Summary: RIFT is worth cataloging not as just another AI x crypto brand or agent toolkit, but as an onchain contest engine for AI inference. The important split in the public materials is between contest creation and funding, participant staking, commit-reveal submission, commit-reveal review, and final onchain reward distribution. That makes RIFT a useful comparison point for Ritual, Giza Protocol, Allora, and other crypto × AI systems because its main control surface is not general-purpose model execution, but stake-backed evaluation and reward routing around model outputs.
What it does:
Lets creators create and fund AI-inference contests on Solana with an onchain reward pool
Requires participants to join contests and stake before submitting outputs
Uses commit-reveal flows for both submissions and reviews rather than plain public posting
Supports contest finalization, reward claiming, stake withdrawal, and refunds through the protocol lifecycle
Burns 1% of gross funding onchain when reward pools are funded, with the remainder routed into the contest pool
Exposes a TypeScript SDK for direct app/backend integration and an agent toolkit with MCP/OpenClaw-compatible tooling layered on top
Frames current deployment posture explicitly as dumbnet, with real onchain contest flows live but without onchain TEE or processor-attestation verification yet
Key claims:
The strongest reason RIFT belongs in the corpus is the explicit decomposition of AI inference into a contest market: funders create reward pools, participants stake and submit outputs, reviewers commit and reveal scores, and settlement happens onchain.
The SDK README makes clear that commit-reveal is not a side feature. RIFT uses separate commit-reveal paths for both submissions and reviews, which makes evaluator privacy and later reveal timing part of the protocol surface rather than offchain ops policy.
The public materials are unusually candid about trust posture. Both the SDK and the agent kit describe current devnet and initial mainnet as dumbnet, and both explicitly say TEE / processor attestation verification does not yet exist onchain.
The 1% funding burn is analytically useful because it means contest creation is not only a coordination action but also a protocol-fee event encoded into funding itself.
The agent kit matters because it shows RIFT is not only targeting human app developers. It already treats backend-signer MCP tooling, commit-reveal blob preservation, and OpenClaw skill integration as first-class infrastructure for machine participants and operators.
RIFT cleared the bar because it makes AI-evaluation markets legible as a separate control plane beneath broader decentralized inference rhetoric: reward-pool funding, evaluator admission through staking, hidden submission/review phases, settlement timing, and the still-missing attestation layer all matter more than the generic AI branding.
Whitepaper: No canonical standalone RIFT whitepaper surfaced in this pass. The strongest primary materials were the official rifted-dev-kit repository and its SDK/agent-kit READMEs collected in ../whitepapers/rift-primary-sources-2026-05-11.md.