Skip to main content
Skip to content
AI

ZK Proofs for AI Agents: Mobile Proving, Costs and Trusted Execution

Review mobile proving, model transformation, confidential execution and verification costs using DN’s evidence record. No invented framework performance scores.

ZK Proofs for AI Agents: Mobile Proving, Costs and Trusted Execution
Reading options: audio, text size & sections

What Matters

Choose the statement that needs verification before choosing the proof system or hardware. Mobile proving is possible for some workloads and unsuitable for others. A zero-knowledge proof, a trusted execution environment and public execution establish different properties. None independently proves that an agent's source data is true, its model is commercially useful or its proposed trade is authorized.

DN Evidence Block

Reviewed 13 November 2026. Primary documentation: Mopro's versioned mobile benchmarks, EZKL's verification workflow and IETF RATS architecture. DN has not run these proving systems, rented accelerators or measured battery use, proof latency, memory or gas. The worksheet below is an original evidence record, not a vendor ranking.

Correction: this revision retires the earlier unsourced framework efficiency scores, fixed prover-time estimates, universal low-cost verification claims and the assertion that quantization preserves over 99% accuracy. The URL and approved feature image remain unchanged.

The Signal: verification starts with a precise claim

A trade recommendation can involve several different claims: a named model produced a result; the result used the specified inputs; those inputs describe the relevant market; and the operator is permitted to act on the result. A mechanism addressing the first claim must not silently certify the other three.

EZKL describes a workflow that compiles supported ONNX models into circuits, creates proofs and verifies the designated computation. The practical evidence record should retain the actual model, settings, input commitment, verification key and result. A proof of a computation is not a general guarantee that the model's prediction is accurate or that the chosen data was appropriate. EZKL system documentation.

DN's thesis is that the valuable product is the link between the exact claim and its acceptance decision. “Uses cryptography” is too broad to support procurement. The reviewer needs to know which statement was checked, which artifacts it binds and which assumptions remain outside it.

Can a phone generate the proof?

Mopro's version 0.3 benchmark page reports device- and circuit-specific results, including completed proofs on mobile hardware and memory-related failures for some heavier configurations. It also distinguishes native and browser implementations and warns about compatibility limits. These are the toolkit authors' reported results, not DN reproductions or a benchmark of arbitrary AI models. Mopro performance and benchmarks.

That evidence rejects a blanket statement that mobile proving is impossible. It does not justify the opposite blanket statement that every agent can prove every model on a phone. Circuit shape, witness preparation, proof system, implementation, device memory and the application's operating constraints all belong in the test.

Record whether “iOS” means a physical phone, a simulator or a compatible application running on a Mac. Preserve cold and warm runs separately. If a published report selects the fastest run, do not relabel it as median latency or the time a typical user should expect.

For an actual acceptance test, choose representative inputs and a maximum permitted completion time before running the benchmark. Keep timeouts, unsupported operations, memory failures and rejected proofs in the results. Include the interaction needed to obtain inputs, prepare the witness, generate and transmit the proof, and verify it. Proving time alone is not end-to-end task time.

Separate quantization error from proof correctness

A constrained representation can differ from the original floating-point model. Decide which model and numerical tolerance the application promises. Then test the transformed model against an appropriate held-out dataset and evaluate decision errors, not only agreement on a convenient example.

A valid proof for the transformed computation does not establish a universal accuracy-retention percentage. A small numerical difference can be harmless for one decision and consequential near another decision's threshold. Preserve both the transformation settings and the application's acceptance rule.

Constructed example: a hypothetical policy allows an action only when a model output is at least 0.50. An original output of 0.501 and transformed output of 0.499 sit on different sides of that rule. This invented example illustrates why aggregate numerical similarity alone may not preserve a consequential decision. It reports no actual model performance.

TEE, proof or public execution: compare the trust boundary

DN claim-to-control comparison, not a platform ranking
ApproachEvidence to inspectSeparate unresolved question
Zero-knowledge proofStatement, circuit/model, public inputs, key and actual verifier resultInput truth, model fitness and authority to act
Remote attestation / TEEAttester, verifier, reference values, policy and freshnessWhich components and data paths lie outside the attested scope
Public executionExecuted code, inputs, chain state and applicable consensus assumptionsConfidentiality, external input quality and permitted use
Combined designEvidence binding each stage to the nextGaps introduced at handovers or alternate execution paths

IETF RFC 9334 separates attesters, verifiers and relying parties and explains the role of evidence, reference values and appraisal policy. An attestation result is therefore interpreted under a defined trust arrangement, not as an unqualified certificate for every output the machine produces. The RFC is an informational architecture. IETF RATS architecture.

DN recommends mapping what each party can see and change. A private execution provider may protect data from one observer while receiving plaintext itself. A verifier might accept an allowed software measurement without checking the economic quality of the task. A public record can expose the exact computation while revealing information the user intended to keep confidential.

These approaches can complement one another. No market-wide migration from proofs to TEEs, or from public to confidential execution, is established here. A procurement decision should follow the actual threat model and workload rather than a claim that one technology universally replaces the others.

Build a cost boundary that survives an invoice

Record successful and failed attempts, setup, proving compute, verification, data publication, transfer, storage and active review. Separate a reused setup cost from a charge paid for every proof. Compare the same accepted task and the same accounting period.

Illustrative gas arithmetic: 220,000 gas at an effective price of 20 gwei per gas is 0.0044 ETH for that gas component. No ETH/USD conversion is assumed. That amount excludes any separately charged data or service component and does not assert that a particular verifier consumes 220,000 gas.

Cost per accepted task = all included attempt and operating costs / accepted task count. If no task is accepted, the ratio is undefined, not free. For a mobile design, report battery and thermal behavior separately where measured. A low cash charge does not prove an acceptable user experience or deadline.

The previous parameter-count-to-latency estimator has been removed because a single linear multiplier cannot establish actual framework performance across different circuits and hardware. The replacement record preserves observations and unknowns instead of manufacturing a precise provider result.

How to use the DN Proof Execution Record

Define one task and one claim. Pin the circuit or model and the software versions. Record the device and runtime, public and private inputs, test population and every result. Preserve an unrun test as unrun. The local worksheet records your statements without checking files, generating proofs or connecting to a wallet.

DN Proof Execution Record

Use non-sensitive aliases. This local worksheet records references and gaps. It does not verify a proof, inspect a device or authorize a trade.

No current record.

Version 1.0.0. No network calls, wallet access or persistent storage. Creation time is not evidence observation time.

Frequently asked questions

Does zkML prove that an AI answer is true?

It can establish a specified computation under its proof assumptions. Input truth, prediction quality and authorization require separate evidence.

Is mobile proving impossible?

No blanket conclusion is justified. Published toolkit results show workload-specific success and failures; test the actual circuit and device.

Does attestation replace every proof?

No. Compare the exact claim, trust assumptions and excluded components. The mechanisms can address different parts of a workflow.

Does this article rank proving frameworks?

No. DN has not run a matched benchmark. The earlier unsourced scores and fixed timing estimator were retired.

Commercial tools and disclosure

For coordinating a non-production review workflow, the existing DN referral links to Make and Taskade remain available. Their official sites describe workflow and project capabilities; this review has not tested either as a proof-system integration. Check permissions, plan limits and ongoing costs before buying.

DN may earn compensation through these links. They do not affect the evidence or establish proving performance. Trading tokens, buying a hardware wallet or subscribing to tax software is not a prerequisite for using this free review record.

Methodology and corrections

Version 2.0, 13 November 2026. This is a primary-documentation review and original evidence workflow. Mopro observations remain author-reported. All numerical examples are fictional arithmetic; no proof, device or paid workload was tested by DN. Contract, model and deployment changes require a new review.

Send corrections with the exact version, claim and supporting evidence to press@decentralised.news. Do not send credentials or confidential witnesses.

Get the most talked about stories directly in your inbox

Join the Decentralised News briefing for independent crypto, DeFi and AI analysis. No spam, unsubscribe anytime.