A deterministic check for acceptance criteria that no test covers

[guillaume-flambard]

Hi everyone,

09-for-sure carries a rule I keep coming back to: "never set
status: implemented until the success condition genuinely passes". Same spirit
in 03-assert and 06-test. It is the right rule. It is also an instruction
addressed to a model, and nothing outside the model checks it.

So I spent a day asking a smaller question: which acceptance criteria in a
repository actually have a test behind them? Not "is the test good", just "is
there one, and which".

I ran the measurement on this repository, at f6c61cf, because it is the
largest public corpus of AIDD task documents and it is written by the people who
designed the format. Reading every document under aidd_docs/tasks:

Criteria found	315 (74 Done-when bullets, 241 checkboxes)
Ticked boxes	202
Ticked boxes with no test title behind them	200
Criteria sharing no significant word with any test	0

Three honest caveats before anyone reads that as an accusation:

1. Most of those 200 are in review.md. A box there means "I read this",
   not "this behaviour is guaranteed". My tool leaves review documents out of its
   default scope for exactly that reason; the number is here because you asked
   for nothing and I would rather show the whole measurement than the flattering
   half.
2. The last row is the real finding. Zero criteria out of 315 fail for lack
   of a candidate: every single one shares vocabulary with some test title in the
   repository. What is missing is not tests. It is the link between a criterion
   and the test that covers it, and that link is one line.
3. The plan template ships a #### Acceptance criteria section that nobody
   fills. Zero such headings in the repository; plans carry ## Guards and
   ## Proof prose instead. That is a fact about the template, not about anyone's
   discipline, and it matters for anything built on top.

The tool is here, MIT, working, CI green, 161 tests:
https://github.com/guillaume-flambard/aidd-guard

It reads ## Done-when bullets and - [ ] checkboxes, reads Vitest and Jest
test titles from the TypeScript syntax tree, and matches the two. Never runs a
test, never imports your code, never calls a model, and gives identical bytes for
identical input. A criterion is linked to a test with an HTML comment, never a
heading, so your own parsers and spec-validator.yml see nothing new:

- [ ] The export produces a JSON file the user can download.
  <!-- aidd-guard:test="writes the export as JSON" -->

It specifies itself in that format and gates its own build on it: 27 criteria,
26 linked by an explicit selector, one declared non-testable with a written
reason.

Two questions for you, and I genuinely do not have a preference:

1. Does this belong anywhere near the framework? A skill in aidd-dev, a
   hook, or nothing at all and it stays a companion tool people can install.
2. Should the plan template say what plans actually say? Any checkbox-based
   check is checking a section people are not writing today.

Happy to do the work whichever way you want it, and happy for it to live under
the ai-driven-dev org rather than mine if that makes more sense. If the answer
is "interesting, but not in the framework", that is a perfectly good answer and I
will say so in the README.