Reputation System

InfraMind’s reputation system is a decentralized trust layer that governs how reliably a node is treated within the mesh. Unlike centralized infrastructure platforms where trust is pre-assigned or hidden behind SLAs, InfraMind’s reputation is fully transparent, composable, and measurable—backed by cryptographic proofs and protocol telemetry.

Every node accumulates a reputation score over time, computed from recent performance across several weighted dimensions. This score influences:

• Job routing priority

• Access to premium workloads (e.g. zkML, streaming agents)

• Escrow multipliers and payment time

• Inclusion in quorum-based or peer-dependent tasks (e.g. training cohorts, fallback validators)

Reputation is earned. It cannot be bought, faked, or delegated.

---

Scoring Dimensions

Each node is evaluated over a rolling window (e.g. 1,000 most recent jobs), with weighted metrics:

Metric	Description	Weight (%)
Success Rate	% of jobs completed without error or timeout	50%
SLA Compliance	Responsiveness to job assignment and deadline	30%
Latency Score	Percentile against global and regional averages	20%

✅ Success Rate

A job is considered successful if:

• Input was accepted

• Container returned output matching the declared output_schema

• A valid proof was submitted before timeout

Nodes below 85% success rate lose access to tiered jobs. Below 70% → slashing triggered.

⏱️ SLA Compliance

Nodes must acknowledge jobs within t_ack (typically 300ms), and return results within t_exec_max. Delays, ghosting, or frequent retries lower SLA compliance.

SLA = (jobs accepted & served within timeout) / (jobs assigned)

Scheduler may introduce grace windows for high-load nodes.

🚀 Latency Score

Latency is normalized within a percentile curve across all active nodes:

• Nodes in top 10% latency (fastest) → bonus multiplier

• Nodes below 25th percentile → deprioritized

• Nodes >500ms median → excluded from real-time workloads

Latency is computed based on proof submission timestamps relative to job assignment and includes container startup time.

---

Composite Reputation Score

Reputation is a composite score in the range 0.0 – 1.0, updated continuously:

Reputation = (0.5 × SuccessRate) + (0.3 × SLA) + (0.2 × LatencyScore)

Example:

Success Rate:   96.2% → 0.962
SLA:            92.8% → 0.928
Latency Pctl:   88%   → 0.88

Reputation = 0.5×0.962 + 0.3×0.928 + 0.2×0.88 ≈ 0.938

Nodes with reputation ≥ 0.95 are labeled Tier-One and routed premium jobs by default.

Nodes with reputation ≤ 0.70 are rate-limited and pushed to low-risk workloads or test environments.

---

Reputation Impact on Protocol Behavior

Protocol Behavior	Influence from Reputation
Job Eligibility	Minimum rep threshold for job tiers
Reward Boosts	Up to +30% payout for top-tier nodes
Escrow Slippage	Lower wait time for payment settlement
Slashing Tolerance	Soft grace period for high-rep nodes
Peer Inclusion	Used in multi-node training consensus

Escrow slippage means that low-reputation nodes may need to wait longer (1–2 epochs) before reward claims are released, to allow for late challenge disputes. High-reputation nodes have faster, often instant, payout cycles.

---

Viewing & Auditing Reputation

Each node exposes its current score:

infra reputation --node 0xABC...

Returns:

Node: 0xABC...
Reputation: 0.938
Success Rate: 96.2%
SLA: 92.8%
Avg Latency: 213ms
Last Updated: 1720301123

Nodes can also query their reputation across epochs:

infra reputation --history

All scores are publicly verifiable via:

https://explorer.inframind.host/nodes

---

Peer Trust for Training Jobs

In future releases (Phase 2+), InfraMind will introduce multi-node job quorums for training, distillation, and voting-based inference. Only nodes above a configurable reputation threshold will be included.

This guards against adversarial or poisoned compute in federated training systems.

Quorum trust score will weigh:

• Per-model success rate

• Staking lock-in

• Historical consistency

---

Decay & Recovery

Reputation decays slowly if a node is offline. Inactivity is not punished immediately, but staleness is factored in after ~72 hours of missed heartbeats.

Recovery is automatic:

• Serve jobs successfully → score rebounds

• Slashing or reputation drop resets the rolling window

• Delegators are notified on large reputation drops

---

Summary

Reputation is InfraMind’s native proof-of-trust. It determines how much the protocol should rely on you, how much you earn, and what jobs you’re eligible for. It's computed in real-time, stored transparently, and built only through provable action—no social weighting, no gatekeeping, no marketing layer. Just computation. Reputation, on InfraMind, is execution made measurable.