/qa-data

Trust nothing about the data just because a Dev session said so. This skill re-derives every fact from the target repo itself — its repo config, its schema, its database — checks the data against what fabrication promised, and files what it finds in the fixed report format. A dataset that passes here can be handed to demos, staging, and tests with confidence.

Ground rules

• Fresh session, re-derived context. No memory of the Dev session survives here (ADR 0004). Everything reads from the repo: docs/agentic-skills/data.md, docs/agentic-skills/fabrication-manifest.json, the schema, the database. If docs/agentic-skills/ does not exist, stop and say /agentic-skills-init must run first; if the manifest is missing, that is itself a finding, not a reason to skip the audit.
• Read-only. Never modify a row, never "fix" a finding in place. Repairs belong to the Dev side; this skill only measures and reports. Every query you run must be a SELECT, a PRAGMA read, or the driver's equivalent check.
• Whole dataset, not a sample. Counts and violations come from queries over every row; spot-checks only illustrate, they never decide.
• Stack-agnostic. The repo's schema files and docs/agentic-skills/stack.md say what the database is and how to reach it. Adapt every query below to the repo's dialect; the schema is the truth, these words are not.

Inputs to re-derive first

1. docs/agentic-skills/data.md — the conventions the data must uphold: profiles and volumes, locale blend band, default seed, marker convention, manifest convention.
2. docs/agentic-skills/fabrication-manifest.json — the recorded run: seed, profile, ID rule, marker, per-table counts and ID ranges. The audit checks the database against this record.
3. The schema — tables, key relationships, which columns may be NULL, from migrations, SQL files, or ORM models.
4. The database itself, reached the way the repo reaches it.

The audit areas