Independent systems laboratory / Since 1991

HeadSoft

We build compact, inspectable systems: a self-contained programming runtime and a governed intelligence substrate designed to make model-assisted work durable, auditable, and measurable.

Carnivore / language runtime Amoeba / AI architecture research

Work in public

Two systems, one engineering standard

Both projects emphasize explicit behavior, low operational overhead, and inspectable mechanisms. One is a language runtime you can use today; the other is an active research program with measured evidence and strict claim boundaries.

PROJECT 01

Carnivore

A lightweight, zero-dependency, batteries-included BASIC runtime. Carnivore pairs a high-level interface with direct mode, devices, graphics, objects, namespaces, and low-level systems access.

Open the browser beta

Carnivore BASIC running in the public browser beta.
The current WebAssembly browser host, captured from the live local build.
PROJECT 02

Amoeba Core v2

An experimental intelligence substrate around language models. Amoeba externalizes evidence, provenance, selection, verification, memory, and cost policy so behavior changes can be inspected and tested without changing model weights.

Review the research and evidence

Amoeba equal-evidence experiment result: raw 18 of 48, flat 14 of 48, governed 48 of 48.
Primary controlled result from the first public working paper. Boundaries and counterevidence are reported with the finding.

Latest publication

Equal evidence, different outcomes

The first Amoeba paper asks a narrow causal question: when a small model, acquired evidence, and work are held fixed, can governed deterministic processing outperform flat evidence injection?

48 casesFrozen process-isolated target family
34-0-14Governed wins, losses, and ties versus flat evidence
219 checksMachine-verifiable paper evidence receipt