Development · Reviewed 6 October 2026
ALPHA is
not finished.
It is an active research and engineering project. This site shows its intended companion experience alongside a development foundation that already exists.
The difficult parts should be visible. Implemented systems, incomplete validation and longer-term ideas are described separately below.
No public release announced
01 — Implemented
A working
local foundation.
These systems have implementation and test evidence in the repository. Their validation applies to the recorded environments and tasks, rather than every possible use.
- Native desktop applications
A SwiftUI application on macOS and a Windows desktop build, with local engine communication and controls for memory, actions and learning.
Repository README · F04 Windows validation- Memory with provenance
A local SQLite store keeps current values, revisions and supporting evidence. Observation and learning use explicit grants and storage limits.
F02 · Memory, permissions and storage- Bounded workflow learning
File-organization routines can be learned from permitted moves, tested before activation and reversed. Corrections become evidence; they do not silently authorize a change.
F03 · Developmental loop validation- A certified local foundation
Dated records document a bound Ministral 3 8B foundation on Windows and a certified 4-bit serving variant on the Mac. This is a local open-weight model, not an ALPHA model pretrained from scratch.
F07 · 26–27 September / D48 Mac serving · 28 September 2026- Measured skills and durable work
Typed skills, concepts and predictive world models have been tested on simulated workspace tasks. User-familiarity and durable-task systems have local test evidence; their phase closure still depends on the checks named in those records.
F09 · 29 September / F10 · 30 September 2026
02 — In progress
Built machinery.
Unfinished validation.
Passing a component test does not prove the complete experience. These are specific boundaries the project has not yet crossed in the reviewed evidence.
- Continual neural development
Growth jobs, retention gates, promotion and rollback are implemented and tested on Windows. The 1 October record reports synthetic-artifact experiments; the real-model completion runs, final full-suite rerun, Mac work and CI remain pending in that record.
F11 Windows evidence · 1 October 2026- Learning unfamiliar applications
The Mac baseline includes measured browser interaction and a certified demonstrated workflow on test material. Its local outcome selector did not meet the qualification threshold. Unseen-application evaluation, visual fallback, Windows work and CI remain open.
F12 Mac completion evidence · 2–4 October 2026- Companion integration
The website’s live character is an isolated motion study. Connecting its presence and expressions to real engine state remains part of the intended interface, rather than a capability shown here.
Website implementation · Character and engine are separate
03 — Exploring
The larger direction.
A companion that carries useful context across devices. Broader application competence. Coordinated long-term work and selective specialization as experience accumulates.
The implementation plan describes these ambitions. A plan is not evidence that they work, and the completed personal intelligence described by that plan is not available today.
Reading the status
Evidence has a date
and a scope.
This review follows the dated validation records where the repository’s introductory text is older. It reports those records; it does not independently repeat the product’s model or application evaluations.
The character renders and film are conceptual. No release date, progress percentage or general intelligence claim is attached to them.
Read the source notes