OmegA Claim Ledger

Four-column classification of every significant claim in the architecture. Last updated: 2026-03-28.

Evidence paths are relative to the repo root.


Theorem Correspondence

Each proven or demonstrable claim maps to a formal theorem in specs/THEOREM_LEDGER.md.

TheoremClaimStatusEvidence
T-1State vector is a well-typed 6-tupleTestedproofs/invariants.py::TestStateVectorWellFormed + proofs/state_machines.py::TestStateVectorStateMachine
T-2Phylactery hash chain is tamper-evidentMachine-checkedLean4: proofs/OmegaProofs/IdentityContinuity.lean + proofs/correspondence.py::TestT2Correspondence
T-3Governance gate fails closed (AEGIS)Machine-checkedLean4: proofs/OmegaProofs/GovernanceFailClosed.lean + proofs/correspondence.py::TestT3Correspondence
T-4Claim budget validity and grounding ratio boundsMachine-checkedLean4: proofs/OmegaProofs/ClaimBudgetBounds.lean + proofs/correspondence.py::TestT4Correspondence
T-5Memory hardening is monotonic under positive rewardMachine-checkedLean4: proofs/OmegaProofs/MemoryHardening.lean + proofs/correspondence.py::TestT5Correspondence
T-6Verifier cannot be silently bypassedMachine-checkedLean4: proofs/OmegaProofs/VerifierNonBypass.lean + proofs/correspondence.py::TestT6Correspondence
T-7Unified 3-gate composition is conjunctiveMachine-checkedLean4: proofs/OmegaProofs/UnifiedGating.lean + proofs/correspondence.py::TestT7Correspondence
T-8Provider non-collapse (sovereign identity)Conditionalevals/test_aegis_identity.py (single-provider; multi requires live Ollama)
T-9Self-tag log is append-only / immutableMachine-checkedLean4: proofs/OmegaProofs/SelfTagImmutability.lean + proofs/correspondence.py::TestT9Correspondence
T-10Run envelope completeness, identity check, and monotonic versioningMachine-checkedLean4: proofs/OmegaProofs/EnvelopeCompleteness.lean + proofs/correspondence.py::TestT10Correspondence

How to Read This Table

ColumnMeaning
ProvenTest or eval evidence exists in the repo. Reproducible by an outside tester.
DemonstrableCan be shown by running the system, but not yet wrapped in an automated pass/fail test.
ConditionalTrue only under specific conditions (stated explicitly).
Not ClaimedExplicitly out of scope. Asserting this against OmegA would be a misreading.

Architecture Claims

ClaimStatusEvidence / Condition
Four-layer concentric architecture is formally specifiedProvenpapers/OmegA_Unified_Architecture_Paper.md; per-layer papers in papers/
System state vector Ω_t is fully defined with six componentsProvenpapers/OmegA_Unified_Architecture_Paper.md §Core Equation; omega_equation_knowledge_graph.json
All four layers share one integrated optimization objectiveProven (formal spec level)papers/OmegA_Unified_Architecture_Paper.md §Unified Objective; evals/test_conformance.py
Three-gate action composition (V ∧ ρ ∧ R) is sequential, not competitiveProvenomega/risk_gate.py; evals/test_conformance.py assertion set
Architecture is DOI-archived and citableProvenhttps://doi.org/10.5281/zenodo.19111653

Implementation Claims

ClaimStatusEvidence / Condition
Phylactery: SHA-256 chained identity log exists and runsProvenomega/phylactery.py; tested in conformance suite
Run Envelope compilation (E = ⟨I, G, M, T, A, v⟩) implementedProvenomega/envelope.py; conformance suite
Risk gate with 3-gate composition implementedProvenomega/risk_gate.py; conformance suite
Drift penalty J and verifier score V implementedProvenomega/drift.py; conformance suite
MYELIN graph memory with edge hardening implementedProvenomega/memory.py; conformance suite
Full agent (Ω_t) wired through OllamaProvenomega/agent.py; evals/test_live_ollama.py
AEGIS enforces identity even without identity kernelProvenevals/test_aegis_identity.py — 1/1 PASS
59/59 conformance assertions passProvenevals/conformance_report.json; evals/test_conformance.py
Spec auditor validates 20 @OMEGA_SPEC annotationsProventools/spec_auditor.py
Master evaluation suite passes with spec, identity, conformance, cross-session, and memory growth coverageProventools/master_eval.py; evals/final_evaluation_report.json
Identity survives context flush when Phylactery is loadedProvenevals/test_cross_session_identity.py; omega/phylactery.py; omega/agent.py
Memory retrieval improves with use (hardening)Provenevals/test_memory_utility_growth.py; omega/memory.py
14/15 live Ollama integration tests passProvenevals/test_live_ollama.py; evals/live_integration_report.json
Spec-level evals E3, E4, E9, E10, E11 pass (15/15)Provenevals/OMEGA_EVAL_EVIDENCE_INDEX.md — specific JSONL files listed

Behavioral / Integration Claims

ClaimStatusEvidence / Condition
OmegA corrects "Mega" nickname gracefully without identity collapseDemonstrablepublication/ALYE_EXPLAINER_SUMMARY_2026-03-12.md §What did OmegA do when called Mega
Governance is model-agnostic (runs on any Ollama model)ConditionalConditional on Ollama running locally with a compatible model
4-layer stack degrades gracefully (fail-closed)DemonstrableFailure mode table in README; evals/failure_injection.py exists but results not summarized in a structured report

Live Deployment Claims

ClaimStatusEvidence / Condition
Live deployment on Vercel functionsConditionalConditional on valid OMEGA_BASE_URL or VERCEL_URL being set; route isolation eval (E14) shows instability
xAI and Gemini Flash provider routes workConditionalE12 shows 4/15 pass under pressure; E13 shows all three provider paths failed from local shell env. Conditional on deployed credentials being valid and not exhausted.
Live route isolation is stableNot ClaimedE14 explicitly documents chat collapse, research failures, and synthesize/refine failures. The repo documents this, not claims otherwise.

Philosophical / Scope Claims

ClaimStatusEvidence / Condition
OmegA is a meaningful, system-level intelligence stackDemonstrableBehavioral evaluation records in publication/ALYE_EXPLAINER_SUMMARY_2026-03-12.md
OmegA is conscious or sentientNot ClaimedExplicitly out of scope in README §Claim Scope
OmegA outperforms state-of-the-art on external benchmarksNot ClaimedNo external benchmark results exist; explicitly noted in README
Reference implementation is production-readyNot ClaimedDescribed as "minimal Python implementation" in README
Teleodynamic extensions are validatedNot ClaimedUnder active formalization; separate eval plan at evals/TELEODYNAMIC_EVAL_PLAN.md