centram

Centram · Developer preview

Information Is Your Creation

Where it runs. What it learns. What it shares — keep those choices in your hands.

Explore Centram ↓See the evidence ↗
Centram emblem: a graphite hexagon with three detached triangular satellites
Explore the live merge demo

Live readout · genome_merge._slerp

‖a‖ parent A—
‖b‖ parent B—
ω angle between—
t interpolation—
‖slerp(a,b,t)‖—

Scrub t to explore the merge operator. These numbers describe interpolation, not measured learning.

Where the project stands, and how to read this page

Developer preview · source build only · no two-machine test yet

measured
A committed result file backs it.
implemented
The code runs. Nothing is claimed about whether it helps.
designed
Specified, not built.
unproven
The project's own record says open, or says negative.

The one section below without a marker makes no claim about Centram.

What it runs on, and what you need to build it
Current setup
Source build on macOS or Linux. No packaged installer exists yet.
Requirements
Python 3.11+, Git, Rust (cargo). Flutter as well, for the GUI.
Hardware
8 GB RAM and about 5 GB free disk for a small model; 16 GB+ recommended. A GPU is optional.
Docker
Not required. It is an optional deployment path for servers and VPSs.

First launch creates a local node identity and API token. Downloading a model is a separate step. Installation is in the docs, which are not published yet →

One address, kept to send release news and nothing else. Unsubscribe by asking.

The current deal

AI moved closer. Control moved farther away.

Much of today's AI runs in someone else's cloud. Your data — and the improvements you help create — stay behind the API.

You don't control the machine

The hardware, price and terms belong to the provider, and change without your permission.

The more it knows, the more you hand over

Personalisation usually means sending documents, messages and habits to someone else's server.

You improve it. Someone else keeps it

Every correction improves a model behind the API.

Own the process

implemented

Learn locally. Verify first. Then connect.

  1. Your machine trains a small add-on, not a whole new model

    That add-on is a LoRA: a small set of extra weights beside a model you never retrain. Raw data never leaves; what is shared is perturbed with statistical noise by default.

    a per-day privacy budget is applied by default

    measured: privacy and collaboration do not yet coexist — with the privacy budget on, the receiving gate takes almost nothing.

  2. Share the learning, not the source data

    What a machine learns becomes a compact update file — a genome. The receiver puts it through a staged gate: trust, rate limit, evaluation budget, replay, hop-chain integrity, structural and statistical checks.

    accept only if the merged model measures better than the one it replaces

    every exchange so far has run between processes on a single host.

  3. Prove the split before trusting the network

    The prototype divides a model's layers into shards held by separate processes.

    TinyLlama-1.1B   22 layers, split across two shards  ·  tokens identical  ·  4.13× slower
    SmolLM-135M   30 layers, split across two shards  ·  tokens identical  ·  1.58× slower

    both shards were subprocesses on one laptop. No cross-machine test has been done.

Proof, not promises measured

Two things have been measured. One has not.

Splitting one model's layers across separate processes reproduces the single-process output token for token, on both models tested. A peer's update measurably improves another peer's model on the same task.

The claim this project exists to prove — that merging what different machines learned beats the best single specialist — is not proven. It has been measured and it did not hold. That is stated here rather than left out, and the detail behind it is held back with the rest of the research until release.

What Centram OS is

designed

Operate the computer you already have through an intelligence of your own.

A design for letting that intelligence use an ordinary operating system safely — through Core's permissions, Module capability, and River's evidence. It is neither an AI chat app nor a bespoke kernel.

The way in is the PC you already use

Files, web, mail, photos, settings. An icon resolves through the public intent path, and behind it memory, permission and execution evidence accumulate as one trail.

Connected, but never handed over

External services get used. The boundary for memory, deletion, recovery and permission still sits with you.

The AI proposes. Core verifies

Model output is not authority. Every read, send or change is re-checked against the granted capability, and the stated purpose is recorded.

Philosophy

designed

Information begins with you.

The memory, output and capability produced in contexts you permitted should be yours to inspect, correct, delete, take with you, and license on your own terms. That is a commitment about what a system owes you, not a claim that you legally own all information.

  1. 01
    CreateThe flow of your work, study and creation is your own context.
  2. 02
    OwnIt becomes memory you can review, delete and export.
  3. 03
    GrowThe value that comes out of it returns to you.
  4. 04
    LilyA personal intelligence that holds that context.
  • store it
  • train on it
  • delete it
  • take it with you
  • set the terms

Future of work

unproven

Conventional AI helps you finish the work. Centram OS is designed to keep what the work taught.

What is demonstrated today is the floor beneath that: a local fine-tuner that cannot make your model worse. The compounding described here depends on pooled learning beating the best single specialist, and the one committed five-seed test at 1.1B went the other way.

Borrow someone else's intelligence

Existing AI services

  • Centred on one question and one answer at a time
  • Memory, history and permission tend to sit with the provider
  • Strong at short-form drafting, summarising and lookup

Own the intelligence yourself

Centram OS, as designed

  • Work memory, output and decision history stay on your machine
  • Lily keeps growing against your particular context
  • The value that comes out of your information returns to you
  1. 01
    Students and researchersPapers, notes and experiment history as your own research record.
  2. 02
    CreatorsVoice, source material and process as a lasting creative asset.
  3. 03
    Small teamsOperational knowledge as memory the team owns.

Modules

designed

Capability can be added. Authority cannot add itself.

A Module is a specialist capability. In time Personal Intelligence will choose them, but the right to read, write, send or act is confirmed by Core every time.

implemented

Personal Search

Searches your memory within the owner's scope.

implemented

Document Assistant

Summarises text you paste, through the local model broker. Pasted text, not documents.

implemented

Translation

Keeps the glossary and the context, and translates only the range you name.

implemented

Coding

Reads a repository and returns proposed changes, tests, diffs and an explanation. It proposes; it does not apply.

designed

OCR

Would turn pages and images into searchable text with provenance attached.

designed

Navigation

Would propose constrained routes from your schedule, maps and device signals.

designed

Physical AI Vision

Would turn sensor input into safe candidate actions.

designed

Federated Learning

Would exchange permitted evidence of improvement without collecting raw data.

Webπ / River

designed

Circulate verifiable capability, not data.

Webπ is not an existing standard; it is Centram's idea of what a network should be. River carries local events and security evidence as typed data, tagged with the authorising grant id. It never issues authority — Core checks again at use time.

shared streams are disabled today. River stays on the device, so nothing circulates yet.

  1. 01
    CreateAn event happens on your machine and is typed at the source.
  2. 02
    RememberIt enters memory you can review, correct and delete.
  3. 03
    GrowLearning is derived locally, as a compact update rather than a copy of the data.
  4. 04
    ShareThe update — never the source — may be offered to a peer. Disabled today.
  5. 05
    ReturnValue that comes back is attributed to the owner of the information.

Trust by design

designed

The person stays in charge.

Centram OS starts from the assumption that AI must not take choices away from people. The permission surface that would carry it is specified and partly built: the daemon refuses what it cannot verify.

  • listed and revocable
  • deny by default
  • auditable

the consent dialog is still in progress.

Concept screen · not a shipping dialog

Allow access to your family photos?

Allowing this does not send any photo anywhere.

scopethe 24 selected photos only
perioduntil this task finishes
grant idrecorded with every use

Where authority sits

  1. 01
    Personal IntelligenceConnects intent, memory and task context
  2. 02
    Centram CoreChecks identity, policy, capability and audit
  3. 03
    ModulesProvide specialist capability inside a limited scope
  4. 04
    RiverCarries events and evidence as typed data
  5. 05
    Host OSPerforms the actual effects on files, processes, network and devices
  6. 06
    WebπA future picture: trading verified capability without giving up ownership

Honest status measured

Where this actually stands.

A node runs, speaks to other nodes on a local network, and serves a local API. Distributed inference and peer-to-peer transfer have been measured and work. Everything above those two results is design: it is written down, much of it is implemented, and none of it has been run as a real network across separate machines.

The central research claim has not been proven. The full ledger — every item, every number, and the files behind them — is held back with the research until release.