Home Protocol Architecture Tokens Roadmap FAQ Whitepaper Economics GitHub ↗
⚠ Core Problem

Guardian centralization risk.Open

This is the most significant structural tension in Prometheus. We document it transparently because pretending it doesn't exist would undermine the trust the protocol depends on.

The problem in one sentence: Using the illustrative planning ranges below, LLaMA 3 70B requires institutional-grade hardware and operating cost. Globally, only a limited number of nodes at this level are realistic. This contradicts the decentralization promise.

Hardware cost reality

All cost figures below are illustrative planning ranges, not current quotes. Acquisition cost may mix new and used purchases; electricity varies materially with regional rates, utilization, cooling, and power limits; cloud pricing varies by provider, commitment, and availability.

Node TypeHardwareIllustrative acquisitionIllustrative annual electricityAccessibility
Light Client Any device, 4 GB RAM $0 extra Minimal Everyone
Guardian (8B) 24 GB NVIDIA GPU (RTX 3090/4090 class) $800–1,500 ~$1,000 Enthusiasts
Guardian (70B) 4× A100/H100 80 GB, 256 GB RAM $60,000–120,000 $8,000–12,000 Data centers only
Validator Standard server + 10,000 KAS ~$500–1,000 ~$500–1,000 KAS holders
Honeypot Any internet-exposed server ~$0 Minimal Everyone

The 8B Guardian remains accessible on a single 24 GB consumer GPU, including RTX 3090/4090-class hardware. The reputation system rewards quality output regardless of model size. The 70B Guardian is categorically different: it requires institutional-grade infrastructure.

Economics

Break-even calculations.

A future Guardian operator would need rewards to cover hardware and operating costs. The following calculations are hypothetical sensitivity scenarios based on planned Year 1 tokenomics, not forecasts or active rewards.

Important caveat: PROM minting, emission, liquidity and trading are not implemented, deployed or active. These values are planning inputs, not current quotes, expected earnings, investment guidance or predictions.

Assumptions (Year 1)
VariableValueSource
Guardian emission Y16,000,000 PROMWhitepaper tokenomics
Guardian scenario50–150 nodesUnproven planning assumption
PROM per Guardian (avg, 100 nodes)60,000 PROM/year6M ÷ 100
Reputation weighting70B earns ~3–5× more than 8BQuadratic voting power formula
Break-even price scenarios
Guardian 8B
Scenario
Illustrative inputs; no active PROM market
Annual hardware cost~$1,000
Electricity~$1,000
Total annual cost~$2,000
Avg PROM earned~20,000
Break-even price$0.10
Modeled threshold$0.10/PROM
Guardian 70B (owned)
Scenario
Illustrative inputs; no active PROM market
Hardware (amortized 5yr)~$18,000
Electricity~$10,000
Total annual cost~$28,000
Avg PROM earned (5× weight)~100,000
Break-even price$0.28
Modeled threshold$0.28/PROM
Guardian 70B (cloud)
Scenario
Illustrative inputs; no active PROM market
Cloud GPU (4× A100)~$1,200/mo
Annual cloud cost~$14,400
Total annual cost~$14,400
Avg PROM earned~100,000
Break-even price$0.144
Modeled threshold$0.144/PROM

Scenario result: Under these selected inputs, 8B has a lower modeled cost threshold than 70B. This does not establish achievable profitability, future PROM value, cloud availability, real model quality or an operating reward path.

Early-mover advantage in numbers
YearTotal Guardian emissionAt 100 nodesAt 500 nodesAt 2000 nodes
Year 16,000,000 PROM60,000/node12,000/node3,000/node
Year 25,400,000 PROM54,000/node10,800/node2,700/node
Year 34,800,000 PROM48,000/node9,600/node2,400/node

The table illustrates how a fixed planned pool would divide across assumed node counts. It does not promise earnings, encourage early participation, or account for eligibility, quality, uptime, abuse controls, market value, taxes or operating-network rules.

Solution Approaches

Six candidate paths for Guardian centralization.

Repository state machines and an 8B-first routing scaffold exist, but none of these approaches is proven as an operated decentralization solution. Merged and exact-main-verified GH-253/PR #254 adds owner-local transition-key succession mechanics only; real-world key ownership, external/decentralized membership authority, Sybil resistance, model quality, public multi-host behavior and economics remain open, and no production authority follows. Product exact main is 5920cb4bb737376977f762beb0d5e3108519c7a0; Prometheus CI 34031999904, Security Audit 34031999907, and Pages 34031999575 pass.

Immediate · High impact
1. Hybrid Routing (8B/70B escalation)
The tested router selects an injected 8B path first and an independent 70B path below confidence 0.70. No real model quality, workload mix, zero-day behavior or automatic production routing has been evaluated.
Target impact: reduce dependence on 70B hardware; the actual routing rate and decentralization effect are unproven.
Immediate · Proven pattern
2. Ensemble Voting (multiple 8B nodes)
The tested local ensemble state machine requires at least five configured Guardian identities and strict-majority ballots. It does not prove independent operators, model accuracy, 70B-equivalent quality or public-network consensus.
Target impact: distribute analysis; membership, Sybil, quality and geographic claims remain unproven.
Medium term · Architecture needed
3. Guardian Pooling Protocol
Multiple operators share a single 70B node. PROM rewards split on-chain proportional to compute contribution. A data center with 4× A100 can serve 10+ operators who each contribute GPU time and receive proportional PROM.
Impact: Lowers 70B entry barrier from $60k to $6k per operator. On-chain split ensures trustless revenue sharing.
Medium term · Ecosystem maturity needed
4. Specialization Sharding
8B Guardians specialize by attack class: Guardian A handles ransomware, Guardian B handles network intrusion, Guardian C handles privilege escalation. Specialization + fine-tuning makes 8B competitive with 70B within its domain.
Impact: Fine-tuned 8B models match 70B quality in their specialty. Natural ecosystem of specialized Guardians emerges organically.
Medium term · Research needed
5. Incentivized Cloud Pooling
DAO provides initial grants (from Dev Pool) to cover first 6–12 months of 70B cloud costs for selected operators. Operators commit to providing public Guardian service in return. After PROM market price matures, grants phase out.
Impact: Ensures 70B presence in early bootstrapping phase without requiring operators to bear full risk during price discovery.
Medium term · Cryptography research
6. PLONK instead of Groth16
Groth16 ZK-proofs can take 5–30 seconds on CPU (Light Client hardware). PLONK proofs are faster to generate and don't require a trusted setup. If Kaspa natively supports PLONK verification, Light Client proof generation drops to <2 seconds.
Impact: 60-second SLA becomes reliably achievable even on weak hardware. Removes GPU requirement for Light Clients entirely.

Current engineering direction: Continue testing the existing hybrid-routing and ensemble state-machine foundations before evaluating them with real models and independent operators. No candidate is approved as a production decentralization solution; pooling, specialization, grants and alternate proof systems require separate design and review.

Bootstrapping

The chicken-egg problem — and how KAS solves it.

Every decentralized network faces the same fundamental challenge: without nodes, there is no protection. Without protection, there are no users. Without users, there is no incentive for nodes. Prometheus has a structural answer to this that most protocols lack.

KAS as the economic bridge

The design uses KAS, never PROM, for validator stake. KAS is an external network asset, but this does not automatically give Prometheus economic security. That requires deployed contracts, correct membership and stake operation, adequate value at risk, monitoring and a live validator network.

Design intent: separating KAS collateral from PROM rewards avoids PROM self-staking. It is an incentive assumption, not proof of day-one security or future PROM value.

Early-mover dynamics

The specification plans 20M PROM in Year 1 and a declining schedule. Actual per-node issuance would depend on implementation, verified contributions, eligibility, abuse controls and active operation. No early-operator return is promised.

Realistic bootstrap phases
Phase 1
Mo 1–3
Planning scenario only. A small operator-controlled validator/Guardian set, initial PROM issuance and a later pool would each require separate implementation, governance, security and deployment evidence. None is active.
Phase 2
Mo 4–12
Target expansion scenario. Public clients, independent Guardians and Dev Pool grants remain contingent on real-model quality, privacy, membership, Sybil, token and operational gates.
Phase 3
Year 2+
Long-term target scenario. Node counts, market liquidity, coverage, pooling and specialization are aspirations, not forecasts or demonstrated outcomes.
Open Problem

Sybil resistance for Guardians.Partial

An attacker who operates 10 A100 GPUs could theoretically create 5 independent 70B Guardian identities and dominate the pre-analysis stage — injecting malicious YARA rules.

Specified or locally tested mechanisms; no Sybil-resistance proof
  • Quadratic voting power: power = (reputation / 100)² × compute. Linear expansion of nodes does not linearly expand power.
  • PoW registration: Difficulty scales with current Guardian count. 100 fake nodes cost 100× the registration effort.
  • Historical success rate: New nodes start at reputation 0.1 (REPUTATION_START). Building influence requires a track record of accepted proposals — which requires honest behavior over time.
  • Validator final gate: Guardians propose, Validators decide. Even if all Guardians collude, 67% of Validators (who have KAS at risk) must also approve.
Open question (not yet designed)

How do we prevent a single operator with 10 A100s from registering 5 independent 70B identities? Current PoW scales with node count but does not verify that two nodes are operated by different people. Proposed approaches under consideration:

  • Require KAS or PROM stake per Guardian identity (financial Sybil cost)
  • Minimum time-weighted reputation before high voting power (slow Sybil)
  • Hardware fingerprinting via ZK-proof (verifiable compute without revealing identity)
  • Geographic distribution requirements (prevents single-datacenter dominance)

Status: This is an open design question. Community proposals are welcome. A final mechanism will be specified before mainnet deployment.

Open Problem

60-second SLA under real-world load.Partial

The repository has a same-host development-stub fixture below 60 seconds. It is not a guarantee or production performance evidence because it has no real model, approved proof, public network load, IPFS availability or operated consensus.

StepTarget timeRisk factorStatus
ZK-proof generation~2sCPU: 5–30s possible on weak hardwareRisk
LLaMA inference~12sHigh load: queue buildup on popular Guardian nodesRisk
Network propagation~1sIntercontinental latency: +100–300ms per hopLow risk
Validator voting window20sFixed — not a bottleneckFixed
Kaspa finality<1sDAGKnight — provenProven

The real risk is ZK-proof generation on low-end hardware. Groth16 on a 5-year-old laptop CPU can take 20–30 seconds — which would push the total time beyond 60 seconds before the threat even reaches the first Guardian node.

Mitigation under evaluation: PLONK proofs (faster on CPU) as alternative to Groth16. Local GPU for Light Clients as optional recommendation. Asynchronous proof generation (Light Client submits hint immediately, proof follows). Final decision in Sprint 9.

Known Limitation

8B vs 70B YARA rule quality.Acknowledged

TaskLLaMA 3 8BLLaMA 3 70B
Known malware familiesNot evaluatedNot evaluated
YARA for common CVEsNot evaluatedNot evaluated
Zero-day analysisNot evaluatedNot evaluated
Multi-stage threatsNot evaluatedNot evaluated
YARA syntax correctnessCompile-only fixture existsNo real-model output evaluated

No independently evaluated real 8B/70B run, calibration, real-sample quality result or actionable authority is published. The 85% threshold and reputation transitions are policy/state-machine inputs, not proof of detection quality or a requirement that 70B solves zero-day analysis.

Open Research

Questions we do not yet have answers to.

The following are genuine open questions where Prometheus would benefit from community research, academic collaboration, or external input. If you have relevant expertise, contributions are welcome.

Immediately actionable
  • Guardian routing mechanism: What is the optimal algorithm for deciding 8B vs 70B escalation? Confidence-score-based? Reputation-weighted? Random sample? Combination?
  • Ensemble protocol design: If 7 of 10 8B Guardians agree on a YARA rule — what constitutes "agreement"? Exact match? Semantic similarity? Confidence-weighted average?
  • PLONK on Kaspa: Does the Kaspa L1 support PLONK proof verification natively, or would this require a protocol change?
Medium term
  • YARA-specialized 8B model: Training a fine-tuned 8B exclusively on YARA syntax and CVE datasets — does this reach 70B quality for rule generation while keeping inference lightweight?
  • On-chain reputation formula: The current formula (quadratic voting power) works well theoretically. Does it hold under adversarial conditions with coordinated reputation farming?
  • Guardian Pooling cryptography: How do multiple operators share a 70B node with on-chain PROM splitting without any single operator being able to defect with the full reward?
Long-term research
  • Sub-1B ZK-capable models: Are there models under 1 billion parameters that can reliably detect specific anomaly types — small enough to run on IoT devices or phones with <2 GB RAM?
  • Decentralized Guardian orchestration: How do Guardians coordinate without a central coordinator — especially for multi-stage threat analysis that requires sequential steps?
  • Privacy-preserving model evaluation: Can we verify that a Guardian's AI model is the authentic, unmodified version using ZK-proofs — without revealing model weights?

Contributing: Open issues for each of these research questions are tracked in the GitHub repository. If you have relevant expertise in cryptography, distributed systems, or ML, your input is genuinely valued. This is not vaporware — the foundation code exists and these are real engineering problems.