# QC — `launch_about.mp4` v2 (Metrum AI · AMD + Dell + Router · what the story is about, external, ~2 min)

> **v2 (Veritasium hook + source-study beats, 31 Oct 2026).** Act 1 is now a hook: *"A cheap model call can
> still produce an expensive job. Retries do that — and almost nobody counts them. Ask what a job costs and
> you will get a token price. That is not the number."* — labelled **C13 · MODEL · ON MEASURED PASS RATES**
> with the retry multipliers on screen, paying off on the headline. Two supervisor-approved measured beats
> from `shared/SOURCE-STUDIES.md` were added to the execution-choice tiers (**S-A3**, **S-B1**). Nothing else
> changed: same length, same claims, same captions structure, same music, same receipt figures.

Owner: **Apollo-launch**. Script: `SCRIPT-about.md` (scene table with claim IDs). Claims: `shared/CLAIMS.md`.
Cost figures: `shared/cost-model/outputs.csv` (`cost_model.py` reference self-check **PASS**; no figure was
retyped from a document). This file records what was checked and the measurement behind each result.

## 1. Delivered

| File | What |
|---|---|
| `deliver/about/launch_about.mp4` | **1920×1080** H.264, 30 fps, **124.500 s**, 3735 frames, AAC 320 kbps 48 kHz stereo, faststart |
| `deliver/about/launch_about.srt` | 22 caption cues, timed to the delivered voice track |
| `deliver/about/launch_about_transcript.txt` | narration transcript with timecodes |
| `deliver/about/sources/` | `SCRIPT-about.md`, the editable HyperFrames composition, the three.js scenes (incl. the new `about-money`), the audio chain, captions |
| `deliver/about/QC.md` | this file |

Authored **at 1080p** (the composition's canvas is 1920×1080), so the type floor is the delivery floor:
24 px labels, 30 px body copy — above the 22 px / 28 px rule.

## 2. The film

Six acts, ~2 min: the premise (agents run to a job) → the bill is not the token bill (**C1**) → own the
baseline on-premises (PLAN) → the AMD execution choices and the Router control point (tier scope chips,
**R-G**) → **the tokenomics headline (MODEL)** → the one question and the close.

**The headline is the Act 5 comparison**, labelled **MODEL** with its assumptions printed on screen
(**C16**): one illustrative workload (**C2**, ILLUSTRATION — 500 seats · 4 always-on agents each ·
17 h/day · 3% escalation) runs at **$419,451/month** all-frontier (**C3**), **$33,111/month** hybrid
on-premises (**C4**) and **$34,644/month** all-cloud (**C5**) — **92.1% below all-frontier** — with
break-even occupancy **74.7%** against a modelled **68.7%** (**C6**). The three wireframe cost columns in
the 3D scene are drawn at honest relative heights; the figures themselves are live text.

No deep evidence chain by design — that is the story film's job. Two measured beats from the specialist
studies are shown, quoting the numbers exactly (**S-A3**, **S-B1**, `shared/SOURCE-STUDIES.md`,
supervisor-approved 31 Oct): **48 of 48 prompts routed as labelled** on the NPU tier and **16 interactive
chat streams on one desk box** on the Strix Halo tier. The tier identity chips (**N1/H1/I1**) still carry
no figures of their own. No orchestration tool, tab, pane, hostname, model or tool internals are named,
and no source page is screenshotted.

## 3. Checks and results

| Check | Method | Result |
|---|---|---|
| Cost model is authoritative | `python3 shared/cost-model/cost_model.py` | **PASS** — reference self-check reproduces occupancy, the three scenarios, payback and break-even; every Act 5 figure read from `outputs.csv` |
| Brand palette — no green or amber | `work/frame_audit.py` on the delivered file, 46 samples across 124.5 s | **PASS** — max banned-hue fraction **0.00000** |
| Label legibility | composition type sizes (authored at 1080p) | **PASS** — smallest label 24 px, body 30 px at 1080p (rule: labels ≥ 22 px, body ≥ 28 px); the three.js scene audit reports no label below the floor |
| Logo | pixel inspection of the composited asset | **PASS** — exact `metrum_logo_white_new.png`, **2244×405**, aspect 5.5407, RGBA alpha preserved |
| Audio master | `ebur128` on the delivered file | **PASS** — **−14.4 LUFS** integrated, **−1.5 dBTP** true peak |
| Music ducking | narration stem vs music bed over the same window | **PASS** — narration −16.1 LUFS, bed **−29.1 LUFS** = **13.0 dB under speech**; the new bed is 120 bpm and reshaped into a build that peaks at ~83 s (two thirds) and resolves at the end |
| Narration present, in order, intelligible | local ASR (`faster-whisper`, CPU, no upload) over the delivered audio, diffed against the script | **PASS** — all 22 sentences present in script order; EPYC / NPU / MI350P / Ryzen / Strix Halo / PowerEdge / PowerEdge pronounced correctly |
| Captions match the delivered voice | SRT generated from the ElevenLabs character alignment | **PASS** — 22 cues, 0:00.600 → 2:03.1; five cues run 17.4–18.3 chars/s against a 17 chars/s advisory ceiling (technical terms at the spoken rate) |
| No forbidden content | `work/text_audit.py hf-about` over every visible string | **PASS** — no the agent team/mascot/pantheon, no hostnames, instance types or regions, no internal tool names, no private links, no production spend, no energy claim |
| Every figure carries a claim ID and label | `work/text_audit.py` + `SCRIPT-about.md` | **PASS** — chips rendered in frame: `C1 · METHOD`, `C2 · ILLUSTRATION · WORKLOAD`, `C3 · C4 · C5 · MODEL`, `C6 · MODEL · OCCUPANCY`, `PLAN · …`, plus the `MODEL · …` assumption line |
| Tier chips carry no figures | review of `act4.html` | **PASS** — N1/H1/I1 appear as scope identities only |
| Router guardrail | review of `act4.html` | **PASS** — `ROUTER = GATEWAY, NOT THE SERVING RUNTIME` with its limits in the same frame |
| HyperFrames contract | `npx hyperframes check` | **PASS** — 0 errors; contrast 14/14 WCAG AA; layout 0 issues across 9 samples |
| Every frame watched | 28 stills sampled across the film plus 1:1 crops | reviewed — a claim-chip/scenario-block overlap found and fixed in Act 5, and the Act 5 opening plate replaced (a saturated gradient was swamping the cost columns) before this render |

## 4. Known limitations

* **No human listened to the mix.** Verified by measurement (loudness, true peak, the 13.0 dB
  speech-to-bed ratio) and by local ASR of the delivered file; how it *sounds* should be heard once.
* **Five caption cues run slightly above the 17 chars/s advisory ceiling** (17.4–18.3); they are timed to
  the voice, so slowing them would desynchronise the captions.
* **The Act 5 figures are MODEL, not measurements.** They are the cost model's illustrative scenario with
  its assumptions on screen; the workload itself is ILLUSTRATION. Nothing in this film is a new measurement.
* `hyperframes check` keeps a handful of type-floor warnings for 24 px mono labels — the framework's floor
  (2.6 % of the short side = 28 px at 1080p) is stricter than PRODUCTION.md's 22 px label rule.

## 5. Re-run

```bash
python3 shared/cost-model/cost_model.py                     # cost model self-check (must PASS)
cd 3D && node render.mjs --scene about-money --hud=0 --seconds 32.0 --fps 30 --w 1920 --h 1080 --out ../work/about/3d/about-money
python3 work/about/captions-about.py                        # SRT + transcript from the TTS alignment
bash work/about/mix-about.sh                                # narration + music bed, mastered
bash work/about/assemble-about.sh                           # HyperFrames 1080p render + mux + captions
python3 work/text_audit.py hf-about && python3 work/frame_audit.py deliver/about/launch_about.mp4
python3 work/about/asr-about.py                             # local ASR of the delivered audio
bash work/about/package-about.sh                            # deliver/about/sources/
```
