Walking the Variance Workbench: What to Inspect in 30 Minutes

Evidence label: Prototype. The Finance Variance Review Workbench is an interactive browser demo on synthetic data. It does not call a model, connect to a ledger, or claim production outcomes.

Why this walkthrough exists

Hiring screens for finance-domain engineering often ask whether you can show a control-aware workflow, not only a slide. This note is the companion to the live demo: what to click, what to verify, and what the prototype refuses to pretend.

Open the demo in another tab.

Minute 0–5: orientation

Confirm the page states synthetic data and in-memory review. Scan the queue counts (total, open, reviewed) and the net variance figure. Those numbers should move only when you change review state—not when the page invents new actuals.

Minute 5–15: zero-plan and sources

Select a record where plan is zero (or locate one via filters if present). Check that percentage handling does not divide by zero or invent a meaningless rate. Open the detail panel and confirm synthetic source identifiers sit beside the record they support. Context text should read as fixture text, not as generated reasoning.

Minute 15–25: review states

Add a short note. Mark the exception reviewed, then escalated (or the reverse). Confirm counts and filters update, and that a reload clears state—there is no durable audit store by design. That limit is intentional for a public sample; a production path would need permission-aware persistence.

Minute 25–30: source and tests

Skim the implementation source and tests. Look for deterministic calculation helpers and explicit review-state transitions. If you cannot find those, the demo is theater.

How this relates to agentic FP&A

Agents that draft variance narrative still need the same spine: visible inputs, ranked exceptions with owned rules, and human approval before anything leaves the team. See AI Agents in FP&A. For the static statement twin of this pattern, see Why the Statement Workbench Is Static on Purpose.