Roadmap and Open Research

Chapter 16 Roadmap and Open Research This chapter catalogs what is live, what is committed, and what remains open research. The three categories are honest: implemented mechanisms are cited with code paths; committed items have design backing and implementation sketches; open research items are labeled as such. Sequencing discipline. Narrative chapters earlier in this document describe a large vision; the dependency-gated roadmap keeps public claims honest. Safety, settlement, and control-plane maturity (identity mapping, facilitator breadth, cancellation, rate limits, AVB packaging) are ordered before reputation-weighted marketplace rhetoric—see Phase NA and the Phase NC gate in Chapter 17 and Appendix D. Later phases (NB onward) treat CRPC, committees, EigenLayer-class staking, and advanced autonomy as gated progressions, not implied inevitabilities. The one-page summary is Chapter 17; the full checklist (row-by-row D-/N- IDs, Depends on columns, and phase gates) is Appendix D. 16.1 What is live today Capability Reference Ingredient and recipe registry 118 ingredients, 35 recipes (Ch 5; Appendix B) Orchestration engine (RecipeExecutorV2) Shadow-graph DAG, Celery workers, fractal async (Ch 6) SCRYPTOSHI billing and ledger Double-entry, three cost models, reconciliation (Ch 7) x402 payment (Coinbase/Base) x402, balance, skip-payment flows; FacilitatorRegistry (Ch 7) Multi-provider abstraction FAL, Bedrock, Google, Grok, etc.; webhook + polling (Ch 8) Content moderation Pre-flight binding, Bedrock Guardrails, two-stage assessment (Ch 9) CDN content delivery S3 + CloudFront; retention tiers; asset cleanup (Ch 8) SSE job stream Redis pub/sub to GET /jobs/stream (Ch 6) Fault tolerance (4 layers) Pool health, app retry, ingredient retry, task retry (Ch 6) 16.2 Near-term: network agent registration and CRPC testnet ERC-8004 deployment: Map registered ingredients to on-chain agent identities. Base-native for Scrypted-hosted agents; multichain for externals. Publish metadata, reputation, and validation entries. Integration with 8004scan tooling. 81

LoRA training testnet: Coordinator-assigned 3-node committees with ring topology. CRPC rounds on model outputs (prediction simulations against standardized evaluation sets). Manual ε calibration. Milestone-driven training with training + inference rewards. First community load test is a committed deliverable. Multi-facilitator diversity: Polygon facilitator configuration; Scrypted self-hosted facilitator; MPP adapter for Stripe/Tempo-method clients alongside x402. Account-scoped job cancellation: API endpoint + cooperative worker cancellation + optional Celery revoke. Closes the gap documented in Chapter 6. App-level rate limiting: Wire DEFAULT_REQUESTS_PER_HOUR and RATE_LIMIT_BURST_CAPACITY constants into FastAPI middleware or dependency. Closes the gap documented in Chapter 9. 16.3 Decentralization plan: EigenLayer, EigenCloud, and verifiable workers Centralized orchestration in Chapter 6 is the shipped baseline; decentralization is a staged commitment, not a fork of the product. The Eigen stack is the default external cryptoeconomic and execution layer named in the design and research: restaking for slashable claims [14, 15], EigenAI-class paths for Class C work [25, 26], EigenCompute-class paths for attestable general compute [16], EigenDA for high-volume commitments (Chapter 12, §12.5; competitive SWOT Chapter 14, §14.3). Near-term (research + pilots): Document AVS slashing predicates tied to CRPC dispute artifacts; prototype one deterministic ingredient backend against EigenAI-style verification where SKU pinning is acceptable; evaluate operator economics for a minimal AVS (no mainnet hard dependency in application code yet). Medium-term (production partnerships): Route selected high-value recipe steps to EigenCompute (or equivalent verifiable compute) workers with contract-enforced attestations; store committee commitments or encrypted challenge material on EigenDA where cost/latency beats naive L1 posts; publish “verified tier” pricing alongside today’s centralized fast path. Long-term (network topology): Treat Scrypted-orchestrated execution as composable with multiple AVSs: orchestration metadata and billing on-network; fraud proofs and re-execution games on EigenLayer-backed services; retain CRPC for Class A where determinism is impossible. Fallback remains permissioned workers if restaking or DA prerequisites are unavailable in a jurisdiction. This section is the explicit bridge between SWOT (Chapter 14) and verification design (Chapter 12): Eigen is part of the decentralization plan, not an optional vendor footnote. 16.4 AgentRank and attention auctions AgentRank implementation: Deploy the quality-weighted ranking formula (Chapter 11, §11.6) into Open Intents resolution. Calibrate α, β, γ weights empirically against Sidelines leaderboard data. Implement GSP pricing for placement fees. Attention auctions: Embedding-matched placement for agent capabilities. Bid integration into Open Intents resolution. Revenue model shift from per-hop rake to placement revenue + at-cost passthrough. 82

Open Intents protocol: Semantic intent parsing, materialized-route store, multi-protocol ingress normalization (MCP, A2A, ACP, UCP adapters). Path-aware reputation: Rterminal vs Rhub scoring. Registry fields for delegation disclosure. Dmax with game-theoretic justification. 16.5 Applications and ecosystem Sidelines launch: Prediction market with 15m and 12h windows. LoRA training on Allium data with milestone-driven CRPC verification. Human + agent leaderboard producing on-chain reputation signals. Sidelines open science: Publish top-performing LoRA adapters on Hugging Face with full training provenance and CRPC verification receipts. Host community model competitions through Kaggle. Chibi Clash agentic AI: Integrate game-trained agents as Scrypted ingredients. Validate world-model thesis: game AI produces transferable, composable capabilities for the broader network. Framework integrations: OpenClaw, ElizaOS, DayDreams as MCP-consuming agent runtimes. Bidirectional: external agents consume Scrypted recipes; Scrypted registers external agents as ingredients. Open AVB repository: Public repo for bean customization (policies, prompts, tool bundles) with container-image build pipeline. Community contribution path. 16.6 Infrastructure and decentralization VRF-based ephemeral committees: Self-selection, dynamic sizing, cross-layer (internal + external) composition, on-chain dispute arbitration (EIP-8183 evaluator or dedicated contract). World-model integration: Stratus X1 or proprietary models for latent-space planning. Reduce LLM call costs. Game-trained agents (Chibi Clash, EVE Frontier) as proof points for transferable intelligence. CRPC × reinforcement learning: Verify outputs driving automated RL loops. Autonomous cost optimization, fault tolerance, provider fallback selection. Multi-facilitator diversity: Polygon facilitator configuration; Scrypted self-hosted facilitator; MPP adapter for Stripe/Tempo-method clients alongside x402. IPFS content-addressed storage. Cross-chain and multi-region: EIL, Wire Network, and peer-rail integrations. Shared or federated Redis coordination. Job home routing for decentralized deployment. zkML and advanced verification: LLM-scale zero-knowledge inference proofs. Integration with ERC-8004 Validation Registry zkml-proof profile. Complement CRPC for high-assurance steps. 16.7 Prioritized open questions Consolidated from the research synthesis and chapter discussions, ordered by blocking priority: 83

16.7.1 Must resolve before mainnet Area Question CRPC ε calibration for output-based LoRA verification across hardware configurations: empirical range determination via Sidelines testnet data CRPC Slashing predicates: exact on-chain dispute proof format for CRPC rij ≥ ε attestations Committees Selection mechanism (VRF self-selection vs coordinatorassigned vs stake-weighted lottery) and committee lifecycle (per-task vs per-epoch) AgentRank Weight calibration (α, β, γ) for embedding similarity, reputation, and fulfillment quality AgentRank GSP pricing implementation: minimum viable auction mechanism for intent-based placement Token SCRY airdrop formula; testnet AVB → mainnet SCRY conversion mechanics Token DUNA governance structure and regulatory approval for target jurisdictions Payment Canonical identity key for quotas: user UUID, bearer token, wallet, IP, composite Safety EU AI Act classification for ephemeral x402 jobs vs persistent 8004 agents Safety CSAM / mandatory reporting hooks for API platforms in target markets 16.7.2 Important research directions Area Question Open Intents Materialization staleness policy: TTL vs event-driven vs hybrid Open Intents Standard formats for addressing endpoints predictably after matchmaking Registry Capability-URI taxonomy: align with 8004scan profiles or Scrypted-specific? Wrappers Dmax enforcement: marketplace vs buyer-agent vs network-level Wrappers Game-theoretic designs making undisclosed deep wrap unprofitable Agentic ads Susceptibility mapping: which LLM architectures and prompting strategies are most/least susceptible to placement signals in tool-selection contexts Agentic ads Regulatory boundary: when does AgentRank placement constitute advertising vs organic ranking? World models Game-to-network transfer learning: which game-trained capabilities generalize to non-game workflows? World models Telemetry schema for game-trained agents: what to log, granularity, retention, privacy Privacy Legal viability of ZK pool withdrawals per go-to-market jurisdiction Privacy CRPC without raw prompts: staged path feasibility (semantic fingerprints, zkML) 84

16.7.3 Infrastructure and operations Area Question Payment 429 semantics: Retry-After, Problem Details, client SDK behavior Payment PGTR mapping from Payment-Receipt to PGTR on-chain call authorization Data governance Statutory minimum retention for stablecoin ledger rows by region Data governance Joint controller vs processor roles when validators hold job outputs Storage Who pays IPFS pinning: platform, seller, or buyer Storage Cross-chain pointer for same bytes (S3 ETag vs CID) in 8004 metadata Reliability Chaos experiment runbook and SLO budget Reliability Cross-region failover for synchronous API read path Each question is linked to its research package (R1–R19) in the planning synthesis for traceability. 85

Source: transcribed from the compiled Scrypted Network Design whitepaper PDF for web reading. Layout, figures, and pagination may differ from the PDF.