Architecture Overview¶

RCT Platform is organized as an 11-layer constitutional AI stack.

Layer Diagram¶

┌──────────────────────────────────────────────────────────┐
│               RCT PLATFORM SDK v2.0.0                    │
│         Intent-Centric AI Operating System               │
└──────────────────────────────────────────────────────────┘

Layer 11: CI/CD & Quality Gates
├─ GitHub Actions (ci.yml + security-scan.yml)
├─ 2,240 passing tests (10 skipped, 3 known - see TESTING_CANONICAL.md) · coverage pending re-measurement · Python 3.10/3.11/3.12
└─ Bandit 0 HIGH · mypy clean

Layer 10: Enterprise Hardening
├─ Pre-commit hooks (black, isort, bandit)
├─ Semantic versioning (2.0.0)
└─ CITATION.cff · CODE_OF_CONDUCT.md · CONTRIBUTING.md

Layer 9: Regional & Compliance
├─ 8 regional pairs (TH, JP, KR, CN, VN, ID, TW, US)
├─ PDPA / PIPA / PIPL / GDPR adapters
└─ core/regional_adapter/

Layer 8: Reference Microservices (5)
├─ intent-loop         ← 7-state pipeline, FDIA gating
├─ analysearch-intent  ← GIGOProtector, adversarial self-refinement
├─ vector-search       ← FAISS + Qdrant backends
├─ crystallizer        ← concept graph, keyword extraction
└─ gateway-api         ← routing, CORS, health

Layer 7: Intent Loop Engine
├─ RECEIVED → VALIDATED → MEMORY_CHECK → COMPUTING
├─ → VERIFYING → COMMITTING → COMPLETED
├─ Cold start: 3–5s   Warm recall: <50ms
└─ rct_control_plane/intent_compiler.py

Layer 6: SignedAI Consensus
├─ S/4/6/8 tier framework
├─ ED25519 cryptographic attestation
├─ 9 HexaCore roles (+ Ollama LOCAL + Groq LPU)
└─ signedai/core/

Layer 5: Delta Engine
├─ 39–90% smaller than full snapshots in measured bytes (grows with history)
├─ SHA-256 per-record deduplication
├─ Full rollback to any past tick
└─ core/delta_engine/

Layer 4: FDIA Engine
├─ F = D^I × A  (constitutional equation)
├─ A=0 → F=0, the gate refuses (on every path that applies it)
├─ No accuracy figure is claimed
└─ core/fdia/fdia.py

Layer 3: Control Plane DSL (15 modules)
├─ intent_schema · dsl_parser · execution_graph_ir
├─ policy_language · observability · replay_engine
├─ jitna_protocol · signed_execution · cli · api
└─ rct_control_plane/

Layer 2: JITNA Protocol (RFC-001 v2.0)
├─ Wire format: { I, D, Δ, A, R, M }
├─ Standard for cross-agent intent negotiation
└─ signedai/core/models.py

Layer 1: Core SDK Primitives
├─ IntentObject · ExecutionGraph · DSLParser
├─ FDIAScorer · SignedAIRegistry · MemoryDeltaEngine
└─ Exportable under Apache 2.0

Data Flow — Single Intent Request¶

sequenceDiagram
    participant U as User/Agent
    participant CP as Control Plane
    participant FDIA as FDIA Engine
    participant SAI as SignedAI Consensus
    participant DE as Delta Engine

    U->>CP: POST /compile (JITNAPacket)
    CP->>FDIA: score_action(intent, data, architect)
    FDIA-->>CP: score ∈ [0.0, 1.0]
    alt score < threshold or A=0
        CP-->>U: 403 Blocked (constitutional gate)
    else score ≥ threshold
        CP->>SAI: verify(action, tier)
        SAI-->>CP: consensus + ED25519 signature
        CP->>DE: record_delta(agent_id, tick, changes)
        DE-->>CP: compressed delta saved
        CP-->>U: 200 OK (signed execution graph)
    end

Key Design Decisions¶

Why the FDIA equation is not a config option¶

The equation F = D^I × A is enforced at the mathematics level. Governance is not a switch — when A=0, F=0 by arithmetic, not by condition check. This makes the guarantee unforgeable.

Why deltas instead of state snapshots¶

Storing full agent state at every tick is O(n×t). Delta Engine stores only changed_fields with SHA-256 content hashing, reducing the stored log by 39% (20 ticks) to 90% (500 ticks) in measured bytes (scripts/measure_delta_engine_real.py; the gain grows with history, and plain zstd over full snapshots compresses further). Full rollback is still possible by replaying the delta chain.

Why ED25519 over RSA¶

ED25519 signatures are 64 bytes vs RSA-2048's 256 bytes, verify 3× faster, and offer equivalent security. Critical for high-throughput multi-LLM consensus where thousands of decisions need attestation per second.