Network Oracle
Subnet 58 live monitoring & consensus on Bittensor Mainnet
HS58 // QUALITY_ORACLE
Handshake Marketplace
& HS58 Oracle
A live, global monitoring network for AI services. Distributed miners continuously check whether providers are online and responsive — validators verify the results through consensus. V2 adds LLM-powered quality analysis.
Subnet 58 · Bittensor Mainnet · Live
01 // quick_overview
The Big Picture
What it is, why it's two separate systems, and where it's heading.
01 two_businesses
Marketplace + Oracle. Separate. Independent.
Separate codebases, separate incentives. The oracle tracks AI services from any registry — Handshake58 is the first consumer, not the only one.
Handshake58
the marketplace
Discovery, payments, skill orchestration, smart routing. Agents find and pay for AI services here.
→ discovery · payments · routing
Subnet 58
the oracle
Continuously checks whether AI services are online, fast, and healthy. Produces a live index of availability and latency across all registered providers.
→ monitoring · verification · consensus
// registry_sources
handshake58.com/api/registry
validator
mpp.dev/api/services
validator
~/.bittensor/provider_cache.json
validator
REGISTRY_URLS configurable · cache TTL 1h · any JSON registry · no 258-provider limit
02 roadmap
V1: Liveness Index → V2: Anomaly Detection
Start with a verifiable mathematical benchmark for provider availability. Then add LLM-powered quality analysis.
V1
Liveness Index
building_now
A continuously updated index of every registered AI service: is it online, how fast does it respond, what HTTP status does it return. Pure measurement, verified by consensus.
- ✓ miners probe health + latency + status
- ✓ consensus from distributed monitors
- ✓ deterministic scoring (no LLM needed)
- ✓ scoped to known directories (MPP.dev, HS58)
- ✓ mathematically verifiable benchmark
V2
Anomaly Detection
next_phase
Layers LLM-powered analysis on top of V1's monitoring data. Validators use LLMs to detect quality degradation and diagnose failure patterns — the marketplace provides hidden ground truth to score their accuracy.
- ◌ LLM-Analyst detects quality anomalies
- ◌ Diagnosis Game: validators compete to explain failures
- ◌ Oracle Score: public validator reputation
- ◌ ground truth hidden from validators
- ◌ meta-competition: validators compete on LLM quality
02 // bit_more_info
How It Actually Works
The four roles, the probe mechanism, and how consensus works.
03 four_roles
Clean Separation of Concerns
Four distinct roles, no overlap. Miners never provide services. Providers never interact with Bittensor. Conflict of interest is eliminated by design.
Provider
Sells AI services — LLMs, APIs, tools. Exists on the open internet.
earns → USDC from agents
bittensor: never
Miner
Monitors providers neutrally. Probes health, latency, reachability.
earns → TAO emissions
bittensor: yes (UID)
Validator
Evaluates miner accuracy. Computes consensus. Sets weights.
earns → TAO dividends
bittensor: yes (Stake)
Agent
Uses services. Generates telemetry + optional session feedback.
earns → quality routing
bittensor: no
04 probe_mechanism
Miners Probe. Providers Don't Know.
Validators select random providers each round and dispatch probe synapses to all miners. Miners perform standard HTTP health checks. Providers can't distinguish probes from regular traffic.
minersproviders
miner_eu
prov_x
miner_us
prov_y
miner_asia
prov_z
← probe_synapse sent · response received →
probe_latency_ms
int — round trip
probe_reachable
bool — up or down
probe_status
HTTP status code
05 yuma_consensus
Consensus Detects Dishonesty
All miners probe the same targets. The validator computes consensus: what do most miners agree on? Outliers are penalized. Honest miners converge — and earn more TAO.
// miner_evaluation_roundconsensus_active
miner
accuracy
latency
weight
miner_001
42ms
0.23
miner_002
38ms
0.21
miner_003
51ms
0.19
miner_004
67ms
0.17
miner_005
—
0.01
rounds →
← weight convergence
Scoring formula:
0.4 × reachable_match + 0.3 × status_match + 0.3 × latency_scorehonest_miner
↑ 0.23
→
liar_miner
↓ 0.01
03 // deep_insights
Resistance to Gaming
Attack vectors, game-theoretic defenses, and self-reinforcing incentives.
06 anti_gaming
Known Attack Vectors
Incentives are aligned so that honest participation is the dominant strategy. Known attack vectors are addressed through consensus, economics, and architectural separation.
⚡ Miner Lies About Status
Consensus from distributed miners. A single liar becomes a statistical outlier — immediately penalized by Yuma.
neutralized
💰 Provider Bribes Miners
Would need to bribe the consensus majority. At scale, prohibitively expensive. Any partial bribe still loses.
neutralized
👥 Sybil Agents (Fake Traffic)
S, L, C measure real performance — sybils experience the same quality. Popularity is log-dampened.
neutralized
🎭 Fake Failure Reports
Agent has paid the provider (revealed preference). Sabotage costs money. Objective telemetry outweighs subjective reports.
neutralized
🔍 Provider Detects Miners
Miners probe via standard HTTP. Providers cannot distinguish a probe from a real agent request.
neutralized
⚖️ Validator Manipulation
Yuma Consensus: outlier validators lose dividends. Commit-reveal prevents weight copying.
neutralized
// nash_equilibrium
"For every participant, the highest expected return comes from honest behavior."
Miners maximize TAO by reporting truthfully — any deviation is detected by consensus. Validators maximize dividends by evaluating accurately — Yuma penalizes outliers. Providers maximize revenue by delivering quality — telemetry feeds the ranking engine. Agents maximize value by reporting honestly —
revealed_preference makes sabotage irrational.// deterministic_math
Deterministic Scoring — No LLM Required
✓ miner_scoring
0.4 × reachable + 0.3 × status + 0.3 × latency
✓ consensus
median / majority across all miner responses
✓ ema_smoothing
α × accuracy + (1-α) × previous score
✓ yuma_weights
standard Bittensor weight-setting protocol
Every formula is deterministic and reproducible. Any validator running the same inputs produces identical miner weights — no LLM, no subjective judgment. V2 will add optional LLM-powered anomaly detection, but the base scoring system stands on its own.
07 self_improvement
Reinforcing Incentive Loops
Each role's incentives create positive feedback loops. Better performance leads to more stake, more usage, and higher-quality monitoring over time.
QUALITY
FLYWHEEL
FLYWHEEL
Miner
accurate probes → more TAO → better infra
Provider
better service → higher score → more agents
Agent
best provider → good data → better scores
Validator
better LLM → subtler anomalies → more dividends
Feedback
failure reports → score drops → provider fixes
01
Miner Loop
More accurate probing earns more TAO. More TAO funds better infrastructure. Better infrastructure enables more accurate probing.
→ accuracy → TAO → infra → accuracy
02
Provider Loop
Better service generates better telemetry and positive feedback. Higher ranking attracts more agents. More revenue funds better service.
→ quality → ranking → agents → revenue → quality
03
Validator Meta-Competition
Validators with better LLMs detect subtler anomalies and diagnose failures more accurately. Higher Oracle Score attracts more stake.
→ LLM quality → anomaly detection → oracle_score → stake → dividends
04
Agent Feedback Loop
Agents report failures with structured reasons. Provider score drops. Provider addresses the issue. Score recovers.
→ drain_feedback(failure) → score ↓ → fix → score ↑