# QC — technical_story.mp4 · “The inside story” (v2)

Owner: Apollo-tech · external inside story of the AMD + Dell + Metrum AI Router launch media.
Runtime **5:40.1** · **1920 × 1080** · 30 fps · single MP4 + `.srt` + transcript.

## 1. What was checked, and how

| Check | Method | Result |
|---|---|---|
| Runtime inside 4–6 min | `ffprobe` on the muxed master | **340.099 s = 5:40.1** |
| 1080p only | `ffprobe`; no 4K master or export exists in `deliver/` | 1920 × 1080, 30 fps, 10,205 video frames |
| Every on-screen figure carries a claim ID + label | source audit of all 26 shots (HUD chip timelines + station labels) + `production/CLAIMS-MANIFEST.md` | 39 claim IDs present, no unlabelled figure |
| Cost figures are outputs of the one cost model | `3D/src/costModel.mjs` generated from `shared/cost-model/outputs.csv`; scenes read it via `value()`/`money()`, which throw on an unknown key | pass |
| Receipt labelling | three columns on screen: frontier `$0.0359` (C11), owned **marginal** `$0.0036`, owned **all-in** `$0.0057` (C12); ratios **≈9.9× vs marginal** and **≈6.3× vs all-in** | pass |
| Per-job token quantities are MODEL, not MEASURED | on-screen line: “PER-JOB TOKENS 3,704 OUT + 74,088 IN: MODEL · ALL DOLLARS: MODEL · 47 CORE-S: MEASURED (A4)” | pass |
| Palette: no green, no amber | every rendered frame audited at render time, plus a 28-frame re-audit of the finished master | **0 frames flagged** |
| Text legibility | world labels enforced to ≥ 2.59 % of frame height (28 px at 1080p); a render *fails* if any label is below it; HUD labels 22 px at 1080p | new scenes: **0 below 28 px** (studies 81 labels, smallest 32.2 px; inside 27 labels, smallest 32.6 px); reused chapters: 0 below 28 px |
| Logo | exact `metrum_logo_white_new.png`, composited in the title and end cards | 2244 × 405, aspect 5.540741 = 2244/405, RGBA — exact |
| **No vendor or tool names** anywhere | regex scan of transcript, captions, all scene sources and both cards | **none present** |
| No the agent team, mascot, pantheon or Greek agent names | same scan | **none present** |
| No tabs, panes, hostnames, machine nicknames or orchestration tool names | same scan | **none present** |
| No production spend | same scan (`campaign budget`, `production spend`, `build bill`, `$5.51`, `$3.19`, `$51`, `gpu-hours`) | **none present** |
| No brief quoting or paraphrase | script review: the film shows the work and the findings; it never quotes the assignment | pass |
| The agents’ own model calls are not claimed to go through the router | stated only as the next step, labelled **PLAN**, on screen | pass |
| Prep-phase numbers quoted exactly, caveats kept | every number on the two study charts traced to `shared/SOURCE-STUDIES.md`; charts redrawn in brand colours, **no source page screenshotted** | pass |
| FastFlowLM licensing | the “non-commercial” wording is not used anywhere (the licensing note is omitted entirely) | pass |
| Audio: −14 LUFS integrated, ≤ −1 dBTP, music ducked ≈13 dB | `ebur128` on the master; duck depth measured inside a speech window on the music bus | **−14.3 LUFS · true peak −1.2 dBFS · duck 13.1 dB** |
| Captions and transcript match the audio | both generated from the same `audio-inside/timeline.json` the narration is placed from | pass |

## 2. Media

```
deliver/technical_story.mp4   1920x1080  30 fps  h264 + aac 320k  340.099 s  221.1 MB
deliver/technical_story.srt   24 cues, 00:00:03.500 -> 00:05:33.420
deliver/technical_story_transcript.md
```

A single 1920×1080 MP4 per film (lineup note, 30 Oct 23:25Z): no `_1080p` duplicate and no 4K file.

## 3. Story arc (Veritasium-style, prep → swarm → routed swarm)

Hook question (0:03.5) → the failed-attempt evidence (0:16.7) → the retry multiplier (0:30.3) →
where the numbers come from (0:42.5) → **surprise 1:** zero throttling yet waiting (0:55.8) →
**the NPU specialist agent's deep dive** (1:15.2–2:19.8: the counters, 58/60, the contention surprise,
the decision models, 48/48 routed laptop→Instinct) → **the Halo specialist agent's deep dive**
(2:19.8–3:09.8: 16 streams, software beats silicon, the quality gate that caught a “win”, modelled
tokenomics) → the sizing questions and the refused claims (3:09.8–3:47.7) → the mental model and
**the twist** (3:47.7–4:20.8) → the scenario and the levers (4:20.8–4:54.9) → **the swarm assembled the
proposal** (4:54.9) → **PLAN: the routed swarm** (5:08.9) → the close (5:21.9) → end card.

## 4. Correction carried forward from external verification

The receipt that previously mixed the owned **all-in** total with the ratio against owned **marginal**
is fixed: three columns, each labelled `MODEL` (C11 / C12), with both ratios stated on screen
(`≈9.9× vs OWNED MARGINAL · ≈6.3× vs OWNED ALL-IN`) and the distinction spelled out
(“MARGINAL = POWER + OPS; ALL-IN ADDS CAPEX ÷ 36 MO”). The per-job token quantities now read `MODEL`
explicitly, with the 47 core-seconds marked `MEASURED (A4)`. The HUD eyebrow reads “one job, **three**
prices”.

## 5. Known limitations

- The palette re-audit samples every 12th second at 960 px wide; a banned hue covering under 0.5 % of a
  frame would not trip it (the render-time audit checks every frame at full resolution).
- No OCR in this container: the on-screen text audit is done on the sources that generate the text, plus
  a 28-frame contact-sheet review of the finished master.
- The abstract node constellation and the routed-swarm diagram are **concept illustrations**, labelled as
  such on screen; they are not live counters or live activity.
- The tokenomics chart is **MODEL**, with its formula and assumptions printed beside it; operations time is
  explicitly excluded and called out as the deciding factor.
- Study speed numbers come from Metrum AI’s own benchmark and quality from a deterministic 60-case suite;
  the NPU results remain **preliminary, one device**, as labelled on screen.
- The title and end cards load GSAP from a public CDN at render time; nothing else in the film reaches the
  network, and every scene is deterministic (frame *N* is a pure function of *N*).
- The `85 %` measured model-spend reduction (`C18` / `R4`) is not in this film; it appears in the deck and
  the 90-second overview.

## 6. Verdict

**PASS.** Inside its length band, 1080p, legible, on palette, correctly labelled — including the marginal
vs all-in receipt — free of internal names and of any production-spend figure, and with captions and
transcript matching the audio.
