# launch_deck — QC report

Deck: `deliver/launch_deck.pptx` (15 slides) + `deliver/launch_deck.pdf`.
Cost-model self-check: **PASS** (imported; reference conformance enforced at import).

## 1. Speaker notes
- slides with notes >= 200 characters: 15 / 15
- notes ending with a `Source:` line: 15 / 15

## 1b. Legibility floor (v2: body >= 14 pt, labels >= 11 pt, titles 28-32 pt)
- smallest text size found anywhere in the deck: 11.0 pt (OK)
- Poppins runs below 14 pt: none
- content slides with a 28-32 pt title: 12 / 12

## 1c. Layout overflow check

- no two live text boxes overlap
- no text box crosses its panel or the footer zone (limit 6.44 in)
- sensitivity verified on a scratch copy: pushing one text box 1.8 in down flagged both a footer overflow and a panel overflow, so the check catches the class of defect it is meant to catch

## 1d. Chart data-label check

- chart data labels checked: 25
- every data label is short enough to render on one line (<= 4 characters,
  and narrower than its bar) — calibrated against the rendered slides, where
  5-character labels wrapped in the short charts and 4-character labels never did
- sensitivity verified on a scratch copy: injecting one 5-character label ("39.00")
  is flagged, so the check catches the wrap class

## 2. Claim-ID coverage
- distinct claim IDs referenced: 79
- claim IDs not defined in CLAIMS.md: none
- source-deck labels present (documented as source labels, never claim IDs): []
- content slides (1-12) carrying a label: 12 / 12

## 3. Forbidden content audit
- the agent team / mascot / pantheon: none
- internal tool or agent names: none
- hostnames / instance ids / paths: none
- private links: none
- our own campaign or production spend: none
- internal coordinator: none

## 4. Cost-figure provenance
- distinct dollar figures on slide faces: 26
- `$150.76` on slide 2 with its scope line (whole pilot / study cost / never a per-job price): yes
- `$150.76` on slide 9 with its scope line (whole pilot / study cost / never a per-job price): yes

## 5. Palette audit (rendered slides, no green / amber / orange / yellow)
- highest banned-hue fraction: 0.000% (slide-15.png); all 15 slides below the 0.5% limit: yes

## 6. Logo and fonts
- logo instances found: 15; aspect 2244:405 preserved: yes
- fonts embedded in the review PDF: BAAAAA+GeistMono-Regular, CAAAAA+Geist-Regular, DAAAAA+Poppins-Regular
- fntdata parts in the PPTX: 5 (embedded for PowerPoint)

## 7. Editability
- native editable charts: 13; editable tables: 1; text boxes: 622; picture plates: 29
- no slide is a flattened image: every slide carries live text (yes)

## 8. Deck inventory

| # | Slide | Claim chips |
|---|---|---|
| 1 | An agent job is a loop. Only the verifier decides. | C1  METHOD · R19  FACT |
| 2 | The bill has five parts. Only one is the model. | A4  MEASURED · C11  MODEL · C13  MODEL |
| 3 | Where the time goes, and what each measure answers. | A5  MEASURED · A8  MEASURED · A14  MEASURED |
| 4 | Policy runs per agent key, in front of the runtime. | R2  MEASURED · R4  MEASURED · R-G  GUARDRAIL |
| 5 | The NPU serves one request at a time. | N2a-c  MEASURED · N4  MEASURED |
| 6 | Strix Halo: a measured team-tier envelope. | H1  MEASURED · H3  FACT |
| 7 | MI350P on Dell PowerEdge: the accepted-work tier. | I1  MEASURED · I2  MEASURED CLAIM · I2c  MODEL |
| 8 | The gate refused a candidate that looked like a win. | F1  MEASURED · F7  MEASURED · F2  MEASURED |
| 9 | The receipt: one accepted agent job. | C11  MODEL · C12  MODEL · A4  MEASURED |
| 10 | On-prem, cloud or frontier — with the assumptions. | C3-C5  MODEL · C6  MODEL · C16  MODEL |
| 11 | Four levers decide the economics. | C13  MODEL · C14  MODEL · C15  MODEL |
| 12 | Every number re-derives. The next step is named. | A18  MEASURED · P9  PLAN · R18  FACT |
| 13 | Appendix A1 — method and scope | A1  MEASURED · A2  MEASURED · C16  MODEL |
| 14 | Appendix A2 — cost-model assumptions | C16  MODEL · C2  ILLUSTRATION |
| 15 | Appendix A3 — sources, objections, open items | R18  FACT · R20  FACT · R21  FACT |

## 9. Scope notes and manual checks

- **Structure (v2).** 12 main slides + a 3-slide appendix (A1 method and scope, A2
  cost-model assumptions, A3 sources / objections / open items). The 13-slide v1 story
  folded the trace and the three measures onto one slide to reach 12, and the three
  claim-register slides left the deck: the full register now ships as
  `deliver/production/CLAIMS-REGISTER.md` and `.pdf`, and the notes reference it.
- **Legibility (v2).** Body text 14 pt, labels and chips 11 pt, titles 28-32 pt; one key
  message per slide with the evidence chart large. Detail that no longer fits on the
  slide moved into the speaker notes.
- **Visual inspection.** All 15 rendered PNGs (`qc/renders/slide-01..15.png`, 144 dpi)
  were inspected; overflow, chart label wrapping, axis ranges and plate contrast were
  fixed and re-rendered.
- **Numbers.** Every cost figure is read from `shared/cost-model/outputs.csv` through the
  cost model itself, not retyped; `$150.76` appears only as a whole-pilot study cost with
  its scope line and never as a per-job price; no project budget or production spend
  appears anywhere.
- **Plates.** Nano Banana Pro stills are dimmed background illustration only, labelled
  `ILLUSTRATION`; every figure and label on top is live text carrying a claim ID.

## Verdict
**PASS** — all checks clean.
