Assay

A hallmark for AI-assisted work: mint durable IDs at intent, refuse the gaps that hide, and let anyone read the proof at a glance.

For seven hundred years, England has answered one question by law: is this real gold, or is it gilt — base metal dressed to pass? The answer is an assay. The metal is tested and struck with a hallmark, so anyone who wasn't in the room when it was made can read its provenance at a glance.

AI just made that question urgent for software.

The first two notes described a machine-checkable thread running through HomesFlow: every intent given a durable ID, that ID surviving visibly into the code and the tests, a gate on every push refusing to let the chain break. It worked. But it was mine — a script wired into one repo.

Now it's a thing you install. It's called SpecAssay.

AI lowered the cost of writing code. It lowered the cost of code that merely looks right by exactly as much.

That second cost is the dangerous one. A function that reads clean, a task marked done, a test running green — none of it tied to an intent anyone approved. It's gilt: a thin gold skin over base metal, produced now at a scale no reviewer can keep up with by reading.

SpecAssay is the assay office for it. A small overlay on GitHub's Spec Kit, nothing forked. Every intent is minted with a durable ID the moment it's written down. Every task must name the ID it carries. Every acceptance criterion must point to a test — or to an openly tracked gap, never a silent one. On every push, a deterministic check assays the whole chain and strikes a hallmark: a trace-manifest, a plain file recording what is proven, what is honest debt, and what is missing. You read it through a viewer called Loupe, the way a jeweler reads a mark.

The refusal is the product. A silent gap — an acceptance criterion with no proof and no admission — is gilt. The assay finds base metal under the gild and fails the build. Not a warning; a refusal. Passing never means zero unfinished work. It means zero hidden unfinished work.

That distinction is the whole value to a CTO. "We have documentation" is a slide. "The documentation governs the work, and the build stops the moment it doesn't" is control.

It holds up cold. Handed one plain-English line in a PRD and nothing else, a zero-context agent carried it to done — minted the IDs, wrote the spec, the code, the named test — and the gate proved the result with no one coaching it. The mechanism does not have to trust the author. It trusts the check.

HomesFlow still runs it, now as the bundle instead of my hand-wired script: eighty-two durable IDs, the thread intact, zero silent gaps. The receipts are in the repo, and the migration that got it there is one reviewable pull request.

None of this is a new methodology. Spec Kit gives the workflow. SpecAssay adds the one thing a slide deck cannot: a hallmark the build refuses to forge.

The bundle: github.com/rdryfoos/specassay. The proof: github.com/rdryfoos/HomesFlow. Or talk it through first: book some time.