How we think, build, validate, and learn.
A working record of real product questions, decisions, tests, and changes—published without pretending the work is more complete than it is.
Featured story
One question worth following closely
We Told a 70B LLM to Protect Your Wallet. It Started Refusing Legitimate Purchases.
A 50-scenario benchmark of decision models for fiduciary shopping agents — Clef-omni vs Jev vs Strands Decider, and why prompting alone backfires.
Read the storyLatest notes
Recent work from the Lab
Taming Generative Drift: How We Built a Drift-Resistant Asset Pipeline for Mythic IP
When exploring visual IP with generative AI, the true bottleneck isn't synthesis speed—it is decision pollution and identity drift. Here is how we adapted version-locking, three-tier directory isolation, and anchor density rules to turn volatile AI generation into a dependable production discipline.
Flat-4+ Architecture: The Nightmare of Scaling AI-Generated Code and Hardware Constraints
When heavily integrating AI agents, we hit a brutal wall of hallucination and GC micro-stutters. We needed a draconian, highly-deterministic 6-layer architecture to survive >100Hz hardware polling. Today, we open-source it.
ElderMate: Why Senior Mode and Family Mode Must Be Designed Separately
Older adults need low-friction actions and unmistakable confirmation. Families need fast understanding, clear priorities, and traceable follow-through. One interface cannot responsibly optimize for both.
ElderMate: Why Family Care Can’t Stop at a Data Dashboard
Families are rarely anxious because they lack another chart. They need to understand what changed, who knows, what happens next, and whether someone has taken responsibility.
The homepage shows a focused editorial selection, not the full archive.
View all 8 articlesAbout the Lab
Different formats. One editorial standard.
We publish what the evidence supports, distinguish observation from inference, and state what remains unresolved.
Rethink
Define the real problem and question why the existing answer is not enough.
Lab Notes
Trace real product work, including failed attempts, supplier feedback, and changed decisions.
Product Breakdown
Examine why a product was designed this way—from trade-offs to business and delivery logic.
Prototype
Show what each hardware, software, 3D, or UX prototype is trying to validate.
Launch
Explain why a product is ready now, who it serves, what was validated, and where its limits remain.