Build records
We are our own first client.
This website — and its sibling product site — was built under the same enforced honesty discipline we hold client work to. That is not a claim we ask you to take on faith. It is git history. Below are the receipts.
These are records of how we govern our own build work. They are not client case studies, and they are not a substitute for the controls we put around a client system.
What does “we are our own first client” mean?
It means this website was built by the same forward-deployed AI harness we build for clients, under the same enforced honesty discipline. The banned-claims lint that blocks a fabricated claim from shipping here is the same class of enforcement surface we build into client AI systems.
The harness
What actually made this build reliable
Not a bigger model. A harness: specialised agents in parallel, an adversarial reviewer, and rules compiled into checks that fail the build. The same disciplines we build into a client's agents and pipelines.
Orchestration
Parallel specialised lanes
Research, strategy, design, build, and review run as separate agents against a shared contract — not one model doing everything in one thread.
Verification
Adversarial review
An independent reviewer pressure-tests the plan and the code before it ships, on the assumption the author missed something — because authors do.
Enforcement
Surfaces, not guidelines
The rules that matter are compiled into checks that fail the build — a banned-claims lint, eval gates — so they survive the next mechanical rewrite.
Audit
The eyes text linting cannot replace
A lint reads code, not pictures. A human-and-model image audit looks inside every asset, because fabricated data hides in a JPEG where no text check can see it.
The receipts
Genuine run-records from this build
Every figure below is verified against the build's own artifacts. Nothing is rounded up — that would rather defeat the point.
Nothing unsupported reaches a page — enforced by the build
A CI-style lint scans every page and fails the build on any claim we cannot substantiate. It exists because AI drafting reliably produces confident, unsupportable statements: on the sibling build’s first run it stopped 75 of them that a manual sweep had missed — prior-employer engagements phrased as if they were our own clients, invented customer counts and success rates, and a star rating in a dead, unshipped file. None of it reached a published page, and the lint still runs on every build.(This very page proves the point: the lint just blocked those real ex-client names from shipping here, even while describing them — so you are reading the redacted version.)
This site was built under that lint from the first commit. Its current run, all 11 rules:
$ npm run lint:claims
✓ claims lint: clean (0 findings)
Fabricated data was hiding inside an image
A lint reads code, not pictures. The image audit on the sibling build found three product screenshots carrying fabricated customer data inside the PNG — including a name the text purge had already removed from the copy, which survived as pixels where no text check could ever see it. All three were pulled.
This site takes the lesson to its conclusion: it ships with zero customer-data images. Structure comes from type and hairline rules, so there is no picture for a false claim to hide in.
A REVISE verdict caught a stale-base error before code shipped
Before any code was cut, an independent review returned a REVISE verdict with one HIGH finding: the work was about to be dispatched from a local branch 116 commits behind main, where files that were real and merged did not exist on disk — the same class of stale-base error that had cost nearly two hours of wasted work weeks earlier.
The fix — verify against origin/main, not the local checkout — was written into the plan before the build began, not discovered halfway through it.
The purge is in the commit history
The rebuild’s own commit sequence on this repo:
purge debug/clerk routes + fabrication legacy pages
port design system v2 + claims lint + honest schema
build Tier-1 pages + wired lead capture
purge remaining gradient residue (15 dead components)
apply operator honesty rulings sitewide
Behind it: 16 work-artifact reports across research, strategy, design, build, and review lanes — the trace of the harness, kept.
The proof is the offer
The discipline we sell is the discipline this site was built under.
The lint that blocks a fabricated claim from shipping on our site is the same class of enforcement surface we build into your AI systems. The adversarial review that caught our stale-base error is how we keep your agents from shipping a confident wrong answer. Harness engineering, eval gates, guardrail surfaces — this page is what they look like when they are pointed at ourselves.
Work with us
Have AI that needs to survive contact with production?
The same harness, review discipline, and enforcement surfaces — pointed at your systems instead of our website.