Skip to main content

System Architecture

1. Architectural Philosophy​

Sentry 2.0 is built on the principle of strict latency decoupling.

In high-frequency decentralized finance, market conditions shift in fractions of a slot (~400ms). If an execution engine attempts to make a remote HTTP request to an AI model while preparing a transaction, the network blockhash expires, liquidity pools move, and the transaction fails.

To solve this, Sentry 2.0 introduces a dual-path architecture:

  • The Deterministic Hot Path (sub-1ms): Executes entirely locally. Evaluates the in-memory Cuckoo filter (~150-300ns) deduplicator, checks deterministic safety gates (SIMD-0047 rent floors, circuit breakers), queries dynamic tip caches, and dispatches directly to Solami Beam's SWQoS TPU socket.
  • The Cognitive Cold Path (Groq LPU): Runs asynchronously out-of-band. Processes telemetry batches, audits execution receipts, performs root-cause classification on failed transactions, and tunes macro policies without ever blocking active transactions.

2. End-to-End System Diagram​

graph TD
    %% Telemetry Layer
    subgraph SENSE [1. Sensing & Ingestion Layer]
        RPC[Solami Private RPC] -->|Block & Slot Sampling| SNAP[Network Snapshot Engine]
        WS[Solami Native WebSocket] -->|Account & Slot Pulse| SNAP
        BLOCK[Live Block Scanner] -->|fetchLiveTransferEvents| STREAM[Real On-Chain Event Stream]
    end

    %% Preflight & Gate Layer
    subgraph GATE [2. In-Memory Preflight (~150-300ns) & Safety Gates]
        STREAM --> DEDUP{Cuckoo Filter Sub-50ns}
        DEDUP -->|Duplicate Detected| SUPPRESS[Suppressed: 0 Gas Spent]
        DEDUP -->|Unique Transaction| POLICY{Deterministic Policy Engine}
        POLICY -->|SIMD-0047 Rent Violation| ABORT[Policy Abort: Circuit Open]
        POLICY -->|Passed Guardrails| HOTPATH[Dynamic Tip Calculator]
    end

    %% Execution Layer
    subgraph EXECUTE [3. Dual-Rail Execution Engine]
        HOTPATH --> ROUTE{Regime Classifier}
        ROUTE -->|High Volatility / Congested| BEAM[Solami Beam SWQoS TPU Rail]
        ROUTE -->|Calm Standard| STDRPC[Solami Private RPC Rail]
        BEAM -->|Validator Inclusion| LAND[On-Chain Landing]
        STDRPC -->|Validator Inclusion| LAND
    end

    %% Auditing & Intelligence Layer
    subgraph VERIFY [4. Provenance & Cold Path Intelligence]
        LAND --> PROV[SHA-256 Chained Provenance Ledger]
        ABORT --> PROV
        SUPPRESS --> PROV
        PROV -->|Ed25519 Signed Receipt| LEDGER[(JSONL Matrix Ledger)]
        LEDGER -->|Out-of-Band Audit| GROQ[Groq LPU Llama 3.3 70B]
        GROQ -->|Macro Parameter Tuning| POLICY
    end

3. Core Subsystems Deep Dive​

3.1 Network Snapshot & Regime Classifier​

  • Source: lib/network-snapshot.ts
  • Function: Samples slot velocity, leader distance, and compute unit consumption across rolling 10-slot windows over Solami Private RPC.
  • Regime States: Classifies the current on-chain environment into one of five states:
    • calm: Normal block utilization, low base tips (under 2,000 lamports).
    • moderate: Steady transaction volume, standard tips (2,000 to 5,000 lamports).
    • congested: Blockspace competition, elevated tips (5,000 to 15,000 lamports).
    • extreme: Heavy DEX competition, rapid tip escalation (15,000 to 50,000 lamports).
    • volatile: Extreme price fluctuations or liquidation waves, maximum SWQoS priority routing.

3.2 Sub-50ns Cuckoo Filter Deduplication​

  • Source: lib/cuckoo-filter.ts
  • Function: Mathematical partial-key Cuckoo filter with 4 slots per bucket and 16-bit FNV-1a fingerprints.
  • Window Eviction: Operates a 150-slot sliding window (SlidingWindowCuckooDeduplicator) that exactly mirrors Solana's 150-slot blockhash lifetime.
  • Performance: Performs lookups in under 50 nanoseconds with zero heap allocation per check, catching transaction storms before they reach RPC.

3.3 Deterministic Policy Engine​

  • Source: lib/policy-engine.ts
  • Function: Enforces strict mathematical invariants that remote AI models cannot override:
    • SIMD-0047 Rent Protection: Enforces an inviolable minimum floor of 650,240 lamports, preventing operator wallet de-allocation.
    • Circuit Breaker State Machine: Automatically transitions from Closed to Open when consecutive simulation or RPC failures exceed thresholds.
    • Maximum Tip Budget: Binds priority fees to predetermined risk limits.

3.4 Solami Beam SWQoS Dispatcher​

  • Source: lib/action-engine.ts and engine/src/beam.rs
  • Function: Communicates with https://beam.solami.dev:11000 to deliver transactions directly to validator TPU sockets over Stake-Weighted Quality of Service lines.
  • Tip Injection: Queries https://api.solami.dev/onchain/tip-addresses dynamically to append micro-tips directly to active ...beam sink accounts.

3.5 Cryptographic Evidence Engine​

  • Source: lib/evidence-engine.ts
  • Function: Generates an append-only ledger where each receipt $R_i$ satisfies:
ReceiptHash_i = SHA-256(i || scenarioId || tip || status || sig || ReceiptHash_{i-1})
EngineSignature_i = Ed25519-Sign(ReceiptHash_i, PrivateKey)

This guarantees mathematical immutability and complete auditability.

3.6 Multi-Model AI Tip Cascade & Failure Taxonomy​

  • Source: lib/tip-engine.ts
  • Function: Multi-provider fallback cascade (Groq LPU -> Anthropic -> Google Gemini -> OpenAI -> Deterministic heuristic) protecting against single-point-of-failure inference timeouts during high-volatility congestion regimes.
  • Bounds: Enforces a strict 1,000 lamport absolute floor, 5,000,000 lamport ceiling under normal conditions, and dynamic escalation when recent failure rate exceeds 50%.
  • Error Taxonomy: classifyTxFailure() maps raw on-chain program failures and RPC errors into five structured classes with deterministic retry recommendations.

3.7 Orca Whirlpool Concentrated Liquidity Execution​

  • Source: lib/orca-swap.ts
  • Function: Direct on-chain liquidity action via @orca-so/whirlpools-sdk. Builds and executes real concentrated liquidity swaps across SOL/USDC, SOL/USDT, SOL/mSOL, and mSOL/USDC pools, routing output transactions directly to Solami Beam SWQoS TPU sockets.

3.8 Dual-Rail Simultaneous Submission​

  • Source: lib/dual-rail.ts
  • Function: Eliminates sequential retry lag by dispatching transactions simultaneously across Solami Beam SWQoS, Jito Block Engine bundles, and Solami Private RPC using Promise.allSettled. Whichever landing rail achieves validator confirmation first is taken; losing attempts are silently dropped.

3.9 RPC Cluster Consistency Diagnostic​

  • Source: lib/rpc-consistency.ts
  • Function: Real-time multi-node benchmarking engine probing Solami Private RPC head slot, response latency, and blockhash progression against public Mainnet-Beta and secondary RPC mirrors. Computes solamiAdvantageSlots quantifying empirical latency superiority.

4. Rust Confirmation Kernel​

The native Rust subsystem (engine/) provides low-overhead, zero-allocation confirmation tracking:

  • main.rs: Orchestrates Tokio asynchronous threads, boots the Solami RPC slot poller, and manages priority transactions.
  • beam.rs: Interfaces with Solami Beam SWQoS endpoints for direct TPU socket streaming.
  • geyser.rs: High-performance Yellowstone gRPC client providing microsecond-precision confirmation events.
  • lifecycle.rs: Multi-stage latency tracking measuring Processed to Confirmed to Finalized deltas.

5. Memory Management and Leak Prevention​

To run indefinitely in production environments, Sentry 2.0 enforces three memory management rules:

  1. Sliding Window Expiration: The Cuckoo filter automatically sweeps entries older than 150 slots. Memory usage stays strictly constant ($O(1)$) regardless of total execution count.
  2. Streaming Log Rotation: Receipts are streamed sequentially to JSONL files on disk rather than accumulated in memory.
  3. Connection Pooling: Solami RPC and WebSocket connections use persistent HTTP keep-alive and WebSocket reconnection backoffs with jitter, preventing socket descriptor exhaustion.