Deepen operational quality evidence

[quantameridian]

What changed

The engine now makes the report date, source data and rule policy explicit. It can read CSV or DuckDB input, validates a versioned rule file, records machine readable run events and writes a manifest with source, policy and output hashes.

Three synthetic scenarios show the complete decision range. A clean input is ready for routine reporting, a second input needs review and the blocked example is not ready. The blocked example produces 39 exceptions and a score of 12 out of 100.

The score is deliberately simple enough to challenge. It does not claim that a passing record is factually correct, and the local DuckDB output is not presented as a hosted warehouse.

Checks

• make qa: 53 tests passed with 96.42 per cent coverage, and Ruff found no issues
• make audit: no known vulnerabilities in an isolated environment
• make benchmark: 100,000 rows and 1,000 deliberate exceptions completed locally
• The sample outputs reproduced without an unexpected diff
• No cache, virtual environment, package metadata, coverage file or DuckDB binary is tracked