Network Roles

NeuroShard nodes self-organize into three distinct roles based on their hardware capacity and stake.

The Three Roles

Driver (Layer 0)

The Driver is the entry point of the training pipeline.

Responsibilities

• Download and process Genesis Dataset shards
• Tokenize input text
• Compute token embeddings
• Initiate forward pass
• Send activations to Workers

Requirements

Requirement	Value
Memory	>=2GB (recommended >=8GB)
Bandwidth	High (downloads training data)
Stake	None required

Reward Bonus

• +20% (1.2x multiplier) on all rewards
• Compensates for data bandwidth costs

Verification

Drivers are held accountable through Genesis Shard Verification:

Shard ID = Hash(Node ID) mod Total_Shards

Any peer can:

1. Download the same shard
2. Recompute Layer 0 output
3. Compare hash with Driver's output
4. Report fraud if mismatch

Worker (Layers 1...N)

The Worker performs the bulk of neural network computation.

Responsibilities

• Receive activations from previous node
• Forward through assigned layers
• Send activations to next node
• Compute gradients for local layers during training

Requirements

Requirement	Value
Memory	Any (determines layer count)
Bandwidth	Moderate
Stake	None required

Reward Bonus

• +5% per layer held (capped at 100% = 2x)
• 10 layers = 1.5x
• 20 layers = 2.0x (max)

Layer Assignment

Workers are assigned layers based on available memory:

layers_assigned = calculate_layer_assignment(
    available_memory_mb=8000,  # 8GB
    architecture=current_architecture,
    safety_factor=0.6
)
# Returns: ~15-25 layers depending on model size

Validator (Last Layer)

The Validator is the exit point and consensus participant.

Responsibilities

• Hold the LM Head (output projection)
• Compute logits from final hidden states
• Calculate loss during training
• Initiate backward pass
• Verify PoNW proofs from other nodes
• Participate in stake-weighted consensus

Requirements

Requirement	Value
Memory	>=2GB
Stake	>=100 NEURO

Stake Requirement

Unlike Drivers and Workers, Validators must stake NEURO to be eligible. This prevents Sybil attacks on the consensus mechanism.

Reward Bonus

• +30% (1.3x multiplier) on all rewards
• +0.001 NEURO per proof validated

Validator Selection

When a proof needs validation, validators are selected using stake-weighted random selection:

Score = Stake x (1 - 0.3) + Random x 0.3 x MaxStake

This ensures:

• Higher stake -> higher selection chance
• 30% randomness -> small stakers still get opportunities
• No monopoly by largest stakers

Validator Slashing

Validators who vote against consensus are slashed at 2x the normal rate:

Normal Slash:    10 NEURO
Validator Slash: 20 NEURO (2x)

100% of validator slashes are burned (not given to reporters).

Role Assignment Flow

When a node joins, roles are assigned based on capacity:

# Step 1: Check capacity
is_high_capacity = available_memory > 8000  # 8GB
is_medium_capacity = available_memory > 2000  # 2GB

# Step 2: Check stake for Validator
has_validator_stake = staked_amount >= 100  # 100 NEURO

# Step 3: Assign roles
if is_high_capacity or (is_medium_capacity and num_drivers < 2):
    assign_layer_0()  # Become Driver

if has_validator_stake and (is_high_capacity or num_validators < 2):
    assign_last_layer()  # Become Validator

# Step 4: Fill remaining capacity with Worker layers
assign_remaining_layers()

Dual/Triple Roles

In early network stages (few nodes), a single node might hold multiple roles:

This is temporary — as more nodes join, layers are redistributed.

Role Economics

Earnings by Role

Role	Base Rate	Bonus	Example Daily (Active)
Driver	Training + Uptime	+20%	~50-75 NEURO
Worker (20 layers)	Training + Uptime	+100%	~80-120 NEURO
Validator	Training + Validation Fees	+30%	~60-100 NEURO
All Three	Combined	Combined	~120-200 NEURO

Optimal Strategy

1. Low Memory (<4GB): Pure Worker, maximize layers
2. Medium Memory (4-8GB): Driver or Worker
3. High Memory (>8GB): Driver + Worker
4. With 100+ NEURO stake: Add Validator role

Checking Your Role

Via Dashboard

Open http://localhost:8000/ and check the "Role" field.

Via API

curl http://localhost:8000/api/stats | jq '.has_embedding, .has_lm_head'

• has_embedding: true -> Driver
• has_lm_head: true -> Validator
• Otherwise -> Worker

Via Logs

[NODE] Node abc123... became a Driver (Layer 0)
[NODE] Node abc123... became a Validator (Last Layer)
[NODE] Assigned 24 layers: [0, 1, 2, ... 23]

Next Steps

• Training Pipeline — How training works
• Proof of Neural Work — PoNW verification
• Staking — How to become a Validator
• Reward Structure — Maximize earnings