Evolving AI — Governed, Evolving AI Runtime (v0 Prototype)

A governed, evolving AI runtime where nothing evolves in place. Every change is a versioned amendment that must be evaluated and approved before promotion.

Architecture

Four distinct roles:

Role	Responsibility	Power
Operator	Executes user tasks with approved tools/memory	Productive
Steward	Analyzes failures/telemetry, generates amendment proposals	Creative
Evaluator	Replays candidate runtimes, detects regressions	Adversarial
Governor	Enforces constitution, gates promotions, rollback	Authoritative

The central constraint: the Steward can propose changes, but cannot unilaterally enact them.

Amendment lifecycle:

PROPOSED → SANDBOX → EVALUATED → REVIEW → APPROVED/REJECTED → PROMOTED

Layout

app/
├── runtime/        # (reserved) manifest/executor/versioning
├── operator/       # executes tasks, cannot self-modify
├── steward/        # failure analysis → amendment proposals
├── evaluation/     # replay suites, metrics, regression detection
├── governance/     # models, gates, Governor, runtime registry
├── memory/         # episodic/semantic stores, governed lessons
├── tools/          # (reserved) tool implementations
└── api/            # FastAPI app + dashboard
constitution/       # constitution.yaml + model
evaluations/        # suites + datasets
prompts/            # versioned prompt artifacts
tests/              # unit / integration / governance / replay

Quickstart

pip install "fastapi>=0.100" "pydantic>=2.5" uvicorn pytest
uvicorn app.api.main:app --reload

Dashboard: http://localhost:8000/dashboard

Or with Docker:

docker compose up --build

Endpoints

• POST /execute — execute a task with the current runtime
• GET /runtime/current — inspect the current immutable runtime
• GET /runtime/list — list all runtime versions
• POST /runtime/rollback/{version} — rollback
• POST /amendment/propose — propose an amendment (prompt/memory only in v0)
• GET /amendments / GET /amendment/{id} — list/inspect amendments
• POST /evaluation/run — run a replay suite against a runtime
• POST /governance/evaluate/{amendment_id} — evaluate → attach evidence → REVIEW
• POST /governance/approve/{amendment_id} — approve + promote (human approval)
• POST /governance/reject/{amendment_id} — reject
• GET /governance/audit — full audit trail
• POST /steward/analyze — run the steward loop (never promotes)
• GET /telemetry/failures / GET /telemetry/runs — inspect telemetry
• POST /memory/lesson + lifecycle endpoints — governed lessons

Demo flow (definition of done)

# 1. Operator repeatedly fails a task class (seed failures)
curl -X POST "localhost:8000/steward/analyze"

# 2. Steward proposes amendments; evaluate one
curl -X POST "localhost:8000/governance/evaluate/<amendment_id>?suite_id=core"

# 3. Human approves → new immutable runtime created
curl -X POST "localhost:8000/governance/approve/<amendment_id>?reviewer=me"

# 4. Old runtime remains for rollback
curl -X POST "localhost:8000/runtime/rollback/v0"

# 5. Audit trail
curl "localhost:8000/governance/audit"

Tests

python -m pytest tests/ -v

Covers: runtime immutability, amendment lifecycle, governance gates, approval/rejection, rollback, evaluation reproducibility, memory lifecycle, audit logging, unauthorized self-modification attempts, and the full end-to-end evolution pipeline.