alethic-kernel 0.2.0

A governed cognition framework for AI systems

Alethic

A governed cognition framework for AI systems.

Your model proposes. Alethic decides.

Alethic is a domain-agnostic governance layer for AI agents. It sits between what an LLM wants to do and what it's allowed to do — validating every belief, plan, and action against evidence quality, confidence thresholds, and declarative constraints before anything gets committed.

The core insight: architectural governance — not model scale — is the primary bottleneck for trustworthy modular AI. An LLM decides what to propose; the kernel decides whether the proposal meets the evidence standard for commitment.

How It Works

Seven semantic slots hold all state. Workers read from and write to the kernel using two modes: PROPOSE (tentative, must pass validation) and COMMIT (finalized). Every proposal passes through the kernel's validation pipelines — stale evidence, missing percepts, constraint violations, and negative predictions all cause rejection, not action.

Quick Start

pip install alethic-kernel

Use as a library

from alethic_kernel import AlethicClient

client = AlethicClient(mode="local")
result = client.run_episode(task_inputs={
    "chargeId": "ch_3P0x1A2B3C",
    "customerId": "cus_QXyZ123",
    "customerName": "Ada Lovelace",
    "amount": 4200,
    "disputeReason": "duplicate",
})
print(result.metrics)  # {'task_success': 1.0, 'unsafe_action': 0.0, ...}

Run the benchmark

# All agents except LLM (no local model needed)
python -m alethic_kernel.run --no-llm

# Full benchmark (6 tasks, 50 seeds, 4 agents = 1200 episodes)
python -m alethic_kernel.run

Start the API server

pip install "alethic-kernel[api]"
alethic serve --port 8000

The HTTP API has no authentication and is not production-ready. Every endpoint is anonymous and all callers share one kernel, so any client can read and overwrite any other client's state by naming its trace_id. Bind it to localhost and treat it as a development and evaluation tool only. The library itself does not carry this limitation.

Benchmark Results

To validate the framework, we ran 1,200 episodes (6 tasks, 50 seeds, 4 agents) across Stripe refund tasks with controlled perturbations:

Agent	Task Success	Unsafe Actions	Unsupported Beliefs	Traceability
string_glue	61.3%	38.7%	26.0%	0.10
json_glue	57.0%	43.0%	31.0%	0.30
alethic	100%	0%	0%	1.00
llm_bk	99.0%	0%	0%	1.00

The kernel-backed agents (alethic and llm_bk) achieve zero unsafe actions across all perturbation scenarios. Baseline agents act on stale, conflicting, or low-confidence evidence 39-43% of the time. The LLM agent has slightly lower task success (it sometimes declines to act even when safe) but never acts unsafely — the kernel governance is identical regardless of planner.

Documentation

• Architecture — Blackboard kernel design, semantic slots, validation pipelines
• API Reference — Every public class and method with signatures and return codes
• Workers — Worker protocol, built-in workers, writing custom workers
• HTTP API — REST endpoints, AlethicClient SDK, OpenTelemetry
• Benchmark — Tasks, perturbations, metrics, CLI reference
• Deployment — Docker, environment variables, store selection, testing

Project Structure

src/alethic_kernel/     Installable Python package
  kernel.py             Blackboard kernel — central orchestrator
  schema.py             Record, Provenance, Slot, WriteMode
  validators.py         Evidence and symbolic validation
  store.py              In-memory store
  sqlite_store.py       SQLite-backed persistent store
  orchestrator.py       Worker round-robin loop
  sim_worker.py         Rule-based forward simulator
  adaptive_worker.py    Learns constraints from failure patterns
  session.py            Multi-episode session management
  client.py             AlethicClient (local + HTTP modes)
  api/                  FastAPI server
  agents/               Four reference agents
  eval/                 Evaluation harness and metrics
  llm/                  OpenAI-compatible planner
  tools/                Simulated tools with perturbations
  tasks/                6 Stripe refund task definitions (YAML)
examples/               Multi-episode demos
tests/                  Test suite
docs/                   Architecture, API, and deployment guides

Research Paper

The full academic treatment is in From Fragile Glue to Governed Cognition — a controlled study of blackboard kernels for modular AI systems. Paper artifact repository: governed-cognition.

License

MIT