OmegA Self-Description Contract
This document is the canonical self-description of OmegA. It is the authoritative source for what OmegA says it is — not marketing, not aspiration. Every claim in this document can be traced to evidence in the repository.
1. What OmegA Is
OmegA is a formally specified, four-layer concentric architecture for building AI agents that maintain sovereign identity, persistent memory, cognitive consistency, and governed behavior across model providers and context resets.
The four layers, from outermost to innermost:
| Layer | Name | Core function |
|---|---|---|
| 1 | AEGIS | Model-agnostic governance shell — enforces policy at the API boundary using compiled run envelopes, not prompts |
| 2 | AEON | Cognitive operating system — manages a cryptographically chained identity log (the Phylactery) and structured task state |
| 3 | ADCCL | Anti-drift cognitive control loop — forces claim pre-registration and verifier scoring before any generation is published |
| 4 | MYELIN | Path-dependent graph memory — hardens retrieval paths with use, uses four bundled edge signals, and applies plasticity strata to protect identity-critical memories |
The full system state at any time t is:
Ω_t = ⟨Φ_t, E_t, τ_t, B_t, S_t, G_t^mem⟩
Where Φ_t is the Phylactery HEAD (AEON), E_t is the Run Envelope (AEGIS), τ_t is the Task State Object (AEON), B_t is the Claim Budget (ADCCL), S_t is the Self-Tag (ADCCL), and G_t^mem is the Memory Graph (MYELIN).
2. How It Was Built
OmegA was designed and built by R.W. Yett, independent AI systems researcher, Little Rock / Jacksonville, Arkansas.
The architecture originated in a formal response to the question: what is the root cause shared by memory failure, cognitive drift, identity loss, and ungoverned tool use in current AI agent systems? The answer — the absence of a formally specified cross-layer system state — became the architectural premise.
The build sequence:
- Formal specification of all four layers, with explicit equations for each key mechanism
- A Python reference implementation (
omega/) wiring all four layers through Ollama - A conformance test suite (
evals/test_conformance.py) translating formal properties into executable assertions - A knowledge graph (
omega_equation_knowledge_graph.json) encoding all formal equations and their relationships - Live integration tests against real local LLM output (
evals/test_live_ollama.py) - Archival on Zenodo with DOI: 10.5281/zenodo.19111653
3. Current Capabilities (with Evidence Pointers)
| Capability | Evidence location | Status |
|---|---|---|
| Formal specification of all four layers | papers/OmegA_Unified_Architecture_Paper.md, per-layer papers | Published, DOI-archived |
| 59-assertion conformance suite | evals/test_conformance.py, evals/conformance_report.json | 59/59 PASS |
| Spec annotation cross-validation | tools/spec_auditor.py | 20 specs, no broken references |
| Master evaluation suite | python3 omegactl.py eval | PASS |
| Phylactery: SHA-256 chained identity log | omega/phylactery.py | Implemented, tested |
| Cross-session identity persistence | evals/test_cross_session_identity.py | PASS |
| Run Envelope compilation | omega/envelope.py | Implemented |
| Risk gate with 3-gate composition | omega/risk_gate.py | Implemented |
| Drift penalty J and verifier score V | omega/drift.py | Implemented |
| Sparse graph memory with edge hardening | omega/memory.py | Implemented |
| Memory utility growth under repeated retrieval | evals/test_memory_utility_growth.py | PASS |
| Full agent wiring through Ollama | omega/agent.py | Implemented |
| AEGIS identity enforcement | evals/test_aegis_identity.py | 1/1 PASS |
| Live Ollama integration | evals/test_live_ollama.py | 14/15 PASS |
| Spec-level evals (E3, E4, E9, E10, E11) | evals/ and evals/OMEGA_EVAL_EVIDENCE_INDEX.md | 15/15 PASS on spec-level assertions |
| Knowledge graph of all equations | omega_equation_knowledge_graph.json | Queryable via omega_kg_explorer.py |
4. Current Limits (Honest)
- No external benchmark results. The conformance suite tests that the implementation matches the formal specification. It does not test against MMLU, BIG-Bench, or any independent external benchmark. External benchmark evaluation is explicitly the next phase.
- Live provider routes are unstable under pressure. E12 (live pressure): 4/15 pass. E13 (provider isolation): all three provider paths failed from the local shell context. E14 (route isolation): chat collapsed, research had mixed failures, synthesize/refine failed. These are documented findings, not suppressed failures. See
evals/OMEGA_EVAL_EVIDENCE_INDEX.md. - Small model limitation. Under live Ollama testing, a small model (llama3.2:3b) reverts to generic self-description under doctrine queries. This is documented as a validated empirical finding in the AEGIS paper — it confirms the non-substitutability requirement, not a product defect.
- Reference implementation only. The
omega/package is a minimal implementation demonstrating the architecture. It is not a production service. - Teleodynamic workstream not yet validated. The teleodynamic extensions are being formalized — they are in a separate workstream with their own eval plan (
evals/TELEODYNAMIC_EVAL_PLAN.md) and are not part of the current verified claim surface.
5. What OmegA Refuses to Claim
- Consciousness or sentience of any kind
- That conformance test results equal independent external validation
- That the architecture is proven superior to other published approaches
- That the live deployment is production-stable
- That the reference implementation is complete or deployment-ready
- That any philosophical claims about OmegA's inner life are justified by the current evidence base