AI SKILL LAB
LABRUStart
BUILD STORY / OPEN PROVENANCE

This site is
our own case study.

Not “made with AI” as a sticker. Below is the actual iteration history: what AI accelerated, what stayed a human decision, which failures were found and which controls appeared because of them.

Open build log ↓Current manifest ↗
AI SKILL LAB / BUILD PROVENANCE● EVIDENCE-SCOPED

01ai_roleACCELERATE

02human_roleDECIDE

03machine_roleVERIFY

04claimsBOUNDED

05historyPORTABLE

06releaseMANIFESTED

IDEAITERATEBREAKPROVE
Build timeline

Not one prompt.
A system of iterations.

This is a milestone snapshot, not the full changelog. Exact current release identity always lives in /_release.json.

R8 · BASELINE

Proof + parent decision layer

The first portable source authority and an honest proof layer without invented testimonials or guarantees.

R24 · COMMERCIAL

Commercial parity

Pricing, packages and claims reconciled into one commercial truth and protected by a parity gate.

R31 · RUNTIME

Behavior, not markup

A real static matcher syntax bug led to runtime smoke tests, no-JS fallback and a verifiable Start flow.

R38 · GOVERNANCE

Security + structured data

Hashed CSP, privacy contracts, metadata/structured-data integrity and semantic gates became release requirements.

R49 · RELEASE

Deterministic builder

Artifacts, manifests and wrappers became byte-for-byte reproducible instead of manually packaged.

R60 · EXPERIENCE

Site becomes the demo

Proof Lab, Project Studio, Pilot Simulator, Brief Compiler, Challenge and Skill Graph turned the site into an interactive proof surface.

R65 · ATTENTION

Product-scene hero

The hero began showing ACTIVE WORKFLOW → DEFINE → BUILD → VERIFY → SHIP instead of an abstract AI illustration.

R66 · PROVENANCE

Build process becomes public

Build history, failures, toolchain and responsibility boundaries become part of the product itself.

R68 · TRANSFER

Performance becomes a contract

Worst-case first view stays around 14 KB Brotli: HTML + shared CSS + LAB runtime. The budget is verified on every release.

AI did

Searched,
generated,
accelerated.

  • researched options and current policies
  • proposed architecture, copy and code
  • searched for source/static divergence
  • built testing and release scaffolding
Human owned

Selected,
bounded,
authorized.

  • defined product intent and commercial truth
  • made visual and semantic decisions
  • set risk / claims / youth-safety boundaries
  • kept final authority over production release
Actual toolchain

Only what
was actually used.

Planned tools are not added to the case study retroactively. A new agent or service becomes a separate verifiable iteration.

01

CHATGPT

Product reasoning, research synthesis, code drafting, QA design and iterative development.

02

WEB RESEARCH

Current policy checks and comparison of modern product/design patterns.

03

LOCAL TEST STACK

Python + Node gates, runtime smoke, CSP hashes, metadata parity and release manifests.

04

GIT / BUNDLES

Exact history, clean checkpoints and portable project recovery.

05

VERCEL

Deployment target. Preview, production and readback are separate release gates.

06

HUMAN REVIEW

Intent, taste, factual claims, risk boundaries and the final decision to release.

Failure log

The system grew most
when something broke.

A useful failure changes the system so the same class cannot pass unnoticed again.

FAILURE → CONTROL

STATIC JS BROKE

The deployable static matcher contained a syntax error.

→ Added syntax gates plus behavioral runtime smoke.

FAILURE → CONTROL

METADATA DRIFT

RU/EN and source/static surfaces diverged in social and SEO metadata.

→ Added deterministic metadata, hreflang and structured-data gates.

FAILURE → CONTROL

WRAPPER CORRUPTED

One release wrapper was found to be malformed.

→ The builder now checks archive SHA, manifest SHA and byte-exact reconstruction.

FAILURE → CONTROL

SANDBOX DISAPPEARED

A working directory disappeared during an iteration.

→ Recovered the exact project from a Git bundle.

FAILURE → CONTROL

DEPLOYMENT BLOCKED

Vercel quota/provider blockers prevented release.

→ Source, build, deployment and readback stayed separate states; a blocker was never reported as PASS.

What this proves

Not “AI built a website.”
We can govern an AI build.

The evidence is not the number of generated lines. It is the ability to evolve, verify, recover, constrain and release a product without losing commercial or semantic truth.

That is the same discipline behind our training and AI projects: Define → Build → Verify → Ship.

Open Proof Lab →Project Studio →