving copy at 115–120 wpm puts this at **33–34s** — 3–4s over. | Word counts vs duration: `R03 66 words / 30s → 29.3s @135wpm, 26.4s @150wpm` — i.e. it only fits by reading *fast*. R05 (67 w / 40s), R09 (69 w / 40s) and R10 (63 w / 60s) all have comfortable headroom; R03 is the only one that does not. | Cut ~10 words from R03. Candidate: drop "It is made in our Mansfield factory and finished before it ever reaches your block." (R05 and R10 already carry the Mansfield line). |

---

## MINOR

| Sev | Class | Reel+Shot | Finding | Evidence | Fix |
|---|---|---|---|---|---|
| MINOR | A | R07 `treatment_note` | R07 abandons the house exterior treatment. AUDIT §5 adopts the ink band with a fine gold hairline as *"the house treatment for all exterior footage"* and §11 calls it *"a recognisable brand device across the deck"* — and the one exterior-led reel is the one that switches to `letterbox-blur`. The deviation is declared with a reason, but it means 6 of the deck's 14 exterior shots use a different device from the other 8. | AUDIT §5 vs `R07.treatment_note`. Exterior shot census: ink band in R01 s1/s24, R02 s1/s30, R03 s1/s12, R05 s1/s16, R06 s1/s14, R10 s1/s18 (12 shots); blur in R07 (6 shots). | Pick one. If the blur is right for exterior-led, use it for the bookends too; otherwise hold the band in R07 and let the type sit in the ink. |
| MINOR | C | R03 | Internal contradiction: `copy_rules` says "No delivery date" and the `vo_script` in the same reel says "Twelve to sixteen weeks". | Both fields, same object. | Reword `copy_rules` to "No fixed delivery date" or drop the lead time from R03's VO. |
| MINOR | C | R05 s3 · R01 s2 · R10 s2 | `0002R-living-sofa-corner-pushin` is not a living-room establishing shot — it is a framed-art-and-bar-stools vignette. R01 s2 puts the room label "Living" on it and R05 s3 opens the NOMI half's whole case on it. The deck's best open-plan wide, `0002-living-tvon-pushin` (TV wall + slat pantry + kitchen + breakfast bar + sofa in one frame, detail 1520), is used **once**, in R02 s11. | Frame grabs t=2.0s of both. | Open R05's NOMI half on `0002-living-tvon-pushin` and demote 0002R to a cutaway. |
| MINOR | C | AUDIT §6 | One of the four cited reasons for excluding the talent footage does not hold: the audit lists *"a gold-framed glass coffee table"* as evidence the room is Hamptons rather than coastal, but the approved coastal living plate 0002 contains exactly that — a glass-topped table on a light metal/oak frame. The other three mismatches (navy bedding, ginger-jar lamps, panelled walls + hydrangeas) do hold, so the exclusion still stands; the reasoning is just one item weaker than stated. | Frame grab `0002-living-tvwall-pan` t=4.5s and `E02-coffee-table-pullback` t=2.5s: glass top, slim pale-oak frame. | Drop that item from the §6 rationale. No change to the exclusion. |
| MINOR | E | R07 | Two of the five place names are outside the stated market. AUDIT §9: *"Market: Queensland — Toowoomba, Hervey Bay, Gold Coast, Cairns, Brisbane, Sunshine Coast, Rockhampton."* R07's persona is "Regional QLD" but s2/s4 are Byron (NSW) and s5/s6 are the Barossa (SA — it is in the filename). Only 2 of 6 shots are on QLD land. | `R07.market`, shot labels, `NOMI-REAL-vineyard-barossa-goldenhour.mp4`. | Either broaden the stated market or drop "WINE COUNTRY" and lead with Toowoomba + Hervey Bay. The audit's own §5 note is that EXT2/EXT3 are the only assets addressing this buyer — spending 4 of 6 shots elsewhere wastes the reel. |
| MINOR | E | deck | No cover/thumbnail frame nominated for any reel; no caption copy, hashtags, location tag or alt text; no music track named or licence cleared (`"ambient-calm"`, `"cinematic-swell"` etc. are placeholders); no tracked destination behind "link in bio"; no end-card design or hold-duration spec despite 10 `endcard: true` flags; no 1080×1920 companion for other channels; no Brad approval gate before publishing. | `grep -ci` on `reels.json`: `thumbnail` 0, `hashtag` 0, `license` 0, `music track` 0. `endcard` appears 10× with no accompanying spec. | Add a delivery-manifest section to `reels.json`. The approval gate is a project hard rule for anything public-facing. |
| MINOR | E | deck | AUDIT §10.5 records "No floor-plan motion graphics… the 40 m² / 21 m² / 65 m² story is currently words, not picture" and lists four plan PNGs as available. No reel uses one — including R05, whose entire premise is the plan difference, and which carries the m² figures as a static title card. | AUDIT §10.5 vs R05 s2/s9 title cards. | A 3s animated plan at R05's hinge is the single highest-value addition in the deck and needs no new footage. |

---

## Verified — no gap found (checked, so it is not re-litigated)

| Class | What was checked | Result |
|---|---|---|
| A | **Plate-map room assignments, every shot.** The contested cases were resolved visually, not assumed. | **All correct.** `B02` is genuinely **bed1-main** — dark-slate fringed throw and fluted-ball lamp match `0003`/`0004` exactly, not the light-blue waffle throw and urn-with-brass-collar lamp of the N-series. `N11-bed-hero`, despite showing framed coastal art on a VJ wall (the bed1-main signature), is genuinely **kin-bed2** — its light-blue waffle throw and urn lamp match `N17`, the approved kin-bed2 anchor, which also has framed art above its bed. `0006` → kin-bed2 ✓ (cane arched robes + oval backlit mirror + lit niche). `0009`/`0010` → kin-bed2 ✓. `0003`/`0004`/`B01` → bed1-main ✓. **No reel intercuts two bedrooms as one room by mis-assignment** — the R06/R05/R01 problem above is a *casting* problem (near-identical pictures), not a mis-labelling one. |
| B | **Every used clip traced to its source plate.** | **Clean.** All 30 element IDs used have a live approved 4K master in `coastal-images-9x16/` root. The `…-2k.png` files sitting in `_rejects/` for K01, K03, K05–K07, K09–K11, N02, N06, E08 are **housekept working files, not rejections** (no failure reason in the filename, and each has a promoted 4K sibling in root) — the genuine rejects carry the reason inline (`…-WRONG-KITCHEN`, `…-REJECTED-not-per-room-design`, `…-INVENTED-…`). Frame-0 of the four clips whose plates *were* superseded (`N04` white-vj-wall, `N03` shelving-wrong-side, `0009` v3-queenbed, `0010` v2-fan) matches the **current** root master in every case — so none carries retired geometry. |
| D | **Source-duration vs `ss` + shot length, all 156 shot instances.** This was the most likely silent failure and it is currently clean — but only just. | **No overruns.** Tightest is **R09 s7** (`K03-panright`, `ss=1.6`, 80f needed, 5.0417s source) with **2.60 frames = 0.108s** of headroom. R08's eight shots each hold 7.00 frames. Note R09's own `pace` field records that at the previous 96-frame tempo `ss=1.0` left one frame of headroom — that was the reason for the live edit, and it is what broke the frame count (BLOCKER 1). Any future increase to `shot_frames` or `ss` in R09 goes negative immediately. |
| D | **All referenced source files exist.** | **30/30 present** across `Approved/` (54 files), `NOMI-4k Asset Library/external-flyovers/` and `NOMI /films /`. No missing filenames, no typos. |
| D | **Letterbox arithmetic.** | **Exact, no fractional pixels.** 1440 × 9/16 = **810** integer. Every exterior source lands on it cleanly: 1280×720 → 1440×810 (1.125× up), 1920×1080 → 1440×810 (0.75× down), 2560×1440 → 1440×810 (0.5625× down). Ink field above/below = (2560−810)/2 = **875 px** each, integer. The choice of 1440 wide over 1080 is correct — 1080 × 9/16 = 607.5, a non-integer band. |
| D | **Frame arithmetic, 9 of 10 reels.** | R01 1440 ✓ · R02 1080 ✓ · R03 720 ✓ · R04 720 ✓ · R05 960 ✓ · R06 840 ✓ · R07 720 ✓ · R08 720 ✓ · R10 1440 ✓. Only R09 fails. |
| A | **Never-adjacent near-duplicate rule** (AUDIT §7.4: `N09 ≈ B01`, `N03 ≈ N12`). | **Not violated in any reel.** N09 and B01 never co-occur. N03/N12 co-occur in R02 (s19/s22) and R03 (s7/s10), always 3 shots apart. Their near-identity is confirmed visually (same lit niche, same oval mirror, same cane-front floating vanity — N03 just wider), so the 3-shot gap is the minimum acceptable and should not be reduced. |
| A | **Move-family adjacency rule** (never two of PAN/IN/OUT/OTHER back to back). | **Not violated in any of the ten reels.** Checked shot-by-shot. |
| C | **On-screen label vs actual frame, all 40 labels.** | **34 verified correct.** Notably: "Round ceramic basin" (BA1 — round vessel basin ✓), "Brass flush plate" (BA13 — plate clearly visible ✓), "Brass-framed shower" (BA7 ✓), "Lit vanity shelf" (BA2 — folded blue + rolled white towels on a lit shelf ✓), "Window seat with storage" (N01 — cupboard doors under the seat ✓), "Boucle sofa" (E09 — genuine boucle texture at macro ✓), "Brass pendant" (E07 ✓), "Breakfast bar" (E01 ✓), "Timber slat feature wall" (E05 ✓), "Handleless oak joinery" / "No pulls. No knobs. No brass." (K10 ✓), "Open plan, 40m² internal" (0002-tvon-pushin — the deck's only true open-plan wide ✓), "Arched cane robe doors" (0010 ✓), "Backlit oval mirror" (N12 ✓), "Lit display niche" (N03 ✓), "Black glass induction, inset flush" (K03-pushin ✓). **"Soft-close drawer bank" on K03-panright is correct** — I expected this to fail, but the frame at t=3.0s does show an open drawer bank with a cutlery insert alongside the cooktop. **"VJ-panelled walls" on 0002-tvwall-pan is also correct** — the TV wall is genuine vertical-board VJ; the timber slat wall is a separate element to its right (E05). The 3 that fail are in the MAJOR table. |
| C | **Kitchen material spec vs frames.** | **Consistent.** K10's splashback reads as a tile grid at first glance but at 1:1 it is genuine narrow **vertical finger tile** foreshortened by the oblique angle — matches K03 and the written spec. Handleless slab fronts with shadow-gap reveals ✓, black glass induction ✓, slimline brushed-stainless undermount canopy above the cooktop only ✓, brass gooseneck ✓, black undermount sink ✓, pale timber plank floor ✓. `K04`'s clip is misleadingly named `zone-island` while its still master is `element-kitchen-window-bench`, but the asset itself is the correct window reference and shows the chunky frame running out to the return. |
| C | **Grade method.** | **Sound as specified for the channel-ratio stage.** `aR = rb_target·B/R`, `aG = gb_target·B/G`, `aB = 1`, normalised so max = 1.0, correctly cannot clip — a cool exterior at R/B 1.003 warms to 1.099 by attenuating blue to 0.912, not by lifting red. Per-clip `rb`/`gb` measurements already exist in `clip-analysis.tsv` for all 54 clips. Only the **luma** stage is broken (BLOCKER 2). |
| D | **Instagram constraints.** | Duration ✓ (longest 60s vs 90s cap; shortest 30s vs 3s floor). Aspect 9:16 ✓. Codec/pix_fmt `libx264`/`yuv420p` ✓. Delivering 1440×2560 is **above** IG's 1080-wide publish ceiling so it will be downscaled server-side — an extra resample generation, accepted here as the price of integer letterbox maths (see the letterbox row). Worth spot-checking one export against a 1080-wide build before committing the deck. |

---

## Could not verify

- **Whether the bathroom is an ensuite, a main bathroom, or both.** The plate names contradict each other (`0011-hallway-cabinet-closed-ensuite` vs `0016-accessible-bathroom-grabrails`) and no `brain_knowledge` row was consulted for this audit. Needs Brad. Flagged MAJOR above rather than assumed either way.
- **"Soft-close" (R02 s17 "Cane bedside, soft-close"; R09 s7 "Soft-close drawer bank").** Soft-close hardware cannot be shown in a still or a 3s clip and no verified spec row states it. Not raised as a finding because the audit does not list it as banned — but it is an unevidenced fitting claim in a reel whose `proof` is "nothing named that is not visible in the shot".
- **"Queen, wall to wall" (R02 s16).** No dimension is measurable from B02's macro framing, and no verified fact states the bed size. The superseded plate name `0009-…-v3-queenbed` implies queen was a deliberate design decision, but that is inference, not a source.
- **Actual audio bed availability/licensing.** No track files exist anywhere in the audited tree; the `music` values are descriptors. Cannot assess sync, length or rights.
- **Whether the exterior building geometry matches the KIN two-bedroom footprint.** Every exterior shows one flat-roof module with a carport/pergola. The interiors are consistent with it for the base NOMI; there is no exterior asset showing a KIN. R05 opens both halves on the same Byron exterior. Not raised as a finding because no reel makes an exterior-size claim, but a KIN exterior is a genuine asset gap.

Every source is 24fps native. Instagram accepts and preserves 23-60fps. The only routes to 30 are frame duplication, which creates the judder the finding warns about, or optical-flow interpolation, which smears fine repeating geometry — and this library is cane weave, timber battens and VJ panelling almost throughout. The remedy is worse than the disease.

Factually wrong, and verified before acting. R07's sources are 8.0 to 10.04s, not 5s. Every shot in the deck was checked for source overrun independently; four genuine overruns were found in R09 and fixed by restructuring it. R07 was never among them.

The reviewer's diagnosis is accepted — price silence has a real conversion cost and is not a neutral choice. The proposed remedy is not: the brain holds three conflicting NOMI figures and three conflicting KIN figures, and publishing a contested price is exactly the risk the policy exists to avoid. Recorded as open commercial risk with a one-edit slot ready in R05 the moment Brad rules.

Adopted once, as R07's signature shot, not globally. Stacked bands fill the frame natively and put two locations in one image, which suits R07's argument exactly. Used on every exterior it would become the device instead of serving one.

The bathroom's low motion is the room's character, not a defect. R08's whole premise is that the family measures 6.6 against the library's 10.2 and should be cut to suit. Ramping it to match the living room would make the deck more uniform, which is the opposite of the brief.

Stay at 24fps. One reviewer argued for conforming to 30fps to avoid a 3:2 pulldown judder. Rejected: every source is 24fps native, and the only routes to 30 are frame duplication (which introduces the judder it claims to avoid) or optical-flow interpolation. This library is dominated by fine repeating geometry — cane weave, timber battens, VJ panelling — which is exactly what optical flow smears. Instagram accepts 23-60fps and preserves it. Interpolating cane weave to dodge a hypothetical pulldown is the larger risk.

A reviewer showed that attenuate-only chroma plus a luma gain is self-defeating on dark cool footage: to reach R/B 1.099 from EXT2's 1.003 attenuate-only must crush blue by 9%, which desaturates while the luma lift raises noise. The gap hunt then measured the luma lift itself clipping 8-19% of pixels to pure white (K01 x1.334 clipped 19.34%, EXT2 x1.290 clipped 17.84%) because those frames already sit at p99.5 = 248-253. Gamma lifts midtones and leaves the highlight shoulder intact.

Higgsfield was authorised for logo animation and deliberately not used. Every generative pass drifts geometry — the one thing a logo cannot do — and the pipeline lessons record repeated geometry drift on far easier tasks. A 1.2s opacity ramp, 3% scale settle and gold hairline wipe is frame-exact, free, identical across all ten reels, and carries zero drift risk.

letterbox-ink was dropped: a 1440x810 band on a flat ink field leaves 68% of the frame dead, which on a phone reads as a reposted landscape video rather than a film still, and Meta's chrome then floats over empty space. Blur-fill keeps every exterior pixel at or near native resolution while filling the frame. The gold hairlines keep it a brand device rather than a generic social reframe.