# AI verification proposal

A proposal built with the Proposal Explorer of the AI Verification Tech Map (https://trustbutveri.fyi/), from its records of 2026-10-09. Interactive version: https://trustbutveri.fyi/explorer/?mechanisms=M-0004,M-0017,M-0001

How to read it: a claim is something one party wants to verify about another's AI hardware or software. A mechanism is a general technique for verifying claims; it is "aimed at" a claim when that is its direct purpose, and "supporting" when it contributes without being aimed at it. A claim is addressed when a mechanism in the proposal is aimed at it and is not excluded by the filters; addressed does not mean verified, so check that mechanism's development status, security evidence and findings. Definitions: https://trustbutveri.fyi/about/methodology/ (roles, properties and findings) and https://trustbutveri.fyi/about/readiness/ (development status).

## Filters

Filters apply to mechanisms only and describe the setting the proposal is for.

None set. Every mechanism on the map was available.

## Overview

One row per mechanism, read from its record. Open failures: critical / significant / minor. The last three columns are the editors' reading of what the verifier sees. Findings are grouped as known failures, scope limitations and open questions. Only known failures count as failures. Counts are an inventory of published findings, not a risk score.

| Mechanism | Development | Security evidence | Prover | Attack testing | Hardware | Open failures | Weights | Inputs and outputs | Training data |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Zero-knowledge proofs of inference | Research demonstration | Published attack testing | Adversarial | Independent red-team | None | 0 / 0 / 0 | hidden | shown | not involved |
| Tamper evidence for verifier devices | Research demonstration | Published security analysis | Adversarial | Analysis | Retrofit device | 0 / 2 / 0 | not involved | not involved | not involved |
| Sampled inference recomputation | Operational use | Published security analysis | Adversarial | Analysis | None | 0 / 2 / 1 | depends | depends | not involved |

## Claims

No claims chosen.

## Mechanisms

### Zero-knowledge proofs of inference

A prover produces a cryptographic proof that an output came from running a committed model on a given input, without revealing the weights. ([Zero-knowledge proofs of inference](https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/))

- Assessment: mechanism family.
- Development: Research demonstration (legacy code R2), assessed for proving a language model's output follows from committed weights, against a cheating prover.
- Security evidence: Published attack testing. Independent evaluation: unassessed. Formal proof: unassessed. Deployment assurance: unassessed.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: none. Prover cooperation: required. Attack testing: independent red-team. Category: Cryptographic & computational.
- What the verifier sees: model weights hidden; inputs and outputs shown; training data not involved. The weights stay committed and hidden; the verifier knows each input and output it checks.

### Tamper evidence for verifier devices

Enclosures, seals and sensors that make physical interference with verification hardware visible, or that destroy the hardware's secrets when tampering occurs. ([Tamper evidence for verifier devices](https://trustbutveri.fyi/mechanisms/tamper-evidence-for-verifier-devices/))

- Assessment: mechanism family.
- Development: Research demonstration (legacy code R2), assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence.
- Security evidence: Published security analysis. Independent evaluation: unassessed. Formal proof: unassessed. Deployment assurance: unassessed.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: retrofit device. Prover cooperation: partial. Attack testing: analysis. Category: Off-chip devices & sensors.
- What the verifier sees: model weights not involved; inputs and outputs not involved; training data not involved. Protects verifier devices; it does not handle model data.

### Sampled inference recomputation

A verifier re-runs a random sample of an AI provider's logged queries on a trusted copy of the declared model and checks the outputs match. ([Sampled inference recomputation](https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/))

- Assessment: mechanism family.
- Development: Operational use (legacy code R3), assessed for checking untrusted workers' activations against the declared model, prompt and precision.
- Security evidence: Published security analysis. Independent evaluation: unassessed. Formal proof: unassessed. Deployment assurance: unassessed.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: none. Prover cooperation: required. Attack testing: analysis. Category: Cryptographic & computational.
- What the verifier sees: model weights depends; inputs and outputs depends; training data not involved. Recomputation needs the weights and sampled requests inside the checking environment. For closed models, the record describes a trusted, confidential environment; disclosure to the verifier depends on that boundary.


## Properties

**Operational use**

- Sampled inference recomputation: Operational use (legacy code R3), assessed for checking untrusted workers' activations against the declared model, prompt and precision

**Built for an adversarial prover**

- Zero-knowledge proofs of inference
- Tamper evidence for verifier devices
- Sampled inference recomputation

**No new hardware needed**

- Zero-knowledge proofs of inference
- Sampled inference recomputation


## Attack testing

Attack testing records published testing for this use. It does not by itself show independent review, a formal proof or that a deployed system is secure.

**Testing history**

- Zero-knowledge proofs of inference: Independent red-team
- Tamper evidence for verifier devices: Analysis
- Sampled inference recomputation: Analysis


## Limits

**Open significant failures**

- Seals are often defeated with simple methods (known failure, demonstrated attack, in Tamper evidence for verifier devices; https://trustbutveri.fyi/mechanisms/tamper-evidence-for-verifier-devices/evidence/flaws/1/) [19][20]. Mechanism-class evidence. Published defeats of general security seals. They warn about proposed verifier-device seals, but do not demonstrate defeat of an AI verification enclosure or sensor. In 1996 a Los Alamos vulnerability assessment defeated all 94 security seals it examined, with 132 defeats in total, using rapid, inexpensive, low-tech methods. It found that seal cost did not predict security. In 2001 Johnston reported that high-tech seals are often easier to defeat than low-tech ones.
- Attack classes outside published models (known failure, open question, in Tamper evidence for verifier devices; https://trustbutveri.fyi/mechanisms/tamper-evidence-for-verifier-devices/evidence/flaws/3/) [13][14][21]. Mechanism-class evidence. The radio compensation result is emulated using measured channel data under a known-reference attacker model. It is not a physical bypass demonstration against an AI verifier enclosure. The authors of the batteryless cover say they cannot assess chemical-solvent attacks, which exceed their expertise, and deem cover removal impractical. Anti-Tamper Radio's reference can drift as the environment or measurement system ages; the authors suggest gradually renewing the reference. A 2025 follow-up by some of the same authors shows, by emulation on measured channel data, that an attacker who knows the reference channel and the needle's effect on it could inject a signal that cancels the change caused by a needle insertion. It proposes a reconfigurable intelligent surface that randomizes the channel as a countermeasure.
- Tolerance for numerical noise leaves a covert channel (known failure, demonstrated attack, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/1/) [22][30][33]. Schemes that accept approximate matches can put an upper bound on an adversary's covert bandwidth, but they cannot close the channel. The weight-exfiltration detector cut exfiltratable information to under 0.5%, not to zero, on a 30-billion-parameter mixture-of-experts model under benign prompt traffic. Its authors called the channel's size under adversarial prompts an open empirical question. An independent study showed that an adversary who controls the prompts roughly doubles the bits leaked per token. Across six models, that cut the slowdown from 146–254 times under benign prompts to 60–118 times. The attack widens the exfiltration bound. It does not target the check that outputs match the declared model.

  Related mechanism: Deterministic and bit-exact inference (R3, not in the proposal). Bit-exact inference would remove the numerical tolerance if exact replay can be deployed with the required weights and configuration.
- Some inference optimizations are not covered (known failure, theoretical argument, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/3/) [23][24]. TOPLOC's authors state that it cannot detect speculative decoding in which a cheaper model does the decoding. They did not test whether it distinguishes types of key-value (KV) cache compression. DiFR was evaluated only on sampling from a single model. Its authors sketch an extension to one speculative-decoding algorithm but do not test it.

**Open minor failures**

- Mixed hardware widens the honest baseline (known failure, open question, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/4/) [23]. When honest reference runs span different GPU types, the spread of benign scores grows. In DiFR's tests on Qwen3-30B-A3B, pooling A100 and H200 runs left Token-DiFR unable to separate the two smallest tested changes, a temperature of 1.1 instead of 1.0 and a simulated top-2 sampling bug, at the target false-positive rate, while cross-entropy separated them. Matched provider and verifier environments, or pooling that weights rare large deviations, restored detection.

**Scope limitations**

- The proof covers a fixed-point approximation, not the floating-point model (scope limitation, open question, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/1/) [1][5][7][11]. Current ZK inference systems prove a quantised version of the network. zkLLM scales values by 2^16 and reports small perplexity changes. Attestable reports quantising matrix multiplications to 8-bit integers while proving other operations in floating point. A verifier therefore learns about the proof-friendly variant, and must separately accept that this variant is the declared model. Trail of Bits built a ResNet-18 backdoor that is dormant in the full-precision model and active after ezkl's quantisation; whether it persists through proving was left for further investigation. A verification system design notes that ZKPs can emulate floating-point operations. Rounding makes floating-point results depend on summation order, so bit-for-bit replay of an accelerator's results needs its original reduction tree. The report calls emulating that tree inside a ZKP an open, intricate problem and asks what it would cost.
- A proof speaks only for the computations that were proven (scope limitation, theoretical argument, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/2/) [12]. Attestable writes that "a proof of some computation is not a proof of all computation", and that a proof cannot discover a datacenter that was never declared. Proofs of inference do not by themselves show that no other workload ran on the same or other hardware.

  Related mechanism: Proofs of useful work for capacity accounting (R1, not in the proposal). The record names proof-of-work accounting as the kind of compute accounting needed to show that proven inference was the only work done.
- The model architecture is disclosed (scope limitation, theoretical argument, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/3/) [1][3]. ZKML "requires that the model architecture (but not weights) is revealed", and zkLLM assumes a publicly known model structure. Architecture can be commercially sensitive.
- Proofs do not bind computational effort (Hollow-LLM) (scope limitation, demonstrated attack, in Zero-knowledge proofs of inference; https://trustbutveri.fyi/mechanisms/zk-proofs-of-inference/evidence/flaws/4/) [10]. Researchers at the University of Southern California show that a proof of inference certifies that an output is consistent with committed weights under the declared architecture, but not how much computation produced it. In their Hollow-LLM attack, a provider keeps the declared architecture and parameter count but commits to "ghost weights". Some layers pass their inputs through unchanged, and wide layers carry the signal in a small subspace, so a much smaller inner model does the real work. The ghost weights satisfy the verification circuit and yield valid proofs.

  The authors ran the attack with the proof procedure of zkGPT, a separate ZK inference system, on a 6-layer, 512-dimensional transformer declared as up to 12 layers and 1,024 dimensions. Outputs were identical to the inner model's, and serving cost stayed at the inner model's level. An honest model of the declared size cost 2.4 times as much to prefill and 3.1 times as much to decode. Proving cost still grew with the declared architecture.

  The authors note that results may be served before any proof, with the provider building the witness only when a call is selected for audit. They describe their constructions as "compatible with state-of-the-art zkLLM pipelines", and state that the attack does not imply a flaw in the proof system itself. They propose challenge-based audits and ablation tests, which raise the cost of cheating but give no guarantee.
- Security depends on inspection protocols (scope limitation, theoretical argument, in Tamper evidence for verifier devices; https://trustbutveri.fyi/mechanisms/tamper-evidence-for-verifier-devices/evidence/flaws/2/) [18][20]. Mechanism-class evidence. An inspection and protocol requirement drawn from safeguards and enclosure studies, not a reported break of a deployed AI verifier. Johnston argues that a seal is no better than the protocols for using it, and that inspectors are usually given little useful information on how to detect tampering. The Sandia survey notes that larger enclosures are hard to inspect fully and that sensor data must be authenticated.
- Only recorded traffic is checked (scope limitation, theoretical argument, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/2/) [22][34]. Recomputation checks that recorded, declared workloads are correct. It cannot show that the record is complete. The published schemes do not cover hidden workloads run on the same compute, or substituted work. Rinberg et al. say their exfiltration-detection scheme cannot stand alone.

  Related mechanism: Network taps and certifiers (R1, not in the proposal). Taps copy and hash all traffic on the monitored links, which bears on whether the traffic record is complete. They do not show what else ran on the same chips.


## Possible additions

Mechanisms on the map, not in the proposal, that the records connect to an unaddressed or partly addressed claim, an open failure or a dependency. Pointers, not recommendations: each brings its own readiness level and findings, and none is claimed to close a failure.

- **Deterministic and bit-exact inference** (Operational use (legacy code R3), assessed for reproducing open-model inference from receipts in Gensyn's information-market service)
  - Bears on the open significant failure "Tolerance for numerical noise leaves a covert channel" in Sampled inference recomputation. Bit-exact inference would remove the numerical tolerance if exact replay can be deployed with the required weights and configuration.
- **Network taps and certifiers** (Proposed (legacy code R1), assessed for committing a complete record of cluster traffic, so declared inference can be checked)
  - Sampled inference recomputation waits on it: In tap-based retrofit designs, recording all inference traffic needs network taps and recomputation servers that can ingest it, in the worst case one recomputation-server network interface per inference front-end interface.
- **Proofs of useful work for capacity accounting** (Proposed (legacy code R1), assessed for bounding the spare capacity of declared hardware that could run training)
  - Zero-knowledge proofs of inference waits on it: Showing that proven inference was the only work done needs a compute-accounting mechanism such as proof-of-work accounting, which is only proposed.


## Dependencies

**Blockers**

- Zero-knowledge proofs of inference: Proving takes about 13 minutes (803 seconds) per 2,048-token forward pass of a 13B model on one A100, and a verification system design calls the overhead heavy. (performance & compatibility) [1][11]
- Zero-knowledge proofs of inference: ZKML and zkLLM prove fixed-point arithmetic, and a verification system design calls emulating an accelerator's original floating-point reduction tree inside a zero-knowledge proof, which bit-for-bit replay needs, an open and intricate problem whose cost is also unsettled. (performance & compatibility) [1][3][11]
- Zero-knowledge proofs of inference: zkLLM's code is unaudited, interactive and archived; the one audited ZK inference library, ezkl, had high-severity circuit soundness bugs before its fixes. (adversarial validation) [2][7]
- Zero-knowledge proofs of inference: Showing that proven inference was the only work done needs a compute-accounting mechanism such as proof-of-work accounting, which is only proposed. (coverage & hidden compute; waits on Proofs of useful work for capacity accounting) [12]
- Tamper evidence for verifier devices: No tamper-evident enclosure has been designed for AI verifier hardware at retrofit scale. (hardware trust) [11]
- Tamper evidence for verifier devices: Battery-backed designs add bulk, limit operating temperature (+10 °C to +35 °C for the IBM 4765) and complicate transport. (performance & compatibility) [14]
- Tamper evidence for verifier devices: Active monitoring needs power, and visual inspection of large enclosures faces access limits. (access & governance) [18]
- Tamper evidence for verifier devices: No evaluation has been published in the AI verification setting. (adversarial validation) [11]
- Sampled inference recomputation: In tap-based retrofit designs, recording all inference traffic needs network taps and recomputation servers that can ingest it, in the worst case one recomputation-server network interface per inference front-end interface. (coverage & hidden compute; waits on Network taps and certifiers) [29][35]
- Sampled inference recomputation: In retrofit designs, the recomputation server must sit inside the prover's data centre, possibly under the prover's physical control, and still be protected from a compromised provider, which Amodo rates 'not on track'. (hardware trust) [22][29]
- Sampled inference recomputation: No independent red-team of a recomputation consistency check has been published (the one independent attack study targets the weight-exfiltration bound), and Amodo rates recomputation red-teaming 'not started'. (adversarial validation) [29][30]
- Sampled inference recomputation: Tolerance-based checks need calibration on trusted hardware and exact knowledge of the provider's sampling procedure, and in one prototype a sampling-implementation mismatch produced large spurious differences. (performance & compatibility) [23][28]
- Sampled inference recomputation: The verifier needs the model weights, so checking a closed-weights model requires a trusted, confidential recomputation environment, which the retrofit designs place inside the prover's facility. (privacy & leakage) [11][23][34]


## What the verifier sees

- Model weights: shown by none; depends on the design for Sampled inference recomputation; hidden by Zero-knowledge proofs of inference; not involved in Tamper evidence for verifier devices; unspecified for none.
- Inputs and outputs: shown by Zero-knowledge proofs of inference; depends on the design for Sampled inference recomputation; hidden by none; not involved in Tamper evidence for verifier devices; unspecified for none.
- Training data: shown by none; depends on the design for none; hidden by none; not involved in Zero-knowledge proofs of inference, Tamper evidence for verifier devices and Sampled inference recomputation; unspecified for none.

## Implementations

- Zero-knowledge proofs of inference: [Attestable zero-knowledge inference prover](https://trustbutveri.fyi/implementations/attestable-zk-inference/) (R1, product); [EZKL](https://trustbutveri.fyi/implementations/ezkl/) (R2, product); [Low-trust AI compute verification system overview](https://trustbutveri.fyi/implementations/low-trust-compute-verification-system-overview/) (R1, proposed architecture); [zkLLM](https://trustbutveri.fyi/implementations/zkllm/) (R2, research prototype)
- Tamper evidence for verifier devices: [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture)
- Sampled inference recomputation: [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture); [DiFR (Divergence From Reference)](https://trustbutveri.fyi/implementations/difr/) (R2, research prototype); [Low-trust AI compute verification system overview](https://trustbutveri.fyi/implementations/low-trust-compute-verification-system-overview/) (R1, proposed architecture); [SASH confidential network logger](https://trustbutveri.fyi/implementations/sash-confidential-network-logger/) (R1, research prototype); [TOPLOC](https://trustbutveri.fyi/implementations/toploc/) (R3, open-source project)

## Sources

1. zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun et al. (2024). https://doi.org/10.1145/3658644.3670334
2. zkllm-ccs2024: code for zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun (2024). https://github.com/jvhs0706/zkllm-ccs2024
3. ZKML: An Optimizing System for ML Inference in Zero-Knowledge Proofs, B.-J. Chen et al. (2024). https://doi.org/10.1145/3627703.3650088
4. NanoZK: Privacy-Preserving Verifiable Inference for Large Language Models via Layerwise Zero-Knowledge Proofs, Z. Wang (2026). https://arxiv.org/abs/2603.18046
5. Proving LLMs at Scale, Attestable (2026). https://attestable.com/blog/proving-llms-scale
6. Verifiable evaluations of machine learning models using zkSNARKs, T. South et al. (2024). https://arxiv.org/abs/2402.02675
7. Zkonduit EZKL Security Assessment, F. Casal et al. (2025). https://github.com/trailofbits/publications/blob/master/reviews/2025-03-zkonduit-ezkl-securityreview.pdf
8. DeepProve-1: The First zkML System to Prove a Full LLM Inference, Lagrange Labs (2025). https://lagrange.dev/blog/deepprove-1
9. Lagrange-Labs/deep-prove (GitHub repository), Lagrange Labs (2026). https://github.com/Lagrange-Labs/deep-prove
10. Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference, C. Gong et al. (2026). https://arxiv.org/abs/2607.28884
11. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). https://intelligence.org/wp-content/uploads/2026/06/A-system-overview-for-near-term-low-trust-AI-compute-verification.pdf
12. Pacing AI Requires Proof, Attestable (2026). https://attestable.com/blog/pacing-ai-requires-proof
13. Anti-Tamper Radio: System-Level Tamper Detection for Computing Systems, P. Staat et al. (2022). https://ieeexplore.ieee.org/document/9833631/
14. Secure Physical Enclosures from Covers with Tamper-Resistance, V. Immler et al. (2019). https://tches.iacr.org/index.php/TCHES/article/view/7334
15. ImpedanceVerif: On-Chip Impedance Sensing for System-Level Tampering Detection, T. Mosavirik et al. (2023). https://eprint.iacr.org/2022/946
16. IBM 4765 Cryptographic Coprocessor Security Module: Security Policy, IBM Corporation (2012). https://csrc.nist.gov/csrc/media/projects/cryptographic-module-validation-program/documents/security-policies/140sp1505.pdf
17. PHYSEC SEAL: Change detection for maximum safety, PHYSEC GmbH (2026). https://www.physec.de/en/solutions/physec-seal/
18. Tamper-Indicating Enclosures, A Current Survey, H. A. Smartt & Z. N. Gastelum (2015). https://www.osti.gov/servlets/purl/1256541
19. Physical Security and Tamper-Indicating Devices, R. G. Johnston & A. R. E. Garcia (1996). https://www.osti.gov/servlets/purl/459707
20. Tamper Detection for Safeguards and Treaty Monitoring: Fantasies, Realities, and Potentials, R. G. Johnston (2001). https://www.nonproliferation.org/wp-content/uploads/npr/81john.pdf
21. Anti-Tamper Radio Meets Reconfigurable Intelligent Surface for System-Level Tamper Detection, M. S. Tabar et al. (2025). https://arxiv.org/abs/2503.14279
22. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). https://arxiv.org/abs/2511.02620
23. DiFR: Inference Verification Despite Nondeterminism, A. Karvonen et al. (2025). https://arxiv.org/abs/2511.20621
24. TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference, J. M. Ong et al. (2025). https://proceedings.mlr.press/v267/ong25a.html
25. PrimeIntellect-ai/toploc (GitHub repository), Prime Intellect (2025). https://github.com/PrimeIntellect-ai/toploc
26. INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning, Prime Intellect Team et al. (2025). https://arxiv.org/abs/2505.07291
27. adamkarvonen/difr (GitHub repository), A. Karvonen (2025). https://github.com/adamkarvonen/difr
28. Scaling Recomputation Inference Verification, Amodo Design (2026). https://amododesign.com/notes/2026-09-02-scaling-recomputation-inference-verification/
29. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). https://amododesign.com/ai-verification/plan-a-sitrep/
30. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). https://arxiv.org/abs/2608.23375
31. SYNTHETIC-2 Release: Four Million Collaboratively Generated Reasoning Traces, Prime Intellect (2025). https://www.primeintellect.ai/blog/synthetic-2-release
32. An Inference Verification Prototype — Stage 1, Amodo Design (2026). https://amododesign.com/notes/2026-06-29-inference-verification-prototype/
33. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). https://arxiv.org/abs/2606.00279
34. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). https://amododesign.com/notes/2026-06-23-verification-algorithms/
35. Verification Plan, R. Dean (2026). https://ai-2040.com/supplements/verification-plan