# Deck QC Desk

> A quality-control desk for investment banking and client presentations. Paste a deck's slide text or
> drop a .pptx; the browser checks every figure on every slide for free, a refresh lane rolls the deck
> to new numbers with a change list the user approves line by line, and a paid QC pass writes the deck
> check report. Every quote and figure the model writes is checked against the slide it cites.
> https://deck-qc-desk.skillsafe.ai/

Deck QC Desk is a web app on SkillSafe derived from two agent skills in anthropics/financial-services-plugins
(Apache-2.0): @anthropics/ib-check-deck (primary) and @anthropics/deck-refresh. It runs on gpt-terra and is
metered per run; the prescan and the refresh change list are free and need no account.

## What the free prescan computes

- Input: slide text with "## Slide N" markers (or --- separators), Markdown pipe tables, or a .pptx that is
  unzipped in the browser - text, tables and chart series data in slide order. Up to 80 slides.
- Every figure with its slide, metric (nearest keyword or table row label) and period (nearest FY / LTM /
  quarter token or the table column header). $500M, $500MM and $500,000,000 are one number; a bare cell
  under a ($M) header is read in $M.
- Cross-slide consistency: the same metric for the same period compared across slides at display precision ($0.5B and $485M agree, $318M and $305M do not)
  (a JavaScript port of ib-check-deck's extract_numbers.py, grouped by metric, unit class and period).
- Arithmetic: growth "from A to B" against the stated %, CAGR with both years, EBITDA margin against its
  revenue and EBITDA (the deck's revenue for that period when the slide omits it), table Total rows.
- Plausibility: "#1", "leading", "largest" claims tested as revenue / stated market size.
- Drift: $M vs $MM, thousands separators vs K, date formats, quarter and fiscal-year label styles.
- Sources and periods: slides with figures or charts and no source line; financial figures with no period.
- Language against the skill's IB terminology table: contractions, exclamation points, casual phrasing,
  first person, unsupported superlatives, vague quantifiers; "Adjusted EBITDA" vs "EBITDA" drift.
- Every flag has an id (P1, P2 ...) that the QC pass must answer.

## What the refresh lane does (free)

- Mapping lines: "old -> new (label)", "label old -> new", tab-separated columns, or a new value alone.
- Finds every instance of each old figure across scale ($485M, $0.485B, $485,000,000), precision ($485.0M,
  ~$485M), unit style ($485MM, $485 million) and bare chart or table numbers.
- Writes each replacement in that instance's own style, and says when the deck's precision rounds it.
- Flags derived numbers next to a changed figure (growth, margins, shares, multiples, totals) and
  recomputes growth, margins and totals where the deck gives the inputs. Nothing is applied until ticked.
- Exports the refreshed slide text, a change log ("Changed" / "Still flagged - did NOT change") and a CSV.
- Copies a find-and-replace list for PowerPoint: one pair per distinct change with its slides, and a
  caution wherever the old text stays on another slide (so Replace All would be wrong).
- Never drops slides: a refreshed deck keeps every slide, including any past the 80-slide prescan limit.

## Limits are never silent

Past 80 slides, past one run's input size, or past the change-list limits, the page says what was left
out, the model is told, every affected export says "Partial", and the download name carries "-PARTIAL".

## Re-checking a deck

A new QC pass of a deck you checked before is compared with your last pass of that deck: findings
resolved, new and still open, on the page and in the report.

## What the paid runs return

One JSON object per run, routed on `task`:

- `check`: `verdict` (not_ready, fix_before_sending, client_ready - derived strictly from severities),
  `headline`, `findings` (severity critical / important / minor, category, slides, exact quote, issue,
  action, fix_from, fix_to, prescan id), `prescan_responses` (every prescan flag confirmed or dismissed),
  `checklist`, `summary`. Number fixes hand back to the refresh lane as a mapping in one click.
- `refresh`: `headline`, `mapping` (reading, confidence, proposed old value for new-only lines), `skip`
  (instances that are a different metric), `extra` (missed instances), `derived` (stale figures with
  recalculations when the policy allows), `narrative` (wording that no longer fits), `questions`,
  `summary`. The review applies to the change list in one click; the user still approves each row.

## Limits

It reads text, not pixels: overlaps, overflow, fonts and contrast need a person. It cannot write a .pptx
back; the user edits from the change log. It knows nothing the deck does not say. Quality control, not
investment advice.

## Links

- App: https://deck-qc-desk.skillsafe.ai/
- API tutorial: https://deck-qc-desk.skillsafe.ai/api.html
- Source skills: https://skillsafe.ai/skill/@anthropics/ib-check-deck and https://skillsafe.ai/skill/@anthropics/deck-refresh