Research

The thesis, and how the lab works.

Why the work exists. Every project and every company the lab produces applies the position stated here.

Thesis

The research position.

The systems that run the physical world, from aircraft to industrial equipment to medical devices, need machine intelligence that is deterministic, auditable, and small enough for the hardware they already carry.

EntroMorphic builds that intelligence. We research methods that make on-device inference fit in kilobytes, run on microcontrollers and stand up to certification, and we prove each result by shipping it in a product.

Methodology

The research-to-spinout lifecycle.

Every project moves through four stages. A project that stalls at a stage is retired.

  1. Research

    We take a hard, specific problem in constrained or safety-critical compute and fund it through paid sponsorship or open-source work. The output is understanding, published openly where possible.

  2. Prototype

    Understanding becomes an artifact: a working system the lab can measure and stand behind. To advance, it must be small and deterministic.

  3. Validation

    The prototype is hardened against the real deployment surface and independently verified. This is where a research result becomes a defensible technical claim.

  4. Graduation

    A validated artifact with real demand becomes a standalone company with its own domain and team. The lab keeps provenance and a stake.

Validated artifacts

Results the lab stands behind.

Each result links to its full report in Writing as the report is published.

All writing
ARTIFACT · FLYNN

8,480-byte equipment-health monitor

Single-file C99, zero heap after init, bit-identical across silicon. On the NASA IMS bearing dataset: zero false positives over 120 hours and about 17 days of advance warning. Graduated as Flynn.

See Flynn
ARTIFACT · LCVDB

100% recall@1 at 6.5M vectors

A ternary vector database. Sharding broke the LSH candidate-recall ceiling, taking held-out recall from 92% to 100%.

Report forthcoming
ARTIFACT · GUANACO

Lossless 3.01× LUT compression

A Ternary-3 lookup table on LFM2.5-350M: 100% average similarity to the original, and faster: 293.8 tok/s against a 278.8 baseline.

Report forthcoming
ARTIFACT · SSTT

Zero-parameter classification

97.53% on MNIST and 86.54% on Fashion-MNIST in about a millisecond, using only ternary signatures and integer table lookups.

Report forthcoming
Provenance

Every claim gets a report.

Each result on this page is one the lab stands behind. The full reports are being written now.

See active work