▸~/lab

The lab.

Less polish, more curiosity. Questions, what was built to answer them, and what happened — including the parts that aren’t finished.

6 ENTRIES · 4 SHIPPED · 2 OPEN

#001

SHIPPED

WEBXR / UNITY / NODE.JS

Patching seven builds without rebuilding one

QUESTION: Can seven compiled Unity WebGL builds be fixed safely without rebuilding any of them?
EXPERIMENT: A Node.js script that patches the minified framework inside each brotli archive — guarded by exact-match counts, a syntax gate, backups and a read-back comparison.
RESULT: All seven builds patched idempotently; re-running reports “already patched”. Originals kept as backups.
LEARNING: When you can’t rebuild, make the patch prove itself before you trust it.

▸ TRY IT

0mspage loads; navigator.xr missing → fetch webxr-polyfill
900msUnity boots first → _InitXRSharedArray()
901msModule.WebXR.onUnityLoaded is not a function ✗
901msboot halts — no error on screen, just a gallery that never starts
1400mspolyfill arrives… nothing left to notify

RESULT · STALLED

CONNECTED TO

#002

SHIPPED

AI AGENTS / PLAYWRIGHT / TESTING

Build the adversary first

QUESTION: What breaks a form-filling agent?
EXPERIMENT: A fictional career site where every hard pattern lives on its own route: custom JS dropdowns, async autocomplete, conditional questions, scroll-to-render fields, server-side validation, SPA navigation, login walls, CAPTCHAs, cookie banners.
RESULT: Every failure mode can be reproduced and tested in isolation by the end-to-end suite.
LEARNING: An agent is only as good as the worst page it has been tested against.

CONNECTED TO

#003

SHIPPED

SEO / RENDERING / AUDIT

What does the crawler actually see?

QUESTION: Why doesn’t Artabia’s blog appear in search?
EXPERIMENT: Fetched each route’s pre-JavaScript HTML and sorted pages into three rendering buckets.
RESULT: Only the homepage rendered. Content pages served a fallback and canonicalised to the homepage; legal pages were empty shells.
LEARNING: Inspect a system from its consumer’s side. Here, the consumer was a crawler.

CONNECTED TO

#004

RUNNING

RAG / SEARCH / INTERACTIVE

Watch the AI decide

QUESTION: Can you see why an AI answered what it answered — before a language model is involved?
EXPERIMENT: Type a question and watch the BM25 retriever score every chunk of Prince’s knowledge base.
RESULT: The same scores drive Ask Prince’s refusal gate: no strong match, no answer.
LEARNING: Explainability is cheapest before the model, not after it.

▸ TRY IT

QUESTION

skill:3d13.20
career:avrlabs10.69
profile:summary9.10
mission:artabia8.20
skill:xr7.86
profile:story7.34
mission:arviewz7.18
profile:stack6.45

│ = refusal threshold (2.2)

GATE OPEN → the model sees the top 5 chunks

CONNECTED TO

#005

SHIPPED

UX / PERFORMANCE

A boot sequence that never blocks

QUESTION: Can a boot animation be pure decoration without ever being a delay?
EXPERIMENT: The boot overlay sits on top of fully server-rendered HTML. It is skippable, shortens to about a second for returning visitors and never runs on deep links.
RESULT: Content, SEO and screen readers never wait for it.
LEARNING: Delight is fine. Blocking is not.

▸ TRY IT

CONNECTED TO

MISSION 000PRINCE.EXE

#006

VERIFY

COMPUTER VISION / PYTHON / OPENCV

Finding the painting in the photo

QUESTION: Can a phone photo of an artwork be turned into a clean, upright image of just the artwork?VERIFY — Unverified — only in: Artwork detection service (Python)
EXPERIMENT: A Flask + OpenCV service: EXIF orientation correction, orientation scoring from Sobel gradients, text-region masking, and multi-strategy Canny/contour detection.VERIFY — Unverified — only in: Artwork detection service (Python)
RESULT: [RESULT OF THE ARTWORK DETECTION SERVICE REQUIRED]
LEARNING: [WHAT YOU LEARNED REQUIRED]

CONNECTED TO

MISSION 001Artabia