CyAI: The Decentralized AI Built by Cysic

An open market for verifiable AI inference — moving from “trust a provider” to “trust an auditable protocol.”

Most developers reach LLMs through centralized APIs and simply trust that the provider runs the promised model, at the advertised precision, billing tokens honestly. That works while AI is a convenience layer. Once inference powers financial workflows, agents, and on-chain decisions, it isn’t enough — the next generation of AI infrastructure needs inference that is traceable, auditable, and verifiable, not just accessible.

CyAI is a decentralized AI inference market. Anyone can supply or consume LLM token inference in a way that is open, trust-minimized, and verifiable. It sits inside Cysic’s ComputeFi vision — turning compute into an open, settleable, verifiable resource — letting developers submit jobs, set budget and latency, and route to providers by price and quality.

The four roles

Role	Responsibility
Provider	Supplies inference via Worker nodes (one model per Worker); prices and competes freely.
Customer	Sends requests; selects on price, latency, model, and quality.
Witness	Randomly audits Provider behavior and collects evidence.
Verifier	Uses LDD verification to check whether execution matches the claimed model.

Pricing is a free market on both sides, with dual-sided incentives (Providers by score and trustworthiness, Customers for using the service) and caps that protect smaller players.

The trust problem

Open supply means untrusted execution. A Provider can cut cost by substituting a smaller model, downgrading precision (FP8/INT8 billed as BF16), serving selectively when it suspects inspection, or overbilling tokens — none of which produce obviously wrong output. A smaller or quantized model still returns text that reads as plausible, so a Customer inspecting only the final answer cannot tell whether the promised model actually ran. So: how does a Customer know the claimed model actually ran?

Verifiable inference with LDD

CyAI audits a small, unpredictable fraction of requests rather than every one — enough to catch large-scale cheating, since profitable cheating has to span much of a workload. The metric is Logit Distance Distribution (LDD): fix the token path and compare the Provider’s logits to a reference model at each step. The intuition is simple:

• Honest execution — differences come from benign numerical noise, so the distance distribution stays stable and concentrated.
• Substitution or aggressive quantization — the differences become systematic and show heavier tail behavior.

No model internals are exposed, no expensive proof rides on every request, and the audit layer adds negligible overhead to normal inference.

Customer → CyAI routing → Provider executes + returns audit witness
                         ↓
              Witness samples requests → Verifier computes LDD
                         ↓
      Network updates provider score / reward / penalty

What verification buys the network

1. Detect model substitution. The token-probability structure is hard to fake, so serving a smaller model surfaces as a statistical shift.
2. Catch precision downgrades. Charging for BF16 while serving FP8 produces a systematic deviation the audit captures.
3. Constrain token overbilling. Tokens are the unit of settlement, so auditability protects Customers and the reward system from overreporting.

CyAI vs. centralized API

	Centralized API	CyAI
Trust	One provider	Cryptographic, trust-minimized
Pricing	Fixed, opaque	Free market
Supply	Single data center	Distributed nodes
Incentives	None	Dual-sided + anti-monopoly caps
Inference trust	Unverifiable	LDD auditing

Built for the agent economy

AI is moving from chat interfaces to agentic workflows — agents that call models, use tools, and execute transactions continuously. An agent acting on a user’s behalf needs verifiable answers to four questions:

• Was the expected model actually used?
• Was the inference service reliable?
• Can the result be traced after the fact?
• If a dispute arises, is there an auditable evidence trail?

Cysic AI’s inference service, Agent Marketplace, and CyHost form a natural stack around this need: a unified inference layer for developers, open compute for Providers, and an auditing foundation underneath. The point is not more compute — it is verifiable compute.

When inference, auditing, settlement, and incentives close into one loop, AI inference becomes a verifiable compute service markets can price, protocols can govern, and users can trust.

Compute with us. Verify with us. Build trusted AI with us.