CORE

Executable constitutional governance for AI-assisted software development.

AI can write code. It cannot be trusted to build complete, correct software autonomously.

CORE surrounds AI with a deterministic governance system. It does not make AI perfect — it makes AI mistakes detectable, traceable, and fixable in a controlled loop.

LLMs operate inside CORE. Never above it.

What CORE Does

CORE is a governance runtime. When an AI agent proposes a change, CORE:

Validates it against constitutional rules before execution
Blocks what a blocking rule forbids — the change halts before it is applied; reporting and advisory rules surface findings and let it continue
Records decisions and the finding → proposal → approval → execution chain (recording is best-effort: a database failure does not stop execution)
Remediates automatically where rules permit; anything outside the narrow safe-auto-approval envelope waits for a human

The aim: AI-assisted development that is auditable and governed. It is not yet proven safe to leave unattended — CORE's own production-readiness verdict is NOT ATTESTED, and unattended reliability (gate G4) is not demonstrated.

The goal is that the governor's job shrinks to two things: check the dashboard (core-admin runtime dashboard), and write constitutional intent. That is the design target, not today's measured state.

What CORE Is Not

❌ Not a generic LLM wrapper
❌ Not an autonomous coding agent you deploy and forget
❌ Not a CI system
❌ Not a replacement for human judgment

CORE is a controlled production pipeline whose purpose is to converge toward working, constitutionally compliant software.

Choose your path

You want to…	Start here	You need
Audit code against rules on your machine	Start a governed project	Python 3.12+
Add ready-made rules to a governed repository	Add a rule pack	Python 3.12+
Block pull requests that break the rules	Audit in CI (GitHub) · pre-commit or GitLab	A repository with a `.intent/`
Bring an existing repository under governance, with rules fitted to it	Govern your own repository	`core-cli` and a running CORE
Run the whole loop — audit, propose, approve, fix, verify — and see CORE govern itself	Run the full runtime	Docker, Poetry; an LLM only for code generation
Let your AI assistant ask CORE and propose changes	Connect your AI assistant	Python 3.12+; a running CORE to submit changes
Check compliance documents against a regulation	Run a GRC gap analysis	The full runtime
Upgrade an existing installation	Upgrade a CORE database	—
Build on CORE's API	API contract	—

CORE ships two command-line tools. core-admin (pip install core-runtime) runs CORE on your machine and operates a CORE installation. core (pip install core-cli) is a small client for a running CORE: proposals, the remediation lane, onboarding. See which one you need.

Know this first

Not production-attested. CORE's own production-readiness verdict is NOT ATTESTED. Do not leave it unattended on code that matters.
Detects, does not prevent, by default. On a single account (solo mode) CORE sees every change to your law and code, but cannot stop you or your assistant from making one. Separate OS accounts per role (solo-separated) let it prevent them; CORE's own server runs that way. CORE does not yet check that the machine matches the declared mode.
No API authentication. The API binds to 127.0.0.1 and trusts every caller. Keep it there.
An LLM is optional. Auditing, rule packs, CI gating and the deterministic fixes need none. Code generation, rule induction and judged GRC verdicts do; CORE names no vendor and works with a local model server or an external API.
Open and complete. Everything on this site is in the open-source product (MIT), free indefinitely. Commercial offerings add usability and scale around it, never a missing capability (ADR-084). Open CORE is the command line, the API and the MCP server; any web interface belongs to core-platform, a separate product.

Understand it

What CORE does — what CORE is for, in plain language
Decisions — every accepted decision and what enforces it