# AI verification proposal

A proposal built with the Proposal Explorer of the AI Verification Tech Map (https://trustbutveri.fyi/), from its records of 2026-10-08. Interactive version: https://trustbutveri.fyi/explorer/?mechanisms=M-0003,M-0014&implementations=M-0003:I-0011

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 readiness and open flaws. Readiness levels R0 to R4 describe one record's public evidence for its assessed use and are never combined. Definitions: https://trustbutveri.fyi/about/methodology/ (roles, properties and flaws) and https://trustbutveri.fyi/about/readiness/ (readiness levels).

## 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 flaws: critical / significant / minor. The last three columns are the editors' reading of what the verifier sees.

| Mechanism | Readiness | Prover | Attack testing | Hardware | Open flaws | Weights | Inputs and outputs | Training data |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Whole-workload recomputation (reproducible packets) / AI 2040 inference-only verification stack | R1 | Adversarial | None | Retrofit device | 0 / 3 / 0 | unspecified | unspecified | unspecified |
| Bandwidth limits and compartmentalization | R2 | Adversarial | Analysis | Retrofit device | 0 / 5 / 0 | not involved | not involved | not involved |

## Claims

No claims chosen.

## Mechanisms

### Whole-workload recomputation (reproducible packets)

A proposed retrofit that isolates data-centre inference units, taps their front-end traffic and recomputes random samples to check that only declared inference runs. ([Whole-workload recomputation (reproducible packets)](https://trustbutveri.fyi/mechanisms/reproducible-computation-packets/))

- Assessment: selected implementation [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/).
- Readiness: R1 Proposed, assessed for showing that retrofitted data centres run only inference.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: retrofit device. Prover cooperation: required. Attack testing: none. Category: Isolation & system architectures.
- 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.

### Bandwidth limits and compartmentalization

Capping or removing network links between groups of accelerators, so that serving within each group still works but large training across groups becomes far slower. ([Bandwidth limits and compartmentalization](https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/))

- Assessment: mechanism family.
- Readiness: R2 Demonstrated, assessed for monitoring inter-node traffic with operator-run software on four GPUs.
- Claims in this proposal: none of them.
- Threat model: adversarial prover. Hardware: retrofit device. 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. Caps traffic between groups of chips; it does not read the traffic's content.


## Properties

**Built for an adversarial prover**

- Whole-workload recomputation (reproducible packets)
- Bandwidth limits and compartmentalization


## Attack testing

Published attempts to break a system, including those that found failures. Testing history does not establish that open flaws are resolved.

**Testing history**

- Bandwidth limits and compartmentalization: Analysis


## Limits

**Open significant flaws**

- The recomputation server must be trusted (theoretical argument, in AI 2040 inference-only verification stack; https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/#flaw-1) [1][2][4]. The plan calls the integrity of the recomputation server an extremely important aspect, and its argument that sampling verifies all outputs assumes that the server's computations and outputs can be trusted. The companion page notes that the server sits inside the prover's facility, possibly under the prover's physical control, and that hardening it against integrity attacks needs significant research. Amodo rates recomputation-server security as not on track.
- Spare compute is not verified (theoretical argument, in AI 2040 inference-only verification stack; https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/#flaw-2) [1]. The plan states that it does not verify that spare compute is unused for unapproved workloads, because this seems very challenging. It relies instead on side-channel bounds and memory wipes, so that the only results that persist are verified inference outputs.
- A recomputation family degrades against prompt-controlling adversaries (demonstrated attack, in AI 2040 inference-only verification stack; https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/#flaw-3) [1][2][6][7]. The plan's companion page names DiFR among the recomputation schemes being tested. An independent study found that Gumbel-based inference verification, the family that includes Token-DiFR, leaks roughly twice as many bits per token when the adversary chooses prompts that disrupt the text's structure, across six models of 1 to 32 billion parameters. The slowdown it imposes on a hidden-signalling adversary falls from 146–254 times under benign prompts to 60–118 times. The attack weakens the bound on hidden information in outputs, which the plan relies on to keep undeclared results from leaving.
- Low-communication training reduces the bandwidth training needs (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-1) [9][13][14]. DiLoCo matched fully synchronous training on 8 workers while communicating 500 times less. Rahman writes that this family of methods theoretically allows large-scale training with less than 100 Mbps. Lucid includes these methods in its bounds, but notes that extreme activation compression, architectures with unusually small inter-layer widths, or modular paradigms could erode the margin.
- Operator control of pod routing collapses the bound (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-2) [9]. Lucid's analysis finds that if the operator can freely assign pods to routers, it could dedicate a whole cell of 100 or more pods to one pipeline stage. The bound then falls to about 90–220x uncompressed and as low as about 25x with compression. The proposed mitigation, auditor-controlled random assignment that is periodically re-randomized, has not been implemented.
- Undeclared local storage raises per-pod capacity (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-3) [9]. More memory or storage per pod helps an adversary. Lucid requires per-pod storage to be declared, capped and physically inspected.
- Training within one pod is not covered (open question, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-4) [9]. Lucid's bounds concern pre-training models larger than the pods are sized for. Training models that fit in one pod, fine-tuning and reinforcement-learning post-training within one pod are outside the modelled threat.
- Parallel scale-up switches are hard enforcement points (theoretical argument, in Bandwidth limits and compartmentalization; https://trustbutveri.fyi/mechanisms/bandwidth-limits-and-compartmentalization/#flaw-5) [12]. In GB200 topologies, GPUs reach GPUs in other nodes through NVSwitches without a NIC on the path. Amodo notes that limits are hard to enforce there because many switches work in parallel, so compromising one or two would bypass the limit.

**Family finding context**

- Context for AI 2040 inference-only verification stack; applicability depends on the finding's scope. Spare compute is outside the scheme (theoretical argument, in Whole-workload recomputation (reproducible packets); https://trustbutveri.fyi/mechanisms/reproducible-computation-packets/#flaw-1) [1][8]. The plan states that it does not verify that spare compute is not used for unapproved workloads, because this seems very challenging. Recomputation checks the correctness of declared work, not its completeness.

  Related mechanism: Proofs of useful work for capacity accounting (R1, not in the proposal). Proposed as one input to accounting for spare capacity on declared hardware.
- Context for AI 2040 inference-only verification stack; applicability depends on the finding's scope. Non-compliant work could be encoded inside compliant-looking packets (theoretical argument, in Whole-workload recomputation (reproducible packets); https://trustbutveri.fyi/mechanisms/reproducible-computation-packets/#flaw-2) [1]. The plan notes that an AI company might try to encode a non-compliant workload inside a workload that looks compliant on the surface.

**Not yet demonstrated**

- Whole-workload recomputation (reproducible packets): R1 Proposed, assessed for showing that retrofitted data centres run only inference


## Possible additions

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

- **Tamper evidence for verifier devices** (R2 Demonstrated, assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence)
  - Whole-workload recomputation (reproducible packets) waits on it: Checking that taps are correctly installed and stay in place at scale is not a solved problem, and hardening the recomputation server inside the prover's facility needs significant research.
  - Bandwidth limits and compartmentalization waits on it: Shaping devices and routing assignments must be trusted by both parties; Amodo has not yet fully analysed resilience to a compromised DPU.
- **Memory wiping and proofs of secure erasure** (R1 Proposed, assessed for showing that no data from earlier work persists in memory the wipe reaches)
  - Whole-workload recomputation (reproducible packets) waits on it: Memory wiping may use existing algorithms, but hardware testing is at an early stage.
- **Network taps and certifiers** (R1 Proposed, assessed for committing a complete record of cluster traffic, so declared inference can be checked)
  - Whole-workload recomputation (reproducible packets) waits on it: Passive optical taps work at 400G, but the 800G and 1600G line rates now arriving in data centres are undemonstrated.
  - Bandwidth limits and compartmentalization waits on it: The verifier must know that all traffic leaving a pod crosses the capped, monitored links.
- **Side-channel suppression for isolated facilities** (R1 Proposed, assessed for bounding physical covert channels out of a verified enclosure)
  - Whole-workload recomputation (reproducible packets) waits on it: There is no plan yet for quickly scaling side-channel defences on a frontier cluster; only early theoretical pieces exist.
- **Sampled inference recomputation** (R3 In production, assessed for checking untrusted workers' activations against the declared model, prompt and precision)
  - Whole-workload recomputation (reproducible packets) depends on it.


## Dependencies

**Missing prerequisites**

- Sampled inference recomputation (R3 In production, assessed for checking untrusted workers' activations against the declared model, prompt and precision), needed by Whole-workload recomputation (reproducible packets)
- Tamper evidence for verifier devices (R2 Demonstrated, assessed for detecting probing of proposed verifier hardware, using server and electronics prototypes as evidence), needed by Bandwidth limits and compartmentalization

**Shared foundations**

- Network taps and certifiers, relied on by Whole-workload recomputation (reproducible packets) and Bandwidth limits and compartmentalization
- Tamper evidence for verifier devices, relied on by Whole-workload recomputation (reproducible packets) and Bandwidth limits and compartmentalization

**Blockers**

- Whole-workload recomputation (reproducible packets): A fully reproducible inference stack needs substantial software and tooling, and per-packet network reproducibility may need considerable software, firmware and possibly hardware work. (performance & compatibility; waits on Whole-workload recomputation (reproducible packets)) [2]
- Whole-workload recomputation (reproducible packets): Passive optical taps work at 400G, but the 800G and 1600G line rates now arriving in data centres are undemonstrated. (performance & compatibility; waits on Network taps and certifiers) [2]
- Whole-workload recomputation (reproducible packets): Checking that taps are correctly installed and stay in place at scale is not a solved problem, and hardening the recomputation server inside the prover's facility needs significant research. (hardware trust; waits on Tamper evidence for verifier devices) [2][4]
- Whole-workload recomputation (reproducible packets): There is no plan yet for quickly scaling side-channel defences on a frontier cluster; only early theoretical pieces exist. (coverage & hidden compute; waits on Side-channel suppression for isolated facilities) [2]
- Whole-workload recomputation (reproducible packets): Memory wiping may use existing algorithms, but hardware testing is at an early stage. (coverage & hidden compute; waits on Memory wiping and proofs of secure erasure) [2]
- Whole-workload recomputation (reproducible packets): Robust red-teaming of recomputation schemes has not started, and most algorithm development remains academic. (adversarial validation) [2][4]
- Bandwidth limits and compartmentalization: No cap that a verifier can check has been implemented or red-teamed. (adversarial validation) [9]
- Bandwidth limits and compartmentalization: The verifier must know that all traffic leaving a pod crosses the capped, monitored links. (coverage & hidden compute; waits on Network taps and certifiers) [10]
- Bandwidth limits and compartmentalization: Shaping devices and routing assignments must be trusted by both parties; Amodo has not yet fully analysed resilience to a compromised DPU. (hardware trust; waits on Tamper evidence for verifier devices) [9][12]
- Bandwidth limits and compartmentalization: Advances in low-communication training could shrink the margin that the cap enforces. (capacity bounds) [9][13][14]


## What the verifier sees

- Model weights: shown by none; depends on the design for none; hidden by none; not involved in Bandwidth limits and compartmentalization; unspecified for Whole-workload recomputation (reproducible packets).
- Inputs and outputs: shown by none; depends on the design for none; hidden by none; not involved in Bandwidth limits and compartmentalization; unspecified for Whole-workload recomputation (reproducible packets).
- Training data: shown by none; depends on the design for none; hidden by none; not involved in Bandwidth limits and compartmentalization; unspecified for Whole-workload recomputation (reproducible packets).

## Implementations

- Whole-workload recomputation (reproducible packets): [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture)
- Bandwidth limits and compartmentalization: [AI 2040 inference-only verification stack](https://trustbutveri.fyi/implementations/ai-2040-inference-only-verification-plan/) (R1, proposed architecture); [RAND secure inference data center (SIDC) design](https://trustbutveri.fyi/implementations/rand-secure-inference-data-centers/) (R1, proposed architecture)

## Sources

1. Verification Plan, R. Dean (2026). https://ai-2040.com/supplements/verification-plan
2. Get Involved in Verification, AI Futures Project (2026). https://ai-2040.com/supplements/verification-plan/get-involved
3. Verifying international AI deals: Plan A, the state-of-play, and what you can do to help, T. Milton et al. (2026). https://amodo.substack.com/p/verifying-international-ai-deals
4. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). https://amododesign.com/ai-verification/plan-a-sitrep/
5. Fitting a Network TAP to our Inference Verification Prototype, Amodo Design (2026). https://amododesign.com/notes/2026-09-15-network-tap-inference-verification/
6. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). https://arxiv.org/abs/2511.02620
7. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). https://arxiv.org/abs/2608.23375
8. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). https://amododesign.com/notes/2026-06-23-verification-algorithms/
9. Traffic Shaping for Workload Classification, Lucid Computing (2026). https://lucidcomputing.substack.com/p/traffic-shaping-for-workload-classification
10. 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
11. De-risking Interconnect Limits for AI Verification, A. Scher et al. (2026). https://techgov.intelligence.org/blog/de-risking-interconnect-limits-for-ai-verification
12. The Tray as a Bandwidth Boundary, Amodo Design (2026). https://amododesign.com/notes/2026-03-16-dpu-bandwidth-limiter/
13. DiLoCo: Distributed Low-Communication Training of Language Models, A. Douillard et al. (2024). https://arxiv.org/abs/2311.08105
14. Does Distributed Training Undermine Compute Governance?, R. Rahman (2026). https://arxiv.org/abs/2605.29359