Validation & WAP — Week 3 CheckWeek 3 of Data Quality, Contracts & Observability: Write-Audit-Publish, table diffing, and the coverage-vs-noise tradeoff — from Dropbox, Meta, LinkedIn, and Checkout.com.

In Write-Audit-Publish, what exactly does the audit gate guarantee that validate-in-place cannot?

Dropbox's validation framework cut data-quality incidents ~95% in a year. Which TWO design choices define it?

Meta's Table Compare exists because of one blunt observation about pipeline changes. Which?

Why do Table Compare and ValiData both invest heavily in normalization (float rounding, array sorting, deterministic sampling)?

LinkedIn's ValiData took validation from ~2 hours to ~15 minutes per pipeline change. What's the mechanism?

ValiData's adoption lesson: usage took off when submission became one-click. What's the general principle?

Checkout.com runs dbt assertion tests AND ML-based anomaly detection. Why both?

Meta's forward-looking argument: AI-assisted code generation makes output validation MORE necessary. Why?