Built by DARKSOL

Synthference

Synthetic inference rights - a general primitive for representing, routing, and settling future compute access.

What it is

Synthference is a concept and protocol direction for turning inference access into programmable rights over future compute instead of raw prepaid credits, provider-locked balances, or debt-backed service guarantees.

A Synthference position can express:

• compute class
• quality band
• latency band
• time window
• max deliverable units
• fallback rules
• settlement outcomes

Why

Today, compute access is usually packaged as:

• API keys
• provider credits
• fixed subscriptions
• marketplace resale of opaque balances

Those models are brittle, provider-specific, and hard to route across heterogeneous supply.

Synthference reframes compute access as a bounded, time-scoped claim on future inference capacity that can be:

• acquired
• routed
• rebalanced
• partially fulfilled
• gracefully degraded
• settled by outcome

Core idea

Instead of promising infinite service and introducing hidden insolvency or hard cutoffs, Synthference models compute access as bounded synthetic rights.

That means:

• no assumption of perfect real-time fulfillment
• explicit fallback behavior under stress
• portable rights across providers and networks
• cleaner secondary markets for future compute
• better agent-native control over inference policy

Primitive shape

A right may define:

• capacity type - text, image, video, embeddings, agent runtime, GPU-seconds
• priority - premium, standard, background
• latency - realtime, interactive, batch
• scope - provider set, model family, privacy or jurisdiction constraints
• window - start, expiry, or rolling epoch
• units - requests, tokens, jobs, frames, seconds
• fallbacks - acceptable substitutions if ideal execution is unavailable
• settlement - full, fallback, partial, expired, failed

Early framing

Synthference is intended as a general primitive, not a single-network product.

It should be usable across:

• centralized inference APIs
• decentralized node networks
• local/private clusters
• enterprise compute pools
• hybrid agentic routing systems

Initial thesis

The right abstraction for inference may not be:

• a token for API usage
• a prepaid provider balance
• or a debt-like service promise

It may be a synthetic right to future compute, with programmable routing and bounded settlement.

Spec

• Schema v0
• Request Envelope v0
• Execution Receipt v0
• Settlement State Machine v0
• System Modules v0
• Supplier Capability Taxonomy v0
• Capacity Assurance Model v0
• Proof Requirements v0
• Worked Examples v0
• Agent Execution Policy v0
• Issuance Paths v0
• Registry Event Model v0
• Right Packaging and Secondary Market Semantics v0
• Lifecycle Diagrams and Flow Guidance v0
• Synthference on Base Architecture v0
• Base Contract Surface v0
• Base Adapter and Finalization Model v0
• Onchain Right Representation v0
• RightRegistry Object Model v0
• RightRegistry Solidity Surface v0
• RightRegistry contract draft

Next

• tighten schema examples and add reference JSON fixtures
• decide whether package objects belong in the base schema or stay as registry extensions
• extend the current Hardhat validation lane with lifecycle, split, freeze, finalize, and transfer-gating tests
• prototype one concrete Base implementation slice (RightRegistry + SettlementAnchor)

Built with teeth. 🌑