Skip to main content

The Lab

Eight lanes. Every number measured, every gap shown.

This is our own R&D, shown honestly — because for this company, undeniable and honest are the same thing. Each lane is one card with its headline number and its tier; open it for the deep write-up, the evidence, and the open edges we keep visible.

  • byte-identicalself-host, gcc ≡ clang
  • 0.0%confabulation (from 87%)
  • 151.7 tok/sQwen in QUANTA == llama.cpp
  • 441×leak killed at compile time
  • 4 of 4compounding axes, at once

The showcase

The eight capability lanes.

🟢 shipped and measured · 🔬 research, measured and partial · 🔭 vision, designed and honest. The bigger claim goes up a tier, never down.

QUANTA — The Forge

Shipped & measured

From the silicon up: a language that proves itself

byte-identicalself-host, diverse-double-compiled (gcc ≡ clang)

Compile it twice, with two different compilers, and get identical bytes. The agreement is the proof there is no backdoor.

Read the write-up

The Calculated Mind

Research

The calculated mind: it reasons or refuses — and 'cannot bluff' is a property under test

87% → 0.0%confabulation, type-level

Watch it refuse to make something up — then cite Canberra from a store that fits on a phone. The refusal is the costly signal.

Read the write-up

Wave-Ternary Models

Research

A mind of waves, weighed in thirds, that remembers

cos 0.7725-fact recall, drift-bounded

Inject a fact mid-stream, ask 25 facts later — it still answers, on a model that does no multiplies.

Read the write-up

Frontier GPU Inference + Training

Shipped & measured

A capable model on hardware you own: inference and training, our own stack

151.7 tok/sQwen in QUANTA == llama.cpp

21 ms × 377 MiB × one 3090 — flat to 128k, pure stack, and the gate is never disabled.

Read the write-up

NovaTerra — The World

Upcoming

NovaTerra: two AI beings that are cryptographically yours

~75%wired (donor bodies)

A world you own — a Friend and a digital Twin, on your own chain and mesh.

Read the write-up

Certifiable by Construction

Shipped & measured

Certifiable by construction: a memory leak is now a compile error

441×arena bug: 1.5 GB → 3.4 MB

The leak doesn't get caught at runtime — it doesn't compile. The rejection is the audit trail.

Read the write-up

The Agentic Operating Model

Shipped & measured

The agentic operating model: propose broadly, dispose certainly

13 / 48hversions shipped

The company is run the way the product is built: propose broadly, dispose certainly.

Read the write-up

The Gate

Research

The gate: tested on every compile, not yet proven sound

tested, not proven soundthe gate's own status — our named core gap (2026-07-11)

Long-context · quantized · any-chip · gate-checked — we hold all four at once. We have not benchmarked every other stack, so we do not claim to be the only one that does.

Read the write-up

The Lab Notebook

The dated archive: what we built, often one to two years early.

These lanes are the destinations. The Notebook is the paper trail — real R&D from an AI-written OS in 2024 to a non-transformer wave model in 2026, each entry dated, tiered, and honest about its failures. Filter it, or read it as a timeline. Some public, some protected.

Open the Lab Notebook

The heart

AI is good. It empowers humans. It leads to an abundant world.

The showcase is the evidence for the belief, not the other way round. We earned the right to be optimists by proving it — tiered, measured, and open about the edges.

How we work

    We use cookies.