# QC — `launch_overview.mp4` v3 (Metrum AI · AMD + Dell + Router launch, external, ~90 s)

> **v3 also adds a Veritasium-style hook and one source-study beat (31 Oct 2026).** Act 1 now opens on a
> counterintuitive measured fact — *"A failed attempt cost almost what a successful one cost: 48
> core-seconds against 47. So why do we price AI by the token?"* — labelled **A4 · MEASURED · CORE SIZING
> PILOT**, drawn as two bars of deliberately equal length, and pays off on the headline. Act 3 gained one
> supervisor-approved measured beat from `shared/SOURCE-STUDIES.md`: **`48 of 48 prompts routed as
> labelled`** (laptop NPU → Instinct GPUs, **S-A3**, MEASURED · one test set), shown on its own panel.
> The hook VO starts at 0.200 s so the longer line still clears the Act 2 cut; nothing else moved.
>
> **v3 (external-verification correction, 31 Oct 2026).** Act 4's receipt labelled the owned column
> `$0.0036` under a bare "total per accepted job", which is the **owned marginal** figure; the owned
> **all-in** figure (`$0.0057`) was only mentioned in the small assumption line. The receipt now shows
> **three** labelled columns — `FRONTIER LIST` / `OWNED · MARGINAL` / `OWNED · ALL-IN` — the ratio is
> stated as **≈9.9× vs marginal · ≈6.3× vs all-in**, and the narration line was re-recorded (Roger) to
> say both owned figures. Because that line is longer, Act 4's slot grew from 20.6 s to 23.5 s and the
> film from 91.0 s to **94.0 s**; the music is the same bed, re-shaped to the new length, and every
> other act's narration, visuals and timing are untouched. All 14 on-screen receipt figures are
> re-verified against `outputs.csv` by `work/verify_numbers.py`.

> **v2 (visual-energy pass, supervisor note 30 Oct 2026).** The story, claims, numbers, narration,
music and captions are unchanged from the approved v1. What changed is the picture: full-frame moving
hero plates with the three.js wireframe screen-blended over them, brighter additive geometry, depth
fog, 3–5× denser particle streams, a new corridor scene, gradient flashes and light sweeps landing on
the music's beats (138 bpm, first beat 0.48 s), and a striking cold open. The v1 results below still
hold — every check was re-run on the v2 render.

Owner: **Apollo-launch**. Script: `shared/SCRIPTS/launch_overview.md` (v1, locked; only trims to fit the
90-second budget — no claim IDs removed, no new numbers). Claims: `shared/CLAIMS.md`.
Cost figures: `shared/cost-model/outputs.csv` (reference self-check **PASS**, re-run before any render).
This file records what was checked, with the command or measurement behind each result.

## 1. Delivered — 1080p only (supervisor instruction, 30 Oct 2026)

| File | What |
|---|---|
| `deliver/launch_overview.mp4` | **1920×1080** H.264 master, 30 fps, **94.000 s**, 2820 frames, AAC 320 kbps 48 kHz stereo, faststart |
| `deliver/launch_overview.srt` | 19 caption cues, timed to the delivered voice track (the hook cues included) |
| `deliver/launch_overview_transcript.txt` | narration transcript with timecodes |
| `deliver/CLAIMS-MANIFEST.md` | every on-screen figure → claim ID, label, source; and what is deliberately absent |
| `deliver/sources/` | editable sources: HyperFrames composition, three.js scenes, audio chain, captions |

**Provenance of the 1080p master.** The composition is authored on a 3840×2160 grid (all type sizes and
the legibility guarantees are defined there). It was rendered natively at 3840×2160 and the delivered
1920×1080 master is a **lanczos downscale of that render** — i.e. a supersampled 1080p master, which is
sharper than a native 1080p rasterisation of the same layout. The 4K file is **not** part of the
delivery and has been removed from `deliver/`. A native-1080p re-render was attempted (a scaled render
root) and rejected: HyperFrames lays scene hosts out against the root box, so the scaled root rendered
only one quadrant. Producing a native 1080p master would mean re-authoring every dimension in all five
acts; say the word and I will do that pass.
| `deliver/launch_overview.srt` | 19 caption cues, timed to the delivered voice track (the hook cues included) |
| `deliver/launch_overview_transcript.txt` | narration transcript with timecodes |
| `deliver/CLAIMS-MANIFEST.md` | every on-screen figure → claim ID, label, source; and what is deliberately absent |
| `deliver/sources/` | editable sources: HyperFrames composition, three.js scenes, audio chain, captions |
| `deliver/QC.md` | this file |

Every source is natively 4K or better: the three.js scenes were rendered at 3840×2160 (not upscaled),
the plates are 4K Nano Banana Pro stills, and all type is live HTML/SVG. No 1080p footage is upscaled,
and no prior-result clip is used in this film at all.

## 2. The film

Five acts: the job and the cost question (0:00–0:15.4) · the loop behind one task (0:15.4–0:35.0) ·
where the work runs (0:35.0–1:00.0) · the accepted outcome and the cost receipt (1:00.0–1:23.5) ·
one technical next step and the end card (1:23.5–1:34.0).

**Exactly two headline numbers**, both labelled in the same frame as the figure:

* **A4 · MEASURED** — failed vs passed SQLite attempts, median **48 vs 47 core-seconds**, drawn as two
  bars of deliberately identical length, with the p10–p90 spread and the A1 scope line
  (`SQLite agents · pinned AMD EPYC 9R14 · Qwen3 Coder Next on Bedrock · not NPU, Halo or MI350P
  validation · no Router-uplift claim`) on screen.
* **C11 / C12 · MODEL** — the cost receipt for one accepted job, in three labelled columns:
  **$0.0359** frontier-priced, **$0.0036** owned **marginal**, **$0.0057** owned **all-in** — stated as
  **≈9.9× vs marginal · ≈6.3× vs all-in**, with the tool / retry / verification terms visibly identical
  across all three columns and the price-assumption line (which defines marginal = power and ops, all-in
  = plus capex ÷ 36 months) on screen for as long as the totals are.

Supporting labelled figures (visually subordinate, never headline-sized): **I2** (~250 concurrent coding
agents per MI350P, 8-card counts labelled MODEL/extrapolated), **C13** (retry multiplier 1.25×–3.03×),
**C1** (the method), **C16** (price inputs), **R-G** (the Router gateway guardrail).

## 3. Checks and results

| Check | Method | Result |
|---|---|---|
| Cost model is authoritative | `python3 shared/cost-model/cost_model.py` | **PASS** — reference self-check reproduces 68.7 % occupancy, $33,111 / $419,451 / $34,644, 1.6 / 33.1 months, 74.7 % break-even; every receipt figure read from `outputs.csv`, none retyped |
| **Every on-screen receipt figure vs the model** | `python3 work/verify_numbers.py` — recomputes each printed value and both ratios from `outputs.csv` | **PASS** — 12/12 line items and 2/2 ratios match (frontier $0.0359; owned marginal $0.0036; owned all-in $0.0057; 9.86× → 9.9×; 6.34× → 6.3×) |
| Brand palette — no green or amber | `work/frame_audit.py` on the delivered file, 46 samples across 91 s | **PASS** — max banned-hue fraction **0.00000**; every status colour is indigo `#465CDA` / orchid `#CC28AF` / red `#FF3132` / grey `#737373` |
| Label legibility | composition type sizes × 0.5 (4K → 1080p) plus the three.js legibility audit | **PASS** — smallest label 24 px, smallest body copy 29 px at 1080p (rule: labels ≥ 22 px, text ≥ 28 px); the three.js scenes audit at ≥ 51.9 px at 2160 px and fail the render below that |
| Logo | pixel inspection of the composited asset | **PASS** — exact `metrum_logo_white_new.png`, **2244×405**, aspect 5.5407, RGBA alpha preserved, composited never generated |
| Audio master | `ebur128` on the delivered file | **PASS** — **−14.5 LUFS** integrated (target −14), **−1.5 dBTP** true peak (limit −1) |
| v2 visual energy (supervisor note) | frame review across the film | **PASS** — hero plates fill the frame and move continuously (3 per act, cross-faded, parallax + push-in), three.js wireframes screen-blended over them, additive glow on every accent, depth fog, 3–5× denser particle streams, a new corridor scene for acts 2/4, and gradient flashes + light sweeps landing on the music's beat grid (138 bpm, first beat 0.48 s). Cold open is a bright moving hero image, not a dim ring on black; no stretch of the film is a near-empty frame |
| Music ducking | `ebur128` of the narration stem vs the music bed over the same 20–30 s window | **PASS** — narration −16.9 LUFS, bed **−29.9 LUFS** = **13.0 dB under speech**; the bed builds from −37 dB to a plateau peaking 56–72 s (≈ two-thirds through) and resolves cleanly to −34 dB by 88 s |
| Narration is present, in the right places, and intelligible | local ASR (`faster-whisper`, CPU, no upload) over the delivered file's audio, diffed against the locked script | **PASS** — 18/18 sentences present in script order; pronunciation correct for EPYC ("epic"), MI350P ("M-I three-fifty-P"), Ryzen, Strix Halo, PowerEdge |
| Captions match the delivered voice | SRT generated from the ElevenLabs character alignment, not from the script | **PASS** — 18 cues, 0:00.600 → 1:29.470, first and last word of each cue anchored to the alignment; 3 cues run 17.4–18.9 chars/s against a 17 chars/s advisory ceiling (technical terms, spoken rate) |
| No forbidden content | `work/text_audit.py` over every visible string in the composition | **PASS** — no the agent team/mascot/pantheon, no hostnames, instance types or regions, no internal tool names, no private links or paths, no energy/tokens-per-joule claim, no budget or production-spend figure, no vendor claim presented as ours |
| Every number carries a claim ID and label | `work/text_audit.py` + `CLAIMS-MANIFEST.md` | **PASS** — chips rendered in frame: `A1 · MEASURED · SCOPE`, `A4 · MEASURED · CORE SIZING PILOT, SQLITE WAVES`, `C13 · MODEL · ON MEASURED PASS RATES`, `I2 · MEASURED CLAIM · METRUM TESTS`, `C1 · METHOD`, `C11 · C12 · MODEL · ONE ACCEPTED JOB`, plus the MODEL assumption line under the receipt |
| Generated imagery is used as plates only | review of `hf/compositions/*.html` | **PASS** — the six Nano Banana Pro stills carry no numbers, labels or logos; all figures are live text; each plate is animated (parallax / scale / cross-fade), never a flat slideshow |
| HyperFrames contract | `npx hyperframes check` | **PASS** — 0 errors; contrast 21/21 WCAG AA; remaining warnings are the framework's 56 px type floor (stricter than this brief's 22 px label rule) and the intentional absence of audio inside the composition (the mix is muxed in `work/assemble.sh`) |
| Every frame watched | 19 stills sampled across the film (3 s → 90.4 s) plus 4K 1:1 crops of the headline and the receipt | reviewed — pacing, cuts, text placement and the two-bar shot read correctly; no placeholder, no black frame, no clipped text |

## 4. Known limitations — named, not hidden

* **No human listened to the final mix.** The audio was verified by measurement (loudness, true peak,
  the 13.0 dB speech-to-bed ratio, the build envelope) and by local ASR of the delivered file; the
  judgement of how the narration *sounds* against the music is a listening check this agent cannot make.
  It should be heard once before publication.
* **Three caption cues run slightly above the 17 chars/s advisory ceiling** (17.4–18.9). They are short
  technical sentences spoken at that rate; the cues are timed to the voice, so slowing them would
  desynchronise the captions.
* **`hyperframes check` keeps 2 type-floor warnings** for mono labels at 48 px (24 px at 1080p). The
  framework's floor is 2.6 % of the short side (56 px at 4K); PRODUCTION.md's label rule is 22 px at
  1080p, which 48 px at 4K satisfies. The labels are declared as telemetry labels, not body copy.
* **The music bed is quiet by design** (13 dB under speech, as directed). Over continuous narration it
  reads as a bed rather than a feature; the build is audible in the gaps and under the receipt.
* **Not a new measurement.** Every number in this film is a labelled claim from `shared/CLAIMS.md`.
  The film adds no benchmark, no capture and no measurement of its own.
* The three.js scenes are rendered at 30 fps (the film's rate) rather than 60 fps; the master is a
  genuine 3840×2160 30 fps H.264 file, and no source was upscaled.

## 5. Re-run

```bash
python3 shared/cost-model/cost_model.py                 # cost model self-check (must PASS)
cd 3D && node render.mjs --scene ov-ring --hud=0 --seconds 15.4 --fps 30 --w 3840 --h 2160 --out ../work/3d/out/ov-ring
#   …same for ov-tiers (25.0 s) and ov-next (10.4 s)
python3 work/captions.py                                # SRT + transcript from the TTS alignment
bash work/mix.sh                                        # narration + music bed, mastered
bash work/assemble.sh                                   # HyperFrames 4K render + mux + 1080p + captions
python3 work/text_audit.py && python3 work/frame_audit.py deliver/launch_overview.mp4
python3 work/asr_film.py                                # local ASR of the delivered audio
bash work/package_sources.sh                            # deliver/sources/
```
