Platform

Architecture

Deep dive into the technical architecture of Platform Network - a distributed validator network for decentralized AI evaluation on Bittensor.

System Overview

Platform Network is Subnet 100 on Bittensor. It orchestrates multiple challenges, manages validator consensus, and distributes rewards to miners based on their performance.

Subnet 100Platform Network Architecture
Orchestration
Platform Serverchain.platform.network
PostgreSQLHTTP APIWebSocket
Challenges
Term Challenge60% Emission
Bounty Challenge40% Emission
Future ChallengesModular
Participants
ValidatorsEvaluate & Consensus
MinersSubmit Solutions
Blockchain
Bittensor ChainWeight Submission (~72min epochs)

Key Components

📡

Platform Server

Central orchestrator running at chain.platform.network. Manages challenge configurations, stores evaluation results, and coordinates validator consensus.

PostgreSQL DatabaseHTTP API (port 8080)WebSocket EventsBridge Proxy

Validators

Node operators who evaluate miner submissions and submit weights to Bittensor. Require minimum 1000 TAO stake for participation.

Docker ContainersPBFT ConsensusStake VerificationAuto-Update

Challenge Modules

Modular Docker containers that define evaluation logic. Each challenge has its own scoring mechanism and emission allocation.

Docker IsolationCustom ScoringMechanism IDsPlug-and-Play

Consensus Mechanism

Platform uses PBFT (Practical Byzantine Fault Tolerance) for validator consensus, ensuring the system remains correct even with faulty nodes.

PBFT Threshold
threshold = 2f + 1 = ⌊2n/3⌋ + 1

For n validators with maximum f faulty nodes.

Byzantine Tolerance
f = ⌊(n-1)/3⌋

System tolerates up to f faulty validators.

Consensus Flow

1Propose

Leader broadcasts action

2Vote

Validators verify & vote

3Commit

Apply if threshold met

Score Aggregation

Miner scores are aggregated across validators using stake-weighted averaging with outlier detection.

Stake-Weighted Average
m = Σ(σv × sv,m) / Σ(σv)

Where σv is validator stake and sv,m is the score.

Confidence Score
confidence = exp(-σ / τ)

Low variance = high confidence. τ = 0.1 sensitivity scale.

Outlier Detection (MAD-based)

Validators with anomalous scores are excluded using Modified Z-Score:

Mi = 0.6745 × (xi - median) / MAD

Threshold: |Mi| > 3.5 = outlier

Weight Submission

At each epoch boundary (~72 minutes), validators submit weights to Bittensor based on aggregated scores.

1

Normalize Weights

wi = si / Σ(sj)
2

Apply Cap (50% max)

wicapped = min(wi, 0.5 × Σw)
3

Convert to u16

Wi = round(wi × 65535)
4

Submit to Bittensor

Per-mechanism weight submission

Security Model

💰

Stake-Based Security

1000 TAO minimum stake creates economic incentive for honest behavior.

🐳

Docker Isolation

All evaluations run in sandboxed containers with resource limits.

🔒

Signature Verification

All API requests signed with validator keys and verified.

🛡️

Outlier Detection

Anomalous validators excluded from score aggregation.

Rate Limiting

100 msg/sec per peer, 5 connections per IP, 1h blacklist.

🔄

Auto-Update

Watchtower ensures all validators run identical versions.

Configuration Parameters

ParameterDefaultDescription
Min Validators3Minimum validators for valid score
Min Stake %30%Minimum stake percentage to count
Outlier Z-Score3.5Modified Z-score threshold
Weight Cap50%Maximum weight per miner
Grace Epochs10Epochs before decay starts
Decay Rate5%Decay per stale epoch
Max Burn80%Maximum burn percentage

Learn More