Skip to content

cognet_integration_plan.md

Latest commit

 

History

History
91 lines (71 loc) · 3.62 KB

cognet_integration_plan.md

File metadata and controls

91 lines (71 loc) · 3.62 KB
Raw

CogNet Integration Plan

Status: Implemented — corpus generator, training pipeline, and 1B model support are all functional. See docs/cognet-integration.md for the user-facing guide. Last updated: 2026-06-20 Owner: Philippe-Antoine Robert (@AFKmoney)

This document explains how the CogNet ↔ AICL integration works. The corpus generator produces spec→code pairs (using AX sub-language behaviors), and CogNet is fine-tuned on them via the cloud_train.py pipeline.

Goal

Bidirectional bridge between AICL specifications and the CogNet cognitive engine (a non-transformer language model with hierarchical memory and cognitive routing). The bridge lets CogNet reason over AICL programs as first-class cognitive representations, and lets AICL programs invoke CogNet as a reasoning backend.

The full design is specced in AICL itself in examples/86_cognet_aicl_bridge.aicl through examples/91_cognet_autonomous_deployment.aicl. Those six example programs collectively form the integration specification.

What is in place today

python/src/aicl/cognet/
├── __init__.py        Public API surface (currently empty — see below)
└── README.md          This file's shorter sibling, viewable from the package

The cognet/ subpackage is reserved. Importing aicl.cognet today succeeds but exposes no symbols. This lets us:

  1. Add the import path in __init__.py without breaking existing code.
  2. Land type stubs and protocol definitions incrementally.
  3. Keep the public API stable across the integration rollout.

What is NOT in place yet

  • The actual bridge implementation (translation layer between AICL AST and CogNet cognitive graph).
  • CogNet model loading / inference glue.
  • Round-trip fidelity tests.
  • A cognet subcommand on the aicl CLI.

Planned rollout

Phase 1 — Bridge protocol (target: v2.2)

  • Implement aicl.cognet.bridge.AICLCognetBridge with two methods:
    • spec_to_graph(source: str) -> CognetGraph
    • graph_to_spec(graph: CognetGraph) -> str
  • Property-based round-trip test: graph_to_spec(spec_to_graph(p)) ~ p
  • Add aicl cognet translate <file> subcommand

Phase 2 — CogNet as reasoning backend (target: v2.3)

  • New Native: section handler in the compiler that delegates to CogNet for runtime reasoning (cognitive routing decisions).
  • Wire runtime.py Risk → Recovery selection through CogNet when a Native: CogNet section is present.

Phase 3 — Autonomous CogNet self-evolution (target: v2.4)

  • Implement the spec from example 87 (87_cognet_self_evolution.aicl).
  • The autonomous compilation loop (aicl evolve) gains a CogNet mode where the LLM-backed ai_generator.py is replaced by a local CogNet model for fully offline self-modification.

Phase 4 — Full AGI loop (target: v3.0)

  • Bidirectional AICL ↔ CogNet loop where AICL specs drive CogNet training data, and CogNet reasoning drives AICL spec evolution.
  • Examples 89–91 (training pipeline, evaluation, autonomous deployment) become executable end-to-end.

How to integrate when ready

When the CogNet runtime is published:

  1. Add cognet-runtime as an optional dependency in pyproject.toml: 
    [project.optional-dependencies]
cognet = ["cognet-runtime>=1.0"]
  2. Implement bridge.py in this folder.
  3. Re-export public symbols from __init__.py.
  4. Add the cognet subcommand in cli.py.
  5. Land property-based round-trip tests in tests/test_cognet_bridge.py.

Until then, this folder is a deliberate placeholder. Do not remove it.