derived-artifactslisted

# Derived artifacts Anything that restates data from elsewhere — a generated page, a report script, a summary table, a count in a doc — is a derived artifact. These rules make it correct by construction rather than correct by luck. Each one comes from a run that got it wrong. ## 1. Go to the structured source, not the generated prose - Before parsing a file, ask what **produced** it. If it is itself generated, read the generator's source of truth instead — parse what the builder parses, or have the builder emit structured data (JSON/CSV) that both it and your artifact consume. - **Never label-match on prose.** Substring matching on rendered text is the bug class: a matcher for `documented` also hit `— derivable from documented…` and reported **64 documented when the real figure was 44**. - Prose is a rendering, not an interface. Its wording changes for editorial reasons, and every such change silently breaks a parser downstream. - If no structured source exists, the fix is to **create one** — add a machine-readable emit to the generator — not to write a tighter regex. ## 2. Assert the parse against the source's declared totals - Every derived number must be checked against a total the source states about itself: a `Totals` table, a summary row, a manifest count. Compare them, and **fail loudly** — raise or exit non-zero — on any mismatch. Never warn and continue. - Print both sides in the failure message: `parsed 64 documented but Totals claims 44`. A b