RIFT is not most usefully understood as a generic AI-agent or inference network. The public materials define it more narrowly and more clearly as an onchain contest engine for AI inference on Solana.
The protocol surface is explicitly split into contest creation, reward-pool funding, participant joining/staking, submission commit-reveal, review commit-reveal, and finalization/claims. That decomposition is the main reason it belongs in the corpus.
Commit-reveal is core to the design, not just an implementation detail. The SDK exposes separate prepare/commit/reveal flows, and the agent kit emphasizes that agents must preserve opaque blobs between those phases.
The protocol already exposes real onchain lifecycle actions — staking, funding, reward claims, stake withdrawal, refunds, and finalization — but the public docs are explicit that attested proof of inference is not yet secured onchain.
The current trust posture is unusually legible because the maintainers call both devnet and initial mainnet dumbnet. The same materials say processor verification and TEE attestation verification do not yet exist onchain.
Funding is itself part of the economics surface: 1% of gross contest funding is burned onchain before the net reward pool is credited.
The agent kit is analytically important because it shows RIFT is already building for machine operators and orchestrators, not just end-user apps. Backend-signer MCP integration and OpenClaw skill support are part of the public architecture.
Specific source signals worth retaining
The repo root README gives the cleanest protocol overview: create contests, fund reward pools, let participants stake and submit outputs, run commit-reveal reviews, then finalize/reward/withdraw on Solana.
The SDK README is the best source for the economic and lifecycle specifics: 1% funding burn, metadata URI model, typed validation, and the concrete module split across contests, participants, submissions, reviews, rewards, and query.
The agent-kit README is the best source for the machine-operator angle: backend-signer runtime model, MCP/OpenClaw integration, and commit-reveal blob preservation as an orchestration requirement.
The repeated dumbnet and missing-TEE-verification language is important enough to preserve verbatim at a high level because it sharply constrains how seriously to take the protocol’s current decentralized inference claims.
Why this entry belongs in the corpus
It adds a distinct AI-inference-evaluation branch that is neither a general execution network nor a pure model marketplace.
It gives the corpus a comparison point where the practical control surface is contest funding, evaluator admission, commit-reveal scoring, and settlement timing rather than semantic agent interfaces or training-data rights alone.
The durable insight is that decentralized AI inference can conceal a narrower and more legible mechanism: stake-backed contest settlement with explicit review and payout phases, plus a still-unfinished proof-of-inference trust layer.