Open Autonomous Intelligence Initiative

Advocates for Open AI Models

Open SGI Architecture Considerations

1) Architecture (recommended)

L0 — UPA Core (axioms)

  • Canonical spec (your A1–A15) in machine-readable form (e.g., JSON/ASN.1 + optional DSL).
  • Provides symbols, semantics, and invariants.
  • Versioned; small, stable.

L1 — UPA Theorem Library (derived rules/patterns)

  • Machine-checkable theorems and design patterns derived from axioms.
  • Includes proof artifacts or test oracles (for lightweight stacks).
  • Tagged by domains: perception, planning, dialogue, governance, etc.

L2 — UPA Services (guardrails & planners)

  • Runtime services that apply axioms/theorems as guards, routers, mappers, and integrators:
    • A7 Context Router
    • A11 Recursive Integrator
    • A12 Cross-Axis Coherence Checker
    • A15 Harmony Estimator
  • Offer clean APIs/SDKs; handle policy, caching, telemetry, proofs.

L3 — Application Adapters

  • Domain-specific wrappers (chat, vision, orchestration, governance).
  • Expose capabilities (“contextual routing”, “deliberation”, “transparency”) instead of axiom primitives.

L4 — Apps & Objects

  • Use capabilities from L2/L3.
  • Optionally get read-only reflection into axiom/theorem usage for explainability or optimization.

2) Exposure policy (when to surface axioms/theorems)

Default: indirect via services.
Expose directly only when needed for:

  1. Planning/Reasoning modules (agents, compilers, schedulers)
    • They benefit from theorem-level hooks (e.g., A7/A11 operators).
  2. Explainability & audit (regulatory, safety)
  1. Objects/apps can retrieve “why” via an Explanation API:
  2. GET /explanations/{decision} → cites (A7, Thm-7.12, A15) with inputs/outputs.
  3. Third-party extensions/plugins
  4. Provide a Theorem Registry API and Guardrail SDK so partners can register new context mappers that must prove or test conformance to A1–A15.
  5. Research mode
  6. Full symbolic access for prototyping; gated in production.

3) Practical interface sketch

Axiom/Thm Registry (read-mostly):

GET  /axioms                 # list A1..A15 (symbol, defn, version)

GET  /axioms/A7              # details + machine schema

GET  /theorems?axiom=A7      # derived theorems under A7

GET  /theorems/Thm-7.12      # formal spec + pre/postconditions

Service layer (capabilities):

POST /context/route          # A7 router

# input: {signal, task, constraints}

# output: {selected_pole, confidence, proof_ref}

POST /integration/compose    # A11 recursive integrator

# input: {levels:[…], policies:[…]}

# output: {plan, invariants_satisfied:[A11,A12], proof_ref}

POST /coherence/check        # A12 cross-axis coherence

# input: {axes, config}

# output: {violations:[…], fix_suggestions, score}

POST /harmony/score          # A15 estimator

# input: {state, axes}

# output: {harmony_score, limiting_axes}

Explainability (reflection):

GET /explanations/{decisionId}

# returns {used_axioms:[A7,A11], theorems:[Thm-7.12], context, traces}

4) Object-level annotations (optional exposure in code)

Use guards and contracts rather than raw axioms:

@AxiomGuard(“A7”)           // ensures context-aware selection

@CoherenceRequired(“A12”)   // rejects configs that break cross-axis constraints

@ProducesHarmony(“A15>=0.7”)// SLO: harmony score threshold

function plan(…) { … }

5) Governance & versioning

  • SemVer for axioms (rarely change); theorems can evolve faster.
  • Proof/Oracle modes: formal proofs where feasible; otherwise strong tests/monitors.
  • Policy bundles: per-domain theorem sets (e.g., “Dialogues-v3”).
  • Telemetry: log axiom/theorem usage with hashes for audit.

6) Examples (how this feels in practice)

  • A7 Context Modulation: Service decides which policy/model to invoke given task + signals; app only sees “best policy + reason”.
  • A11 Recursive Integration: Service composes subplans across layers (edge → service → cloud) while enforcing integrative invariants.
  • A12 Multi-Axis Coherence: Pre-deployment checker blocks configs that maximize one axis (latency) while catastrophically harming another (safety).

This provides clean DX for builders, strong guarantees for safety/consistency, and transparent reasoning when needed—without forcing every app to become a theorem prover.