FAQs.

Provenance of execution: proof that this model produced this output, unforged, established by an independent committee recomputing the job and agreeing on the output hash. It never means the answer is true. It is reserved for consensus alone: a TEE result is attested, never verified.

An unbiasable VRF beacon seats each committee, so no one, including AIOS, chooses who checks a job. Each Miner commits its hash before any reveal, and the canonical result is the hash held by a work-weighted supermajority, more than two thirds. Weight comes from verified work, never capital. Exact sizes are calibrated with the security reviews before mainnet, so we do not publish guesses.

Onchain, in consensus: Native Models, small and integer-quantized (BitNet-class), toy scale by physics, with a credible ceiling around a 2 to 3 billion parameter ternary model. Off-chain: larger models. On the pinned integer AIOS Module they recompute byte-identically; beyond that they verify statistically, by sampled recomputation, never exact-hash.

No supermajority, no settlement. The job re-routes to a fresh committee while your escrow sits untouched, caller-owned and non-custodial. If routing is exhausted, you get a full refund. You are never charged for an answer the chain could not agree on.

No. Floating-point non-determinism across heterogeneous hardware makes exact consensus over frontier models impossible, ever. Onchain models are small integer-quantized Native Models. Larger off-chain models verify statistically, or run privately in attested enclaves without a correctness proof.

Yes, for grounded calls. A corpus registers on the Data Internet with a Merkle commitment, and the committee recomputes the retrieval itself. The proof carries a retrieval manifest, so you can check which corpus and which chunks grounded the answer. Verified data means provenance, never that the data is true, and agreed retrieval is not a claim of the best retrieval.

No. No standing stake and no token slashing. Register once (a burned fee at mainnet, 10,000 $AIOS at genesis and gliding down; free and allowlisted on the testnet), then earn per matched job, solo. A wrong result earns nothing; honest downtime and determinism bugs are never punished. Only proven fraud, a failed honeypot or liveness griefing ejects a miner ID.

Yes, one: block-ordering Producers post a small, equivocation-slashable producer-bond. It earns no yield, gates no inference, and is slashed only for double-signing. The mining layer stays stakeless, and AIOS:STAKING secures nothing.

After the proof testnet is live: a fixed-supply ERC-20 on Ethereum, launched with the working testnet as evidence. 300,000,000 forever, no mint function. 50% liquidity, 33.33% mining pool, 10% staking pool, 6.67% team with a 12-month cliff and 36 months linear. A flat 1% transfer tax with an immutable ceiling that only ratchets down.

The AIOS Module, the Neural VM, a BitNet-class reference Native Model, the registry, Miner software and committee consensus, on operator and allowlisted nodes. Free, allowlisted mining. No token and deliberately no points program. The point is one artifact: deterministic integer inference recomputing byte-identically, before anything trades on it.

No. Upgrades are proposed by a foundation multisig and adopted, or refused, by the Producers who actually run the chain, at a coordinated height. Owning $AIOS means the asset, your models and your keys, not a ballot. The multisig is operator-controlled at launch, and that is disclosed.

Mode B runs a model in a confidential GPU enclave on inputs nobody else sees and returns a signed attestation that the registered code ran untampered. An attestation proves execution, not correctness, and its trust root is chip vendors. So each job is value-capped, needs a multi-vendor quorum, and is badged attested, never verified.

Bridges are the most attacked surface in crypto, and ours is no exception. AIOS shrinks it: one asset, a simple lock-and-mint with the L1 as sole mint authority, supply checked across both homes every block, on an audited external provider. AIOS writes no bridge cryptography. A smaller surface is still a surface; it ships only after a dedicated security gate.

No. A proof shows which model ran and that the output is unforged. It does not stop a hostile page or prompt from steering an agent; prompt injection is a separate, unsolved problem. The apps answer it with agent defaults off, permission tiers and confirmation gates, not a badge.