# launch_overview — claim & source manifest

Every figure that appears on screen in `launch_overview.mp4`, with its claim ID and label from
`shared/CLAIMS.md` (Calliope, v1) and the source it comes from. Nothing in the film is outside this
list; nothing here is a new measurement.

| On screen | Claim | Label shown | Source |
|---|---|---|---|
| `A1 · MEASURED · SCOPE` — SQLite coding agents on pinned AMD EPYC 9R14 cores (AWS), Qwen3 Coder Next on Amazon Bedrock, 25–26 Oct 2026; *not NPU, Halo or MI350P validation; no Router-uplift claim* | **A1** | MEASURED (scope) | `inputs/agenticbench-core-sizing-pilot.html` `#cpuonly`, `#model` |
| FAILED 48 core-seconds vs PASSED 47 core-seconds, median per attempt (p10–p90 44–103 vs 42–112) | **A4** | MEASURED | pilot `#outcomes` |
| 20% / 33% / 50% / 67% failures → 1.25× / 1.49× / 2.00× / 3.03× cost per accepted job | **C13** | MODEL (on measured pass rates) | `shared/cost-model/outputs.csv` §4.5 |
| `LOCAL DECISION · NPU` / `TEAM WORKLOAD · STRIX HALO` / `SHARED INFERENCE · MI350P` — tier identities only, no figures | N1/H1/I1 scope chips (no numbers) | (no figure shown) | `shared/CLAIMS.md` §1–3 |
| ~250 concurrent coding agents per MI350P; 8-card node counts extrapolated | **I2** (+ I2c) | MEASURED CLAIM (Metrum tests) · MODEL for the 8-card extrapolation | `shared/CLAIMS.md` §3 |
| `48 of 48 prompts routed as labelled` — laptop NPU → Instinct GPUs, router on the edge tier | **S-A3** | MEASURED (one test set) | `shared/SOURCE-STUDIES.md` §A3 (supervisor-approved 31 Oct) |
| `ROUTER = GATEWAY, NOT THE SERVING RUNTIME` + limits (no guaranteed quality, atomic budgets or automatic privacy) | **R-G** | FACT (guardrail) | PROMPT.md "EVIDENCE THAT MUST SURVIVE" |
| cost per accepted job = (inference + tools and sandbox + retries and failed attempts + verification + operating overhead) ÷ accepted jobs, latency stated | **C1** | METHOD | PROMPT.md; `shared/COST-MODEL.md` §1 |
| one accepted job = 3,704 output + 74,088 input tokens, 47 core-s | **C2 / C11** | MODEL (quantities: 47 core-s MEASURED) | `shared/COST-MODEL.md` §3–4 |
| receipt, **frontier list** column: $0.0326 + $0.0024 + $0.0008 = **$0.0359** | **C11** | MODEL | `outputs.csv` `receipt_frontier`, model §4.4 |
| receipt, **owned marginal** column at 68.7% busy: $0.0004 + $0.0024 + $0.0008 = **$0.0036** | **C12** | MODEL | `outputs.csv` `receipt_owned_marginal`, model §4.4 |
| receipt, **owned all-in** column at 68.7% busy (capex ÷ 36 months): $0.0024 + $0.0024 + $0.0008 = **$0.0057** | **C12** | MODEL | `outputs.csv` `receipt_owned_allin`, model §4.4 |
| ratio **≈9.9× vs marginal** · **≈6.3× vs all-in** | **C12** | MODEL | derived: 0.03588/0.003639 = 9.86 → 9.9×; 0.03588/0.005656 = 6.34 → 6.3× |
| price assumptions: GPT-6 Terra list $1/$5 per M, 20:1 input:output, 90% cached at 10% of list, one 8× MI350P node at 68.7% occupancy | **C16** | MODEL / VENDOR CLAIM (list prices) | model §2 |
| `NEXT: cost per accepted job, measured end to end — with the latency stated`; dashed = not yet measured | **P9 / A20** (next step) | PLAN | PROMPT.md slide 12; pilot `#followup` |

## Deliberately absent

- **No figure for Metrum's own production spend, campaign budget or tooling cost** (`F3`, `F6` and every
  ledger line are excluded from all media).
- The pilot's whole-study cost (`A6`, $150.76) is **not** used in this film: the 90-second overview carries
  the non-dollar measured pair (A4) and the receipt instead, so the study total cannot be misread as a
  per-job price.
- No the agent team, mascot, pantheon roster, hostname, instance type, region, internal tool name or private link
  (external film).
- No energy-efficiency or tokens-per-joule claim (no NPU power was measured).
- No vendor claim presented as a Metrum result.

## Claims added from `shared/SOURCE-STUDIES.md` (v3, supervisor-approved 31 Oct 2026)

Two measured beats from the specialist studies were added, quoting the numbers exactly and keeping their
caveats. Both are **MEASURED by Metrum AI** on a single test set / a single box — never presented as
general results. Media rules for this material are followed: no orchestration tool, tab, pane, hostname,
model or tool internals are named, and no source page is screenshotted.

| On screen | ID | Label | Source |
|---|---|---|---|
| `48 of 48 prompts routed as labelled` — laptop NPU → Instinct GPUs, router on the edge tier (Act 3) | **S-A3** | MEASURED (one test set) | `shared/SOURCE-STUDIES.md` §A3 |

*(The Strix Halo beat `16 interactive chat streams on one desk box`, **S-B1**, is used in
`launch_about.mp4`.)* These two IDs are not yet in `shared/CLAIMS.md` — a request to add them is filed in
`shared/REQUESTS.md`.

## Plate inventory (v2) — ILLUSTRATION, no numbers, no logos

Every plate is a Nano Banana Pro still from `shared/IMAGES/` used **full-frame** with a continuous
parallax / push-in and the three.js wireframe screen-blended over it. None carries a number, a label
or a logo: every figure on screen is live text.

| Act | Plates (in order) |
|---|---|
| 1 | `hero-agent-swarm-city` (cold open) → `agent-loop-schematic` → `hero-token-river-datahall` |
| 2 | `hero-token-river-datahall` → `retry-loop-metaphor` |
| 3 | `hero-npu-silicon-macro` → `hero-halo-team-desk` → `hero-mi350p-exploded` → `hero-router-gate` |
| 4 | `hero-receipt-unfurling` → `cost-receipt-plate` → `hero-router-gate` |
| 5 | `hero-accepted-job-vault` → `hero-mi350p-exploded` |

The HERO-01…HERO-08 plates were rendered by the image studio for this film (see `shared/IMAGES/INDEX.md`).
The six extra plates requested in `shared/IMAGE-REQUESTS.md` (IMG-30…IMG-35) were not needed once the
HERO set landed; the request is left on file as the fallback it named.

## Reproduce

```bash
python3 shared/cost-model/cost_model.py            # reference self-check (must PASS) + every C figure
python3 work/verify_numbers.py                     # every on-screen receipt figure vs outputs.csv
python3 work/text_audit.py                         # banned-content scan over the film's on-screen text
python3 work/captions.py                           # SRT + transcript from the TTS alignment
python3 work/frame_audit.py deliver/launch_overview.mp4   # palette + legibility audit of the render
```
