# 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-0015,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 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Memory wiping and proofs of secure erasure | Proposed | Published security analysis | Adversarial | Analysis | None | 0 / 1 / 0 | not involved | not involved | not involved |
| 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

### Memory wiping and proofs of secure erasure

Overwriting a device's memory in a way a verifier can check, so that data from earlier, undeclared work cannot persist in memory the wipe reaches. ([Memory wiping and proofs of secure erasure](https://trustbutveri.fyi/mechanisms/memory-wiping-and-secure-erasure/))

- Assessment: mechanism family.
- Development: Proposed (legacy code R1), assessed for showing that no data from earlier work persists in memory the wipe reaches.
- 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: Isolation & system architectures.
- What the verifier sees: model weights not involved; inputs and outputs not involved; training data not involved. Overwrites memory with verifier-chosen data; it does not handle model data.

### 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**

- Memory wiping and proofs of secure erasure
- Sampled inference recomputation

**No new hardware needed**

- Memory wiping and proofs of secure erasure
- 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**

- Memory wiping and proofs of secure erasure: Analysis
- Sampled inference recomputation / DiFR (Divergence From Reference): Analysis


## Limits

**Open significant failures**

- Outside help during challenges (known failure, theoretical argument, in Memory wiping and proofs of secure erasure; https://trustbutveri.fyi/mechanisms/memory-wiping-and-secure-erasure/evidence/flaws/2/) [3][4]. Classic proofs of secure erasure assume the device is isolated during the protocol. Bursuc et al. relax this to a bound on how close a helper can be, enforced by round-trip times. In data centres, remote memory access has round trips of about 1–2 µs, against about 70–200 ns for local DRAM. The MIRI overview therefore says verification depends on ruling out RDMA by latency or physical disconnection.

  Related mechanism: Timed challenge-response and memory-occupation challenges (R2, not in the proposal). Timed challenges bound how far away a helper can be by how quickly it must answer.

  Related mechanism: Bandwidth limits and compartmentalization (R2, not in the proposal). Removing or capping links between groups of accelerators limits remote memory access during a challenge.
- Statistical tolerance leaves a covert channel (known failure, demonstrated attack, in DiFR (Divergence From Reference); https://trustbutveri.fyi/implementations/difr/evidence/flaws/1/) [10][15][17]. 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/) [10][15][17]. 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/) [10][18]. 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/) [9][19]. 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/) [9]. 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/) [9]. 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**

- Memory the wipe cannot reach (scope limitation, open question, in Memory wiping and proofs of secure erasure; https://trustbutveri.fyi/mechanisms/memory-wiping-and-secure-erasure/evidence/flaws/1/) [5][6]. Amodo's inventory of a GB200 system lists many memory stores beyond GPU HBM and host DRAM. It notes that SSD controller DRAM sits on a private bus that host commands cannot read or write, and that its optimized algorithm leaves 25 GiB of HBM unattested. It also asks how switch memory could be wiped.
- Gap between erased and total memory (scope limitation, theoretical argument, in Memory wiping and proofs of secure erasure; https://trustbutveri.fyi/mechanisms/memory-wiping-and-secure-erasure/evidence/flaws/3/) [3]. Bursuc et al. note that memory left between the erased region and the device's full memory could hold data, and that their bounds are tighter only against a restricted adversary.

**Open questions**

- 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/) [9]. 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**

- Memory wiping and proofs of secure erasure: Proposed (legacy code R1), assessed for showing that no data from earlier work persists in memory the wipe reaches


## 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.

- **Bandwidth limits and compartmentalization** (Research demonstration (legacy code R2), assessed for monitoring inter-node traffic with operator-run software on four GPUs)
  - Bears on the open significant failure "Outside help during challenges" in Memory wiping and proofs of secure erasure. Removing or capping links between groups of accelerators limits remote memory access during a challenge.
- **Timed challenge-response and memory-occupation challenges** (Research demonstration (legacy code R2), assessed for detecting whether a GPU is doing other work)
  - Bears on the open significant failure "Outside help during challenges" in Memory wiping and proofs of secure erasure. Timed challenges bound how far away a helper can be by how quickly it must answer.
  - Memory wiping and proofs of secure erasure waits on it: Timed challenges must exclude remote memory and other helpers.


## Dependencies

**Missing prerequisites**

- Timed challenge-response and memory-occupation challenges (Research demonstration (legacy code R2), assessed for detecting whether a GPU is doing other work), needed by Memory wiping and proofs of secure erasure

**Blockers**

- Memory wiping and proofs of secure erasure: Wipes take time: tens of minutes for a pod's volatile memory and hours for SSDs, displacing work. (performance & compatibility) [4][5][6]
- Memory wiping and proofs of secure erasure: Timed challenges must exclude remote memory and other helpers. (coverage & hidden compute; waits on Timed challenge-response and memory-occupation challenges) [3][4]
- Memory wiping and proofs of secure erasure: All memory stores in a system must be inventoried and wiped at the same time. (coverage & hidden compute) [5]
- Sampled inference recomputation: The verifier needs the model weights, so outsiders cannot use the method to verify providers of closed-weights models. (privacy & leakage) [9]
- 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) [9][12]
- 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) [14][15]


## What the verifier sees

- Model weights: shown by none; depends on the design for none; hidden by none; not involved in Memory wiping and proofs of secure erasure; unspecified for Sampled inference recomputation.
- Inputs and outputs: shown by none; depends on the design for none; hidden by none; not involved in Memory wiping and proofs of secure erasure; unspecified for Sampled inference recomputation.
- Training data: shown by none; depends on the design for none; hidden by none; not involved in Memory wiping and proofs of secure erasure; unspecified for Sampled inference recomputation.

## Implementations

- Memory wiping and proofs of secure erasure: [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture); [Low-trust AI compute verification system overview](https://trustbutveri.fyi/implementations/low-trust-compute-verification-system-overview/) (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. Verification Plan, R. Dean (2026). https://ai-2040.com/supplements/verification-plan
2. Secure Code Update for Embedded Devices via Proofs of Secure Erasure, D. Perito & G. Tsudik (2010). https://link.springer.com/chapter/10.1007/978-3-642-15497-3_39
3. Software-Based Memory Erasure with Relaxed Isolation Requirements, S. Bursuc et al. (2024). https://ieeexplore.ieee.org/document/10664348/
4. 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
5. Memory Wipes - Performance Analysis, Amodo Design (2026). https://amododesign.com/notes/2026-07-01-memory-wiping/
6. Improving Disk Wiping Speed for Memory Wipes, Amodo Design (2026). https://amododesign.com/notes/2026-09-14-disk-wiping-speed/
7. Amodo-Design/PoSE-Memory-Wiping (GitHub repository), Amodo Design (2026). https://github.com/Amodo-Design/PoSE-Memory-Wiping
8. Empirical Evaluation of Memory-Erasure Protocols, R. Gil-Pons et al. (2025). https://www.scitepress.org/Papers/2025/135548/135548.pdf
9. DiFR: Inference Verification Despite Nondeterminism, A. Karvonen et al. (2025). https://arxiv.org/abs/2511.20621
10. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). https://arxiv.org/abs/2511.02620
11. adamkarvonen/difr (GitHub repository), A. Karvonen (2025). https://github.com/adamkarvonen/difr
12. Scaling Recomputation Inference Verification, Amodo Design (2026). https://amododesign.com/notes/2026-09-02-scaling-recomputation-inference-verification/
13. Amodo-Design/Inference-Recomputation-Prototype (GitHub repository), Amodo Design (2026). https://github.com/Amodo-Design/Inference-Recomputation-Prototype
14. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). https://amododesign.com/ai-verification/plan-a-sitrep/
15. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). https://arxiv.org/abs/2608.23375
16. An Inference Verification Prototype — Stage 1, Amodo Design (2026). https://amododesign.com/notes/2026-06-29-inference-verification-prototype/
17. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). https://arxiv.org/abs/2606.00279
18. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). https://amododesign.com/notes/2026-06-23-verification-algorithms/
19. TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference, J. M. Ong et al. (2025). https://proceedings.mlr.press/v267/ong25a.html