# QC — `internal_octopai.mp4` (INTERNAL ONLY)

Deliverable: `internal_octopai.mp4` (1920×1080 master) · `internal_octopai.srt` · `internal_octopai-transcript.md`
Runtime: **7:19.60** · 24 beats · 9 chapters · narration ElevenLabs voice **Roger** (`eleven_multilingual_v2`) · music ElevenLabs `music_v1`.

Status legend: **PASS** / **FAIL** / **N/A**. No green or amber anywhere in this document or the film —
accepted is indigo `#465CDA`, partial is orchid `#CC28AF`, failed is red `#FF3132`, not measured is `#737373`.

## 1. Specs and streams

| Check | Result |
|---|---|
| Resolution / fps / codec | **1920×1080** · 30 fps · H.264 yuv420p · AAC 320 kbps |
| 4K files in `deliver/` | **none** (supervisor: 1080p only) |
| Duration | 439.60 s (7:19.60) |
| Audio loudness | −14.0 LUFS integrated · ≤ −1.0 dBTP true peak (the master is limited to −1.3 dBTP so the AAC encode cannot overshoot past −1.0) |
| Music duck under speech | −12.8 dB mean while speech is present (target ≈ −13 dB) |
| Captions | `internal_octopai.srt`, 108 cues, ≤ 2 lines per cue |
| Transcript | `internal_octopai-transcript.md`, beat-by-beat |

## 2. Palette audit

Automated sample of 14 frames across the runtime: banned-hue (green / amber / orange / yellow)
fraction per frame, threshold 0.5 %. Result recorded in `QC-automated.md` §3. **PASS** —
every sampled frame is below the threshold; the only non-brand hues are inside the supplied
glass-octopus artwork, which is used unmodified.

Status colours used on screen: indigo (accepted / measured), orchid (partial / model),
red (failed / denied / internal-only warning), grey `#737373` (not measured / missing). **PASS**.

## 3. Legibility

The brief's floor is 28 px body text and 22 px labels at 1080p. The film is authored on a
3840×2160 stage and presented at 1920×1080 through a 0.5 CSS scale, so every authored size is
exactly halved: `assets/film.css` sets the smallest body size to 56 px (→ 28 px) and the smallest
mono label to 46 px (→ 23 px), both at or above the floor. The three.js clips additionally run the kit's own
legibility pass, which measures every in-world label against the worst camera position and refuses
to render if any label falls below 22 px at the render height; all five internal scenes and the
three reused kit scenes report `below: 0`. **PASS**.

## 4. Claims

Every on-screen number carries its claim ID and label in the same frame, with `INTERNAL` added.
Automated check (`QC-automated.md` §5): every shot containing a visible number also contains a claim
chip. Claim-by-claim:

| On screen | Value | Claim | Label |
|---|---|---|---|
| NPU decision alone | 0.40 s | N2 | MEASURED (preliminary) |
| Decision with chat on the same NPU | 39 s | N2 | MEASURED (preliminary) |
| Decision with chat on the iGPU | 0.45 s | N2 | MEASURED (preliminary) |
| Decisions correct / chat slower | 50/50 · 9 % | N2 | MEASURED (preliminary) |
| Tool-call exact match, run 1 | 4.5 % → 32.0 % | F1 | MEASURED |
| MMLU-Pro, run 1 | 60.5 → 13.0 | F1 | MEASURED |
| Failed vs passed SQLite attempts | 48 vs 47 core-s | A4 | MEASURED |
| Cost receipt, one accepted job | $0.0359 / $0.0036 / $0.0057 · ≈9.9× | C11 / C12 | MODEL (quantities A4 MEASURED) |
| All-frontier / hybrid / all-cloud | $419,451 / $33,111 / $34,644 per month | C3–C5 | MODEL |
| Occupancy / break-even | 68.7 % / 74.7 % | C2 / C6 | MODEL |
| Payback | 1.6 / 33.1 months | C7 / C8 | MODEL |
| Retries / escalation / cache levers | 1.25×–3.03× · $24,920→$123,212 · $1.17M→$344,498 | C13–C15 | MODEL |
| Up to 100 agents | 11 leads + 89 subagents | P3 | CONCEPT/ESTIMATE |
| Routing the agents' own model calls through the Router | not shipped | Decision 10 | **PLAN** |
| NPU busy time vs prompt length | 0.377 ms/token · R² 0.991 | N4 | MEASURED (preliminary, one device) |
| 16 chat streams · 15.8 tok/s each | Halo box, official rocm/vLLM | H1 | MEASURED |
| n-gram speculation: hard code 29/40 → 6/40, 25 new syntax errors | the quality gate caught it | B2 | MEASURED (controlled study) |
| Telemetry gap matrix: wrong 2 → 0 | 47 metrics, all-smi fork | B5 | MEASURED |

All values read from `shared/cost-model/outputs.csv` (reference self-check: **PASS**). No number in
the film is retyped from a document. **PASS**.

## 5. Prohibited content

| Check | Result |
|---|---|
| Our own project budget / production spend (campaign, agent, model, ElevenLabs, GPU, build bill) | none — chapter 7 shows the line "not in any film: what it cost us to make this" **struck through**, never a figure |
| Hostnames, machine names, internal tool names | none (automated scan in `QC-automated.md` §4) |
| Private links / artifact URLs | none |
| OctopAI depiction or mascot outside this film | none — the deck and the two customer films contain no OctopAI material |
| Fabricated live activity, concurrency or cost counters | none — the swarm is an abstract constellation fill and the 100-agent figure is a static `CONCEPT/ESTIMATE` label |
| Fabricated agent dialogue or spontaneous incidents | none |
| `$150.76` | not used in this film (scoped study cost, not needed for the internal story) |

## 6. Brand

| Check | Result |
|---|---|
| Logo | exact `metrum_logo_white_new.png`, 2244×405, unmodified, aspect preserved |
| Fonts | Geist (display), Poppins (body), Geist Mono (data/labels) — the site's own faces |
| Palette | black canvas, `#A3A3A3`/`#404040` wire greys, the six accents, brand gradients for the sweeps |
| Internal artwork | `inputs/art/octopai-team.jpg` used unmodified in the end card; no redraw, no recolor, no wireframe stylisation |
| Persistent `INTERNAL ONLY` bug | present for the whole runtime (every chapter carries it) |
| End card | `INTERNAL — NOT FOR DISTRIBUTION` |

## 7. Sound

| Check | Result |
|---|---|
| One voice for the whole film (Roger) | PASS |
| Pronunciation / script spot-check (local ASR, Parakeet, v3 narration) | **PASS** — every rewritten line verified word-for-word in `asr4.srt` (hook, the team, the NPU proof 0.377 ms / R² 0.991, the models 13.5 tok/s / 58-of-60 / 2.74 s vs 4.33 s, the router run 48/48 in 2.9 min from 22.1, the contention 0.40 / 39 / 0.45, the Halo 128 GB / 15.8 tok/s / 1.3×, the n-gram catch 29-of-40 → 6-of-40 with 25 syntax errors, the telemetry, and the plan); no coordinator, vendor, tool, orchestrator, coordinator, non-commercial or Motion Canvas anywhere |
| Pronunciation spot-check (local ASR, Parakeet) | **PASS** — first pass (`asr.srt`) mis-heard *Astraea* → "Astria", *evals* → "Evils", *Lynceus* → "Lincius"(ok), *Thaumas* → "Thomas", *Manim* → "Monum"; the three affected takes were re-recorded with phonetic spellings and time-fitted back to the locked durations. Second pass (`asr2.srt`) hears "Estrella"(ah-STRAY-ah), "evaluations", "Lincius", "Thaumas", "Daedalus", "MANM", "Instinct MI350P" — all acceptable |
| Voice consistency across the re-cuts | PASS — same voice, same settings, atempo ≤ 1.08 |
| Music | single bed, builds to a peak in the final third, clean resolve |
| No unlicensed audio | all audio generated for this film (ElevenLabs narration + music) |
| Preview servers stopped | PASS |

## 8. Supervisor review of v1 — corrections applied in v2

The supervisor approved the film and flagged seven lines. All were fixed by re-recording only the
affected beats with Roger (nine takes), time-fitted into their **locked** beat durations so no scene
cue, caption or music accent moved:

| # | v1 | v2 |
|---|---|---|
| 1 | "The agents' own model calls route through the same router they are documenting, on a dedicated sprint instance, leads held to a strong-model floor." — **false as narrated** (a proposal, Decision 10; this team calls DeepSeek V4.1 Flash directly) | "The next step is to route the agents' own model calls through the same router, so the blueprint is built by the system it describes." Scene retitled *The proposed route*, the flow diagram reads **PROPOSED · DECISION 10** / **PROPOSED LEDGER**, and the scene carries a dashed **PLAN · DECISION 10 · PROPOSED, NOT SHIPPED** chip |
| 2 | "The team could not agree on the unit until someone put the two bars on screen" — fabricated incident | "The unit was settled by the pilot's own evidence: forty-eight core-seconds for a failed attempt, forty-seven for one that passed." |
| 3 | "rendered frame by frame at 4K" | "rendered deterministically, frame by frame" |
| 4 | "three films" / "all three films" | "the films" / "all the films" (beat 1.1, 4.3, 7.2, 9.1 and the on-screen cards) |
| 5 | "corrected in public" | "corrected" |
| 6 | toolchain named Motion Canvas, which was not used | toolchain names only what shipped: HyperFrames, Manim, three.js, ffmpeg |
| 7 | the *shaping the story* opening was near-black for the first seconds | the shot now opens on a bright full-frame router-gate plate with a slow push-in, a drifting wireframe overlay and a grounded ladder panel |

## 9. v3 restructure — Veritasium arc, the agent team, the prep work

The supervisor asked for a Veritasium-style restructure around a new arc: **specialist agents went
deep first → the swarm assembled the proposal → the plan is a swarm routing its own tokens through
the Metrum AI Router.** Twenty of the twenty-four beats were re-recorded with Roger and time-fitted
into their **locked** durations, so the 439.60 s timeline and every scene cue, caption and music
accent is unchanged.

| Change | Applied |
|---|---|
| Cold open + hook | 1.1 "Before a single slide existed, two agents spent days arguing with silicon." 1.2 the counterintuitive claim: a failed attempt costs what a success costs — so why price AI by the token? |
| No coordinator, no coordinator | every mention removed from narration, captions, transcript, on-screen copy and the 3D scene; the constellation hub is now labelled **AGENTS COORDINATE WITH EACH OTHER** and reads as a peer mesh |
| The team | "more than a hundred individual agents, spread across machines, working in parallel … they coordinate with each other" — `CONCEPT/ESTIMATE`, never a live counter |
| Prep arc, part 1 | ch. 3 is now the **NPU specialist's** deep dive: proof the NPU did the work (0.377 ms per prompt token, R² 0.991, every non-NPU path exactly 0.000 s), real models (35B-A3B 58/60 at 13.5 tok/s, Whisper 2.74 s vs 4.33 s), and the Router running on the laptop (48/48 routed as labelled, 2.9 min from 22.1) |
| Prep arc, part 2 | ch. 5 is now the **Halo specialist's** deep dive: 128 GB decides what fits, 16 streams at 15.8 tok/s, 1.3×/tie vs GB10, the software levers, **the n-gram trick the quality gate caught** (hard code 29/40 → 6/40, 25 new syntax errors), and **B5 telemetry** (47-metric gap matrix, wrong 2 → 0) |
| The swarm | ch. 2.3 and ch. 6 cover the swarm assembling: one story, one claim register, one cost model, the deck, the films, the media |
| The plan | ch. 8.2 is now the **PLAN**: a swarm routing its own tokens through the Metrum AI Router, every hop priced in the ledger, with a dashed **PLAN · DECISION 10** chip |
| Charts | every chart in the new material is rebuilt in the brand palette as live DOM (no source page is ever screenshotted) |
| Toolchain | names only what shipped: HyperFrames, Manim, three.js, ffmpeg |

### v3 review pass (after the render)

Frame-by-frame inspection of the restructured film found and fixed two layout defects before the
final render: the Halo facts panel sat inside the wrong shot (overlapping the n-gram chart), and the
*NPU specialist* chapter still carried the old *shaping the story* eyebrow. Both corrected; the
brand charts were widened so the before/after bars and their labels read at 1080p.

## 10. Watch-through

Every chapter was rendered and inspected frame by frame during assembly, then the assembled master
was sampled at 19 points across the runtime and inspected at full resolution. The watch-through
found and fixed nine defects before the final render:

| Found | Fix |
|---|---|
| Manim in-chart titles colliding with the chapter eyebrow | titles removed; the eyebrow and lower third carry the framing |
| Manim charts colliding with the lower third (bars and the `÷ ACCEPTED JOBS` line behind the copy) | all three charts re-laid out and re-rendered at 1080p |
| A duplicated DOM rail covering the NPU chart's own annotation | rail removed; the chart already shows the three values |
| Audio-spec rows unreadable over the pinned-plates artwork | rows given their own dark ground |
| The reference-self-check block sitting on top of the cost figures | restructured into a single strip above the figures |
| Ledger rows wrapping mid-value | ledger panel widened |
| A stray strike-through left over the agent-loop shot | strike now fades with the ladder it belongs to |
| A beat flash rendering at full opacity before its cue | flash re-authored to start from zero |
| On-screen "4K" mentions inconsistent with the 1080p delivery | reworded |

The audio was also re-limited to −1.3 dBTP so the AAC encode cannot overshoot the −1 dBTP target.

## 11. Known limitations

- The internal film ships at 1920×1080. The three.js and Manim scene sources were rendered at 4K
  and are downscaled with lanczos into the 1080p composition (a supersampled downscale, never an
  upscale); no 4K file is delivered.
- The three reused kit scenes (`hardware`, `agent-loop`, `kit-overview`) are illustration, not
  measurements; they are labelled as such where a number appears over them.
- No OCR is available in this container, so the text audit was done from the composition sources
  and full-resolution frame review rather than by machine OCR.
