Hoster Policy

Rules, requirements, and enforcement. Read this before registering. Violations can result in reduced traffic, probation, or permanent ban.

Availability requirements

Your server must maintain consistent uptime to stay in the routing pool.

Scores recover automatically when your server performs well again. Health score uses exponential moving average — one bad period won't destroy your score permanently.

Model authenticity rules

You must serve exactly what you claim to serve. Violations are detected automatically and result in immediate suspension.

What is verified at registration

1. Your /health endpoint returns HTTP 200
2. Your /v1/chat/completions returns a valid OpenAI-format response
3. Your model does not self-identify as GPT-4, Claude, Gemini, or any commercial model
4. Your model produces a response fingerprint (checked for drift over time)

What is verified every 10 minutes

1. Consistency probes — 5 factual questions with known correct answers (math, geography, history). Must pass ≥40%.
2. Response shape validation — every response must have valid choices[0].message.content
3. Identity fingerprint — we track the model's "personality" over time. Sudden changes flag for review.

Prohibited model practices

• Proxying commercial APIs — registering an endpoint that forwards to OpenAI, Anthropic, Google, etc. Instant permanent ban.
• Misrepresenting model size/quality — claiming to serve Llama 3 70B while running a 7B model. Detected via latency profiling and output quality.
• Returning cached/static responses — serving the same response regardless of input. Detected by consistency probes.
• Running a broken model — a model that produces garbled, empty, or nonsensical outputs consistently. Auto-deactivated.
• Rate limiting NeuralGate traffic — throttling the gateway's requests while serving direct customers faster.

Blocked endpoint types

The following cannot be registered regardless of circumstances:

Enforcement tiers

Violations result in graduated enforcement based on severity and history:

Pending earnings are held during suspension. They are released if reinstated, forfeited if banned.

Trust score explained

Your trust score (0–100%) determines your share of traffic. It's a composite of:

All scores use exponential moving average (EMA) — recent performance matters more than old performance.

Example — health score after failures (starting from 1.0):

After the circuit closes and your server passes the probe, the EMA resumes from where it left off. A bad 10-minute window takes about the same time to recover from. Events older than ~25 minutes have less than 10% influence on the current score.

Acceptable use

• ✅ Open-weight models (Llama, Mistral, Gemma, Qwen, Falcon, etc.)
• ✅ Fine-tuned versions of open-weight models (must disclose in model alias)
• ✅ Quantized models (GGUF, AWQ, GPTQ) — just list the actual base model
• ✅ Multiple models on one server — register each separately
• ✅ Models you trained yourself on your own data
• ❌ Commercial models (GPT, Claude, Gemini, Command, etc.)
• ❌ Models with non-commercial licenses used commercially without permission
• ❌ CSAM, illegal content, or models fine-tuned to produce it

Appeals

If you believe a moderation action was applied incorrectly:

• Email abuse@computeshare.servequake.com with your hoster ID and the reason
• Include evidence: server logs, model card, hardware specs
• We respond within 48 hours
• Permanent bans for commercial API proxying are not appealable

Automated monitoring summary

All checks run automatically — no manual intervention needed to maintain good standing: