Proof-gated evolution for AI agents

Agents mutate.
Evidence decides.

Variaxiom separates an agent's ability to propose a change from the authority to make that change inheritable. Mutation is easy. Governed inheritance is the hard part.

REJECTED

Functionally successful, constitutionally invalid

The candidate passes its functional checks but asks for network.unrestricted without an external grant.

PROMOTED

Same implementation, bounded authority

The candidate remains inside its existing capability lease, carries independent evidence, and retains a rollback target.

Architecture

Governed evolution, not in-place self-rewriting.

Variaxiom is a protected process for creating, testing, selecting, inheriting, and retiring disposable agent phenotypes.

Mutable outer plane

Models, skills, tools, workflows, memory strategies, and harness variants may become candidates.

Protected evidence plane

Independent evaluators, hidden tests, lineage, budgets, and artifact identity constrain promotion.

Human sovereignty

Greater capability never silently grants greater authority. Permissions remain delegated, scoped, and revocable.

Run locally

Clone, run, and verify. No API key.

git clone https://github.com/artemiosu/variaxiom.git
cd variaxiom
bash scripts/demo.sh
bash scripts/verify.sh

The pre-alpha seed is intentionally narrow. It proves an evidence and authority boundary; it does not claim production autonomy or inevitable recursive improvement.