centramContribute

Open call · contribute

Six roles, and the first task in each is already written.

Day one is the same whichever one you pick: clone the repository, install the daemon, run its test suite, start a node against a throwaway data directory, and curl its status endpoint. Then you take the task at the end of your role's document.

The model runs on your own machine and learns from your own files. What peers are built to exchange is a compact weight-delta — a genome — not the files. No node has yet exchanged one with a node on another machine.

AGPL-3.0 Developer preview One operator

Before you spend time here

The acceptance rule was written before the run. The run did not clear it.

The hypothesis: pooling weight deltas across nodes produces a model better at cross-domain tasks than the best single specialist, and than a generalist trained on the union of the same data. The rule below is pre-registered, the artifact it decided is committed, and the verdict it returned is a negative.

Route-D validation run not confirmed

modelTinyLlama-1.1B-Chat-v1.0
seeds, clean5
merged student0.357
best single specialist0.438
union-trained generalist0.364
decide_verdictlost_to_single −0.080

verdict_underpowered: false — this is a clean negative, not a run that was too small to tell. In the same artifact, the loss-space result did hold (competence_gate_regret_mean = −0.630, amplified on all five seeds); it was the bridge from loss to task that failed. The single-node product does not depend on any of this: a node trains locally, gates every incoming genome against its own held-out evaluation, and by construction cannot be made worse by what it receives.

The full ledger is the honest-status document, which opens §Q with a correction retracting its own earlier headline. Read it before the README or this website.

Where the work is

Every command in these documents was run before it was written down.

One document per role, in the repository under docs/onboarding/ — on the working branch, not yet merged to main. The outputs quoted in them are real ones, and each document ends with a glossary, because the vocabulary is the real barrier here.

Developers

Backend and systems. The daemon is one Python process: HTTP API, a binary P2P protocol, a local chain, a ledger, an event bus.

Touches
The daemon: HTTP API, peer networking, CLI and economics
First task
Two filename bugs in cli/, one of them silent data loss

AI / ML engineers

Genomes, merge operators, the evolution engine, the module protocol and distributed inference.

Touches
The daemon’s model and module layers, and the experiment harness
First task
Measure the evaluation noise floor and publish it

Security

Venom: the dam, the T5 security tributary, antibody genomes, governance. Names in the protocol do not get renamed.

Touches
Security, peer networking and the threat model
First task
Repair the Test column of the threat model and add the missing regression tests

Researchers

Run experiments, read artifacts, decide what the evidence supports. Prose-only numbers do not count here.

Touches
Experiments, result reports and the English documentation
First task
Reproduce the committed negative, then close the artifact gap it depends on

Product & UI

Four separate Flutter applications — dashboard, chat, wallet, developer tools — and this website.

Touches
The apps, the wallet and this site
First task
Fix two stale empirical claims in www/ and surface the honest state in the dashboard

Community & ecosystem

Onboarding, documentation, licensing, the marketplace and how the project talks about itself in public.

Touches
Public documentation, this site, the marketplace and licences
First task
Make the front door work: a clone URL that 404s, a test count off by 4×, two dead doc links

How to reach us

Not open yet.

There is no public repository, issue tracker or discussion board today, so there is nowhere to point you. The application form is closed while the Participation Terms are rewritten. This page collects no personal data and runs no scripts of its own.

What happens next

Start small.

  1. We match your interest to open work.
  2. We agree on scope and time.
  3. You take one small, verifiable task.

Access is granted gradually, based on need and demonstrated work. We cannot promise when we will reply.