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: LLaMA 3 70B requires hardware that costs $60,000–120,000 to acquire and $8,000–12,000/year to operate. Globally, only 50–200 nodes at this level are realistic. This contradicts the decentralization promise.

Hardware cost reality
Node TypeHardwareAcquisition costAnnual electricityAccessibility
Light Client Any device, 4 GB RAM $0 extra Minimal Everyone
Guardian (8B) RTX 4070 Ti+, 16 GB VRAM $800–1,500 ~$1,000 Enthusiasts
Guardian (70B) 4× A100/H100, 128 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 is accessible — RTX 4070 Ti hardware is within reach for enthusiasts, and 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.

For a Guardian operator to run a node economically, PROM earnings must cover hardware and operating costs. The following calculations use conservative assumptions and are based on Year 1 emission.

Important caveat: PROM has no price before mainnet (May 5, 2026). These calculations use target price scenarios — not predictions. Actual economics depend entirely on market adoption.

Assumptions (Year 1)
VariableValueSource
Guardian emission Y16,000,000 PROMWhitepaper tokenomics
Active Guardians (realistic Y1)50–150 nodesConservative estimate
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
$0.017
PROM/token to break even in Year 1
Annual hardware cost~$1,000
Electricity~$1,000
Total annual cost~$2,000
Avg PROM earned~20,000
Break-even price$0.10
Profitable above$0.10/PROM
Guardian 70B (owned)
$1.83
PROM/token to break even in Year 1
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
Profitable above$0.28/PROM
Guardian 70B (cloud)
$3.50+
PROM/token to break even in Year 1
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
Profitable above$0.144/PROM

Key insight: The 8B Guardian is economically the most attractive entry point. Break-even at $0.10/PROM is achievable if the network gains meaningful adoption. Cloud-hosted 70B is actually more capital-efficient than owning hardware — but depends entirely on PROM market price.

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

Early operators earn significantly more PROM per node — both because emission is higher in Year 1 and because fewer nodes compete for rewards. This is the strongest economic argument for early participation despite the uncertain PROM price.

Solution Approaches

Five paths to solving Guardian centralization.

None of these solutions is implemented yet. They represent the current best thinking on how to address the problem. Community input is explicitly invited — this is an open design question.

Immediate · High impact
1. Hybrid Routing (8B/70B escalation)
8B Guardians handle standard cases. 70B escalation only triggers when confidence score falls below threshold (e.g. <0.70). Most threats are known variants — 8B handles them well. True zero-days escalate automatically to 70B.
Impact: 70–80% of work stays at 8B level. 70B needed only for edge cases. Decentralization dramatically improved.
Immediate · Proven pattern
2. Ensemble Voting (multiple 8B nodes)
Instead of one 70B node, 5–10 independent 8B Guardians analyze the same threat and vote on the YARA rule via majority consensus. Disagreement triggers 70B escalation. Ensemble accuracy approaches 70B quality for common threats.
Impact: Zero new hardware requirements. Same quality via consensus. Strongly increases Guardian count and geographic distribution.
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 recommendation: Solutions 1 (Hybrid Routing) and 2 (Ensemble Voting) can be implemented immediately with existing architecture. They are the highest-impact, lowest-risk interventions. Solutions 3–6 require additional protocol design and will be addressed in Sprints 9–14.

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 critical insight: KAS already has economic value. On day one of mainnet, Validators stake KAS — an asset worth over $1 billion in market capitalization that has been established since 2021. This means Prometheus inherits real economic security immediately, without waiting for PROM to develop a price.

Why this matters: Most new protocols bootstrap security with their own token that has zero initial value. Prometheus bootstraps with KAS — a proven asset. Security comes from established value on day one. PROM can develop a price organically, under no pressure to be the primary security mechanism.

Early-mover dynamics

The deflationary emission schedule creates a natural bootstrapping incentive. Year 1 emits 20M PROM across fewer nodes than Year 3. An operator who joins in Month 1 earns dramatically more PROM per node than one who joins in Year 2 — at lower competition and higher token scarcity.

This is not artificial scarcity. It is a direct consequence of the network growing while emission decreases. The economic rational actor calculation strongly favors early participation even under price uncertainty.

Realistic bootstrap phases
Phase 1
Mo 1–3
Team-operated nodes. 10 Validators (team) + 10–20 Guardian 8B (team) + 2–5 Guardian 70B (team/partners). Security level: known threats covered. First PROM minted. DEX pool opens. Price discovery begins.
Phase 2
Mo 4–12
Community expansion. +500 Light Clients. +50 Guardian 8B (community). +10 Guardian 70B (institutions/data centers). Dev Pool grants cover early 70B operators. Security level: good coverage of common threats + zero-day capability.
Phase 3
Year 2+
Self-sustaining network. 10,000+ Light Clients. 500+ Guardian 8B. 30–50 Guardian 70B. PROM market price established. Full security coverage. Guardian Pooling Protocol active. Specialization emerges organically.
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.

Current countermeasures (implemented)
  • 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 60-second end-to-end guarantee has been proven in E2E tests — but under controlled conditions with no network latency, no load, and stub implementations for ZK-proof and AI inference.

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 familiesAdequate (with fine-tuning)Reliable
YARA for common CVEsPossible with fine-tuningReliable
Zero-day analysisInsufficient aloneRequired
Multi-stage threats (>128k context)Context window too small128k token context
YARA syntax correctnessImprovable via YARA-specific fine-tuningHigh quality

The 85% confidence threshold combined with the reputation system provides a natural quality filter — 8B Guardians that produce poor rules lose reputation and influence over time. However, for genuine zero-day threats, 70B remains necessary. This is why maintaining at least a small number of 70B Guardians in the network is critical even if 8B dominates in count.

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.