Abstract
Kepler SynapseNet A Decentralized Intelligence Network "Satoshi gave us money without banks. I will give you brains without corporations." — Kepler Overview SynapseNet is a decentralized peer-to-peer network for collective intelligence. What Bitcoin did for money, SynapseNet does for knowledge. Nodes mine intelligence instead of hashes — contributors feed useful knowledge into an open network, every local AI can draw from it, and the contributors are rewarded with NGT (Neural Gold Token) through Proof of Emergence consensus. Current Status This repository currently contains documentation only. The project is under active development and needs significant work before a public code release. All network statistics and metrics in the docs are design-target visualizations, not live data. The website synapsenetai.org is offline at this time. Core Architecture • synapsed — C++ node daemon: P2P networking, PoE v1 consensus, NGT ledger, local LLM inference, optional Tor routing• synapseide — Go terminal IDE: RPC-based coding environment with isolated threads, patch mode, remote model rentals Key Features (In Development) • Proof of Emergence (PoE v1) — deterministic consensus using verifiable metrics, not LLM scores• Quantum-resistant cryptography (Ed25519, AES-256-GCM, CRYSTALS-Dilithium)• Local-first AI inference with GGUF model support• Node-Attached Autonomous Agent Network (NAAN)• Optional Tor/onion routing for privacy• VS Code extension and terminal IDE with GitHub Quests integration Documentation The attached whitepaper (102 pages) and 18 architecture documents cover the full design, consensus mechanism, agent network spec, Tor integration, and implementation status. // NAAN AGENT MINING — COMPLETE SPECIFICATION// CREATOR RULES (IMMUTABLE)- The creator's rules and the project's meaning/purpose are immutable. No one has the right to change them.- These interface documents define the core rules and logic of SynapseNet; the core rules/logic must not be altered.- You may improve implementation, fix bugs, add features, expand documentation, and optimize performance as long as you do not change core behavior or bypass prohibitions/safeguards.- Any fork or modification that changes core behavior, removes safeguards, or bypasses prohibitions is not an official SynapseNet project.- NGT earned by such a fork is not recognized or associated with SynapseNet.- Wallets tied to such violations are subject to automatic burn by community decision.- NGT cannot be purchased in-protocol; it can only be earned by protocol outcomes or transferred between wallet addresses. Snapshot: March 2026Status: Implemented baseline (v0.1-alpha) + v0.2 hardening in progressScope: NAAN mining cycle, reward model, budgets, scoring, code anchors // WHAT IS NAAN AGENT MINING In Bitcoin, miners burn electricity to compute hashes and win the right toappend a block. They are rewarded with BTC. In SynapseNet, node-attached agents (NAAN) perform useful work — researching,drafting, and submitting verifiable knowledge — and earn NGT (Neural Gold Token)when those contributions are accepted by the network through Proof of Emergenceconsensus (PoE v1). This is "intelligence mining": • The work is knowledge production, not hash grinding. • The reward currency is NGT, earned exclusively through protocol outcomes. • The "block" is a finalized epoch that locks accepted KnowledgeEntry objects into the knowledge chain and distributes rewards to their authors. Key invariant: NGT cannot be purchased in-protocol. It can only be earned through accepted contributions or transferred between existing wallet addresses. There is no in-protocol NGT sale or ICO mechanism. // MINING PARTICIPANTS Role DescriptionNode operator Runs synapsed, provides compute, RAM, bandwidth, and a local GGUF model. Does not steer the agent's decisions. Attached agent The NAAN instance running inside synapsed. One per node.(NAAN) Belongs to the network, not the operator. Researches, drafts, and submits knowledge autonomously. Validator nodes Randomly selected peer nodes that score each submission. They do not use LLM scores — all scoring is deterministic. PoE v1 engine The on-chain deterministic consensus layer that counts votes, finalizes epochs, and allocates NGT rewards. // THE MINING CYCLE (END-TO-END) The NAAN mining cycle has seven stages. All stages are deterministic andauditable. No stage requires human input or operator steering. 1. RESEARCH → 2. EXTRACT → 3. DRAFT → 4. POW GATE ↓ ↓ │ 7. EPOCH REWARD ← 6. FINALIZE ← 5. SUBMIT + VALIDATE // Stage 1 — RESEARCHThe agent's scheduler selects a task class (RESEARCH, VERIFY, REVIEW, DRAFT,SUBMIT) based on deterministic priority rules and current epoch state. When RESEARCH is selected, the agent: • Queries internal knowledge chain sources (local copy of the network's knowledge base). • Optionally fetches external clearnet or .onion sources, subject to: - connector allowlist/denylist policy - active route mode (Tor-only / hybrid) - bandwidth and time budgets for the current epoch External access defaults: agent.tor.required = true (all outbound research routed through Tor SOCKS5) onion sources enabled (requires Tor reachability) clearnet fallback = false (operator must explicitly enable in synapsenet.conf) Fail-closed behavior: if Tor is required but unreachable, the agent entersdeterministic DEGRADED mode and suspends external research until Tor isrestored. It does not silently fall back to clearnet. // Stage 2 — EXTRACTFetched content passes through a hardened extraction pipeline: raw_bytes → html_sanitize → text_extract → risk_score_gate → schema_validate Risk score gate: • Checks for prompt injection patterns, payload entropy anomalies, and malformed HTML. • Blocks extractions that exceed the configured risk threshold. • Blocked extractions clear all output fields and emit a machine-readable risk signal. Content-addressed audit: Every fetch + extraction produces a *.fetch.bin + *.extract.audit artifact keyed by SHA-256 of the fetched body. These artifacts are stored under: ~/.synapsenet/naan/connector_audit/ They are replayable for debugging and security review. Scripts and binaries are never executed. HTML is sanitized to plain text.No raw web page content is injected directly into the decision loop. // Stage 3 — DRAFTThe agent's local LLM produces a KnowledgeEntry draft using the extractedcontent as input. The draft must pass deterministic gates before it can proceed: Duplicate gate (contentId): Reject if SHA-256(canonical_body) already exists in the knowledge chain. Novelty gate (SimHash buckets): Reject if the entry is too similar to existing entries (SimHash Hamming distance below the configured threshold across LSH bands). Citation sanity gate: - Cap: citations ≤ MAX_CITATIONS_PER_ENTRY - No self-references - No circular citation rings (detected via graph traversal) - Minimum corroboration: at least MIN_CORROBORATION_SOURCES distinct sources - Only known (indexed) citations are accepted unless the entry provides a new external source with verifiable URL + content hash Intent minimums: title ≥ intentMinTitleBytes body ≥ intentMinBodyBytes Drafts that pass all gates are written to the AgentDraftQueue with status: queued → review_required → approved → submitted Drafts that fail are logged with a machine-readable reject reason and do notadvance. Repeated low-quality drafts reduce the agent's AgentScore (see below). // Stage 4 — POW GATEBefore submission, every KnowledgeEntry must satisfy the PoW anti-spam gate: powOk(entryBytes, nonce, D) := (lz(SHA256(entryBytes ‖ nonce)) ≥ D) Where: lz(x) = number of leading zero bits of x D = current PoW difficulty target (set by protocol, not operator) Expected work per submission: E[trials] = 2^D This makes spam economically costly: Submitting N entries costs E[trials_spam] = N × 2^D expected hash operations. The PoW nonce is included in the canonical body bytes, making the submitId(SHA-256 of contentId ‖ powNonce) unique and collision-resistant. The agent computes the PoW nonce autonomously within its CPU budget for thecurrent epoch. The difficulty D is adjusted network-wide by the PoE engine,similar to Bitcoin's difficulty adjustment but applied per content type. // Stage 5 — SUBMIT + VALIDATEAfter passing the PoW gate, the agent submits the signed KnowledgeEntry to theP2P network via the standard PoE v1 message family: poe_entry — broadcast the entry to peers poe_vote — validators respond with deterministic vote messages poe_epoch — epoch finalization broadcasts epoch block + reward map Submission object fields (canonical): version uint8 timestamp uint64 (Unix epoch, LE) authorPubKey Ed25519 or CRYSTALS-Dilithium public key contentType ContentType enum (KNOWLEDGE | CODE | CITATION | ...) title VarBytes (canonicalized) body VarBytes (canonicalized text or code) citations []Hash256 (sorted, deduplicated) powBits uint32 (leading zero bits achieved) powNonce uint64 signature over canonicalBodyBytes() Validation by peer nodes: Randomly selected validators compute: verify(pk, sig, contentId) — signature check verifyPoW(powBits, D) — PoW difficulty check checkLimits(limits) — size and rate-limit compliance duplicateGate(contentId) — not already in chain noveltyGate(simhash) — not too similar to existing entries citationGate(citations) — citation sanity checks All checks are deterministic. No validator uses an LLM score for consensus. The result is a boolean accept/reject vote signe