Thesis

AI runs on trust.

The case for proving it instead, with every risk stated.

CategoryMachine-intelligence L1
PostureMechanism-first
The BetPoI at mainnet grade
Risks Carried8, in full
01Trust

The trust problem

01 · The Problem

A black box returns text

You cannot check which model ran, whether the provider quietly downgraded it, or whether it ran at all. The trust root of the whole AI economy is a provider's word. That is tolerable while people read the output; it stops being tolerable when agents act on it.

  • Which model ran, or whether it was swapped for a cheaper one
  • Whether the inference ran at all, or was cached, truncated or forged
  • What you were actually billed for
02The Bet

Proof of Inference as consensus

The pieces are reachable and each has lineage. What nobody has done is fuse them into a chain where verified inference is the economic engine.

  1. 01 / Determinism

    Pin the runtime

    The AIOS Module: pure integer, fixed point, batch-invariant. The same bytes out on every conforming node.

  2. 02 / Committee

    Recompute independently

    An unbiasable VRF seats a Miner committee. Each Miner recomputes on its own hardware.

  3. 03 / Agreement

    Commit, then reveal

    Hashes are committed before any reveal, so copying the majority is closed off. The work-weighted >2/3 hash is canonical.

  4. 04 / Settlement

    Proof on the block

    Miners on the canonical hash are paid; a wrong result earns nothing. The call settles with its proof.

03Four Pillars

Four pillars in one proof

AI is models plus data plus compute. Each pillar exists somewhere on its own; the fusion is the product.

Pillar / 01

Chain-state as model

Small integer-quantized models live in chain state and run deterministically in consensus. Toy scale by physics, labelled that way.

Pillar / 02

Blockspace as inference

Jobs are verified by committee consensus and committed by a work-gated producer set. Buy verified inference the way you buy gas.

Pillar / 03

The Model Internet

Every model and adapter gets an address, a price, an owner and a verification mode bound to Proof of Inference.

Pillar / 04

Data as context

Corpora register as owned, priced assets and the committee recomputes the retrieval, so an answer proves which corpus and which chunks grounded it.

04Posture

Mechanism first

We Lead With

Verified by consensus.

  • The mechanism: inference clears by independent agreement
  • The category: the machine-intelligence Layer 1
  • Provenance of execution, stated as exactly that
  • The wedge: the verification binding, never being first
We Never Say

What it cannot prove.

  • A frontier chatbot runs onchain
  • Verified means correct
  • Verified data means the data is true
  • There is zero stake anywhere
  • Verification solves agent security
05Risks

All eight risks

A thesis that hides its failure modes is an advertisement.

Risk / 01

Proof of Inference is unproven at scale

No project makes verified inference the consensus mechanism in production. The best precedent is a single-developer testnet whose throughput numbers failed independent verification. Throughput at our target model size is an open benchmark, not a promise.

Risk / 02

Token-first bleed, mitigated, not zero

The launch is sequenced proof-first: the testnet before any token, then an ERC-20 with the live testnet as evidence. The residual risk is real: a testnet is not mainnet, the token trades before native fee revenue, and proof of mechanism is not proof of demand.

Risk / 03

Model registries are not greenfield

Onchain model registries exist elsewhere. AIOS claims no first-mover advantage; the wedge is the verification binding, that a call is proven to have run the registered model.

Risk / 04

Frontier LLMs cannot run in consensus, ever

Floating-point non-determinism across heterogeneous hardware makes exact consensus over frontier model math impossible. Onchain models are small and integer-quantized; an FP16 model verifies statistically, never exact-hash. The pitch is never a chatbot onchain.

Risk / 05

Bridge risk

The permanent Ethereum to L1 bridge is a high-value target, and bridges are the most exploited component in crypto. AIOS writes no bridge cryptography, uses an audited external provider and keeps it single-asset. That reduces the surface; it does not remove it.

Risk / 06

No-stake mining is a weaker base than staking

Security rests on an honest majority of committed work. The only at-risk value is a burned registration fee: it prices a Sybil fleet, not each committee seat. Bounded, not closed, by value caps, a burn floor, raised honeypots and per-entity caps. The fallback is bonded mining.

Risk / 07

The private mode shifts trust to chip vendors

Mode B proves attested execution, not correctness, and its trust root moves to chip vendors. Quote-forging attacks with physical access are demonstrated, not hypothetical, so: a multi-vendor quorum, a hard value cap at submit, never in consensus, never called verified.

Risk / 08

Verifiable retrieval is unproven engineering

The data pillar holds only if embed and retrieve are byte-deterministic across hardware. Float embeddings and approximate indexes are not; the integer-embedding, exact-kNN path is a target, not a shipped guarantee. Verified data means provenance, never that the data is true.

Also true. There is one standing security stake, the producer-bond: small, equivocation-slashable, on ordering nodes only. And verified inference does not solve agent security; prompt injection is a separate, unsolved problem.
Weights Are Programs · Inference Is Consensus

Read the whole case

The thesis is the why. The mechanism pages are the how, including where it can fail.