# 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-0009,M-0001&implementations=M-0001:I-0002

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 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Hardware-enabled guarantees (flexHEG) and guarantee processors | Proposed | Published security analysis | Adversarial | Analysis | New chip design | 0 / 3 / 0 | hidden | hidden | hidden |
| Sampled inference recomputation / DiFR (Divergence From Reference) | Research demonstration | Published security analysis | Adversarial | Analysis | None | 0 / 1 / 1 | unspecified | unspecified | unspecified |

## Claims

No claims chosen.

## Mechanisms

### Hardware-enabled guarantees (flexHEG) and guarantee processors

A proposed add-on for AI chips: an auditable guarantee processor, sealed in a tamper-protected enclosure, that would check and enforce agreed rules on chip use. ([Hardware-enabled guarantees (flexHEG) and guarantee processors](https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/))

- Assessment: mechanism family.
- Development: Proposed (legacy code R1), assessed for checking and enforcing training-compute limits on chips, against adversaries up to states.
- 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: new chip design. Prover cooperation: required. Attack testing: analysis. Category: On-chip & hardware-enabled.
- What the verifier sees: model weights hidden; inputs and outputs hidden; training data hidden. The guarantee processor sees the chip's traffic inside a sealed enclosure and reports only whether rules were kept.

### Sampled inference recomputation

DiFR checks that an inference provider ran its declared model by comparing output tokens or activations with a trusted re-run using the same random seed. ([Sampled inference recomputation](https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/))

- Assessment: selected implementation [DiFR (Divergence From Reference)](https://trustbutveri.fyi/implementations/difr/).
- Development: Research demonstration (legacy code R2), assessed for checking that outputs match the declared model, precision and sampling settings.
- 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 unspecified; inputs and outputs unspecified; training data unspecified. This Explorer has no asset-specific exposure assessment for this implementation. Check its source and deployment assumptions.


## Properties

**Built for an adversarial prover**

- Hardware-enabled guarantees (flexHEG) and guarantee processors
- Sampled inference recomputation

**No new hardware needed**

- 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**

- Hardware-enabled guarantees (flexHEG) and guarantee processors: Analysis
- Sampled inference recomputation / DiFR (Divergence From Reference): Analysis


## Limits

**Open significant failures**

- State attackers can likely defeat current secure enclosures (known failure, theoretical argument, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/1/) [2][4]. The flexHEG authors write that "nation-state attackers can likely compromise the best current secure enclosures", and that the marginal cost of circumvention per device is hard to estimate. RAND similarly judges that anti-tamper measures "would not be insurmountable for a determined and well-resourced adversary", although they raise costs and can reveal tampering.
- Firmware-only retrofits rely on Secure Boot, which fault injection can bypass (known failure, theoretical argument, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/2/) [2]. Part II notes that the most common attack on Secure Boot replaces the firmware and applies a voltage glitch while the signature is being checked. It also notes that sophisticated actors may use microprobing or laser voltage probing to read key registers.
- FLOP accounting can be laundered through external data (known failure, theoretical argument, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/4/) [2]. Results of earlier or parallel workloads could be hidden in the "external data" fed to a device, which would falsify the total FLOP count unless the inputs are explained or time delays are imposed.
- Statistical tolerance leaves a covert channel (known failure, demonstrated attack, in DiFR (Divergence From Reference); https://trustbutveri.fyi/implementations/difr/evidence/flaws/1/) [8][13][15]. Statistical schemes can put an upper bound on an adversary's covert bandwidth, but cannot close it. In the companion exfiltration study, the detector cut exfiltratable information to under 0.5% under benign prompt traffic. It did not cut it to zero. Its authors called the channel's size under adversarial prompts an open empirical question. An independent study attacked that detector, which uses the same Gumbel-margin statistic. An adversary who controls the prompts roughly doubled the bits leaked per token. Across six models, this cut the slowdown from 146–254 times under benign prompts to 60–118 times. The attack widens the exfiltration bound. It does not target DiFR's check that outputs match the declared configuration.

**Family finding context**

- Context for DiFR (Divergence From Reference). Findings from the mechanism family appear here as context. They apply to an implementation only when its own record lists them, under the conditions stated there. 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/) [8][13][15]. 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.
- Context for DiFR (Divergence From Reference). Findings from the mechanism family appear here as context. They apply to an implementation only when its own record lists them, under the conditions stated there. Only recorded traffic is checked (scope limitation, theoretical argument, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/2/) [8][16]. 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.
- Context for DiFR (Divergence From Reference). Findings from the mechanism family appear here as context. They apply to an implementation only when its own record lists them, under the conditions stated there. Some inference optimizations are not covered (known failure, theoretical argument, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/3/) [7][17]. 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.
- Context for DiFR (Divergence From Reference). Findings from the mechanism family appear here as context. They apply to an implementation only when its own record lists them, under the conditions stated there. Mixed hardware widens the honest baseline (known failure, open question, in Sampled inference recomputation; https://trustbutveri.fyi/mechanisms/sampled-inference-recomputation/evidence/flaws/4/) [7]. 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.

**Open minor failures**

- Mixed hardware widens the honest baseline (known failure, open question, in DiFR (Divergence From Reference); https://trustbutveri.fyi/implementations/difr/evidence/flaws/2/) [7]. For Qwen3-30B-A3B, pooling honest runs across A100 and H200 GPUs and parallelism setups broadened the honest score distribution. Token-DiFR then failed 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 did. The authors report that matched provider and verifier environments, or pooling that weights rare large deviations, restore detection.

**Scope limitations**

- Many important rules cannot be checked on-chip (scope limitation, theoretical argument, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/3/) [1][3]. Malicious intent "is not a technical property observable on-chip", and misuse depends on what is done with a computation's results. A guarantee processor cannot easily tell whether a network is the whole system or one expert in a mixture-of-experts system. Part III judges that a fully local ruleset "may not be entirely feasible" for the same reason.
- Coverage stops at flexHEG-equipped chips (scope limitation, open question, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/6/) [1][3]. Motivated actors will always be able to use some compute that is not flexHEG-equipped. Recalling existing consumer GPUs would likely be impractical, and reaching perfect coverage, or conclusively proving that no secret government data centres exist, would be "practically quite difficult".

  Related mechanism: Chip registries and manufacturing records (R1, not in the proposal). Accounts for which chips exist and who holds them.

  Related mechanism: Remote detection of data centres (R1, not in the proposal). Looks for undeclared facilities that hold other chips.

**Open questions**

- Supply-chain diversion and hidden backdoors (open question, open question, in Hardware-enabled guarantees (flexHEG) and guarantee processors; https://trustbutveri.fyi/mechanisms/flexheg-guarantee-processors/evidence/flaws/5/) [2][3]. Components could be diverted before a guarantee processor is added, and backdoors could be introduced during design or manufacturing. Open-source designs and physical scans of randomly selected chips are proposed as countermeasures. Part III proposes international oversight of production and extensive testing of a random sample of finished devices.

  Related mechanism: Chip registries and manufacturing records (R1, not in the proposal). Records each chip's identity and owner from the fab onwards, which bears on diversion before a guarantee processor is fitted. It does not address hidden backdoors.
- Speculative decoding and multi-model sampling not evaluated (open question, open question, in DiFR (Divergence From Reference); https://trustbutveri.fyi/implementations/difr/evidence/flaws/3/) [7]. The algorithms and experiments cover sampling from a single LLM. Speculative decoding was not evaluated. The authors sketch an extension to one speculative-decoding algorithm, without experiments. They note that other variants would need modified verification and extra metadata.

**Not yet demonstrated**

- Hardware-enabled guarantees (flexHEG) and guarantee processors: Proposed (legacy code R1), assessed for checking and enforcing training-compute limits on chips, against adversaries up to states

**Need new chip designs**

- Hardware-enabled guarantees (flexHEG) and guarantee processors


## 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.

- **Tamper evidence for verifier devices** (Research demonstration (legacy code R2), assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence)
  - Hardware-enabled guarantees (flexHEG) and guarantee processors waits on it: State-level attackers who hold the hardware can likely compromise the best current secure enclosures.
- **Chip registries and manufacturing records** (Proposed (legacy code R1), assessed for a checkable record of which chips were made and who declared owning them)
  - Hardware-enabled guarantees (flexHEG) and guarantee processors waits on it: Governing all relevant chips depends on knowing where they are, through chip registries and detection of undeclared facilities.
- **TEE remote attestation for AI workloads** (Operational use (legacy code R3), assessed for showing which software ran to a party that distrusts the operator holding the hardware)
  - Hardware-enabled guarantees (flexHEG) and guarantee processors depends on it.


## Dependencies

**Missing prerequisites**

- TEE remote attestation for AI workloads (Operational use (legacy code R3), assessed for showing which software ran to a party that distrusts the operator holding the hardware), needed by Hardware-enabled guarantees (flexHEG) and guarantee processors
- Chip registries and manufacturing records (Proposed (legacy code R1), assessed for a checkable record of which chips were made and who declared owning them), needed by Hardware-enabled guarantees (flexHEG) and guarantee processors

**Blockers**

- Hardware-enabled guarantees (flexHEG) and guarantee processors: Integrated flexHEG needs substantial help from the accelerator manufacturer, and the authors estimate 3.7–7.9 years, from when the manufacturer starts work, for such hardware to displace other accelerators in frontier development. (access & governance) [2]
- Hardware-enabled guarantees (flexHEG) and guarantee processors: State-level attackers who hold the hardware can likely compromise the best current secure enclosures. (hardware trust; waits on Tamper evidence for verifier devices) [2][4]
- Hardware-enabled guarantees (flexHEG) and guarantee processors: Rival states would need to trust the design and manufacture of guarantee processors and enclosures, for example through open design, redundant processors from each side or oversight of production. (hardware trust) [1][3]
- Hardware-enabled guarantees (flexHEG) and guarantee processors: Restricting future rule updates would need a formal language for rules, which the authors judge most likely infeasible for early flexHEG versions. (protocol soundness) [1]
- Hardware-enabled guarantees (flexHEG) and guarantee processors: Governing all relevant chips depends on knowing where they are, through chip registries and detection of undeclared facilities. (coverage & hidden compute; waits on Chip registries and manufacturing records) [3]
- Sampled inference recomputation: The verifier needs the model weights, so outsiders cannot use the method to verify providers of closed-weights models. (privacy & leakage) [7]
- Sampled inference recomputation: The verifier must know and match the provider's sampling procedure, and in one prototype a sampling mismatch in a newer vLLM version produced large spurious logit differences. (performance & compatibility) [7][10]
- Sampled inference recomputation: No independent red-team of DiFR's consistency check has been published, Amodo rates recomputation red-teaming 'not started', and the one independent attack study targets an exfiltration detector built on the same statistic. (adversarial validation) [12][13]


## What the verifier sees

- Model weights: shown by none; depends on the design for none; hidden by Hardware-enabled guarantees (flexHEG) and guarantee processors; not involved in none; unspecified for Sampled inference recomputation.
- Inputs and outputs: shown by none; depends on the design for none; hidden by Hardware-enabled guarantees (flexHEG) and guarantee processors; not involved in none; unspecified for Sampled inference recomputation.
- Training data: shown by none; depends on the design for none; hidden by Hardware-enabled guarantees (flexHEG) and guarantee processors; not involved in none; unspecified for Sampled inference recomputation.

## Implementations

- Hardware-enabled guarantees (flexHEG) and guarantee processors: none on the map
- 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. Flexible Hardware-Enabled Guarantees for AI Compute, J. Petrie et al. (2025). https://arxiv.org/abs/2506.15093
2. Technical Options for Flexible Hardware-Enabled Guarantees, J. Petrie & O. Aarne (2025). https://arxiv.org/abs/2506.03409
3. International Security Applications of Flexible Hardware-Enabled Guarantees, O. Aarne & J. Petrie (2025). https://arxiv.org/abs/2506.15100
4. Hardware-Enabled Governance Mechanisms: Developing Technical Solutions to Exempt Items Otherwise Classified Under Export Control Classification Numbers 3A090 and 4A090, G. Kulp et al. (2024). https://www.rand.org/pubs/working_papers/WRA3056-1.html
5. Secure, Governable Chips: Using On-Chip Mechanisms to Manage National Security Risks from AI & Advanced Computing, O. Aarne et al. (2024). https://www.cnas.org/publications/reports/secure-governable-chips
6. Hardware-Enabled Mechanisms for Verifying Responsible AI Development, A. O'Gara et al. (2025). https://arxiv.org/abs/2505.03742
7. DiFR: Inference Verification Despite Nondeterminism, A. Karvonen et al. (2025). https://arxiv.org/abs/2511.20621
8. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). https://arxiv.org/abs/2511.02620
9. adamkarvonen/difr (GitHub repository), A. Karvonen (2025). https://github.com/adamkarvonen/difr
10. Scaling Recomputation Inference Verification, Amodo Design (2026). https://amododesign.com/notes/2026-09-02-scaling-recomputation-inference-verification/
11. Amodo-Design/Inference-Recomputation-Prototype (GitHub repository), Amodo Design (2026). https://github.com/Amodo-Design/Inference-Recomputation-Prototype
12. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). https://amododesign.com/ai-verification/plan-a-sitrep/
13. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). https://arxiv.org/abs/2608.23375
14. An Inference Verification Prototype — Stage 1, Amodo Design (2026). https://amododesign.com/notes/2026-06-29-inference-verification-prototype/
15. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). https://arxiv.org/abs/2606.00279
16. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). https://amododesign.com/notes/2026-06-23-verification-algorithms/
17. TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference, J. M. Ong et al. (2025). https://proceedings.mlr.press/v267/ong25a.html