editor

ADR-0020 — Every feature starts with a research spike and ends with an audit ​

Status: Accepted

Context ​

The Product Owner wants assurance that, at the end of the project, everything makes sense: mathematics correct, every source real and still valid, no copied code, no duplicate, documentation consistent. Research done story by story discovers missing or unverifiable sources late (EN-004: REF-GG-FILLET could not be read and a derivation was written mid-story).

Decision ​

Every feature follows the same frame, enforced by the backlog test:

  1. Research spike first (SP): before any implementation story, documentary research for all the feature's stories — each formula, algorithm and behavior gets a verified source, a derivation, or a research request. Its output: rows in docs/references.md, derivations, a research note, and stories updated (criteria, @see targets, blocked points).
  2. Implementation stories (US, EN), as before.
  3. Audit (AUD): a full check of the feature in depth and of the whole project in breadth, following the checklist of audit.md: mathematics recomputed, sources re-verified, provenance of the code, duplicates, consistency of code, domain, ADRs and backlog. Findings are fixed in the audit or become stories.
  4. Validation (VAL): the Product Owner validates the feature's acceptance criteria.

Regression and end-to-end tests are not part of this frame yet: they come when the applications exist (Product Owner, 2026-10-08).

Consequences ​

  • Two more stories per feature; a feature cannot be closed without its audit.
  • Sources are found before coding, so a story is rarely blocked mid-way.
  • F01, refined before this decision, receives its spike (for its remaining stories) and its audit now.

References ​

ADR-0017, audit.md, research protocol