Proposal Explorer

Add claims to see which ones the mechanisms address.

In production

• R3 Sampled inference recomputation for checking untrusted workers' activations against the declared model, prompt and precision

Built for an adversarial prover

Bandwidth limits and compartmentalization and Sampled inference recomputation

No new hardware needed

Sampled inference recomputation

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

• Sampled inference recomputation: Analysis

Family findings

• Bandwidth limits and compartmentalization

  Context for AI 2040 inference-only verification stack. These findings concern the family or other implementations; applicability must be checked against their stated scope.

  • Low-communication training reduces the bandwidth training needs in Bandwidth limits and compartmentalization

    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. ^{8 9 10}

    Theoretical argument · Significant · Open. On the record

  • Operator control of pod routing collapses the bound in Bandwidth limits and compartmentalization

    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. ¹⁰

    Theoretical argument · Significant · Open. On the record

  • Undeclared local storage raises per-pod capacity in Bandwidth limits and compartmentalization

    More memory or storage per pod helps an adversary. Lucid requires per-pod storage to be declared, capped and physically inspected. ¹⁰

    Theoretical argument · Significant · Open. On the record

  • Training within one pod is not covered in Bandwidth limits and compartmentalization

    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. ¹⁰

    Open question · Significant · Open. On the record

  • Parallel scale-up switches are hard enforcement points in Bandwidth limits and compartmentalization

    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. ¹¹

    Theoretical argument · Significant · Open. On the record

Open significant flaws

6 flaws in 2 mechanisms

• The recomputation server must be trusted in AI 2040 inference-only verification stack

  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. ^{1 2 4}

  Theoretical argument · Significant · Open. On the record

• Spare compute is not verified in AI 2040 inference-only verification stack

  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. ¹

  Theoretical argument · Significant · Open. On the record

• A recomputation family degrades against prompt-controlling adversaries in AI 2040 inference-only verification stack

  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. ^{1 2 6 7}

  Demonstrated attack · Significant · Open. On the record

• Tolerance for numerical noise leaves a covert channel in Sampled inference recomputation

  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. ^{6 7 20}

  Demonstrated attack · Significant · Open. On the record

  Related mechanism R3 Deterministic and bit-exact inference: Bit-exact inference would remove the numerical tolerance if exact replay can be deployed with the required weights and configuration. A pointer, not evidence that this flaw is mitigated. Add

• Only recorded traffic is checked in Sampled inference recomputation

  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. ^{6 21}

  Theoretical argument · Significant · Open. On the record

  Related mechanism R1 Network taps and certifiers: 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. A pointer, not evidence that this flaw is mitigated. Add

• Some inference optimizations are not covered in Sampled inference recomputation

  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. ^{12 13}

  Theoretical argument · Significant · Open. On the record

Open minor flaws

1 mechanism with minor findings

• Mixed hardware widens the honest baseline in Sampled inference recomputation

  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. ¹²

  Open question · Minor · Open. On the record

Not yet demonstrated

R1 Bandwidth limits and compartmentalization

Mechanisms on the map that are not in the proposal. Pointers, not recommendations. ?

• R3 Deterministic and bit-exact inference Add
  • Bears on the open significant flaw “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.
• R1 Network taps and certifiers Add
  • Bears on the open significant flaw “Only recorded traffic is checked” in Sampled inference recomputation. 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.
  • Bandwidth limits and compartmentalization waits on it. Passive optical taps work at 400G, but the 800G and 1600G line rates now arriving in data centres are undemonstrated.
  • 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.
• R2 Tamper evidence for verifier devices Add
  • Bandwidth limits and compartmentalization 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.
• R1 Memory wiping and proofs of secure erasure Add
  • Bandwidth limits and compartmentalization waits on it. Memory wiping may use existing algorithms, but hardware testing is at an early stage.
• R1 Side-channel suppression for isolated facilities Add
  • Bandwidth limits and compartmentalization waits on it. There is no plan yet for quickly scaling side-channel defences on a frontier cluster; only early theoretical pieces exist.
• R1 Whole-workload recomputation (reproducible packets) Add
  • Bandwidth limits and compartmentalization waits on it. A fully reproducible inference stack needs substantial software and tooling, and per-packet network reproducibility may need considerable software, firmware and possibly hardware work.

Missing prerequisites

• R1 Whole-workload recomputation (reproducible packets) needed by Bandwidth limits and compartmentalization Add

Shared foundations

• Network taps and certifiers relied on by Bandwidth limits and compartmentalization and Sampled inference recomputation

Blockers

11 blockers recorded

• Bandwidth limits and compartmentalization
  • 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
  • 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
  • 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}
  • 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
  • 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
  • Robust red-teaming of recomputation schemes has not started, and most algorithm development remains academic. Adversarial validation ^{2 4}
• 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 ^{1 4}
  • 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 ^{4 6}
  • 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 ^{4 7}
  • 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 ^{12 17}
  • 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 ^{12 21 22}

From the family or selected implementation's record. ?

Model weights

Depends on the design for Sampled inference recomputation.

Unspecified for Bandwidth limits and compartmentalization. Check the implementation record.

Inputs and outputs

Depends on the design for Sampled inference recomputation.

Unspecified for Bandwidth limits and compartmentalization. Check the implementation record.

Training data

Not involved: Sampled inference recomputation.

Unspecified for Bandwidth limits and compartmentalization. Check the implementation record.

Bandwidth limits and compartmentalization

• R1 AI 2040 inference-only verification stack Proposed architecture, AI Futures Project
• R1 RAND secure inference data center (SIDC) design Proposed architecture, RAND

Sampled inference recomputation

• R1 AI 2040 inference-only verification stack Proposed architecture, AI Futures Project
• R2 DiFR (Divergence From Reference) Research prototype
• R1 Low-trust AI compute verification system overview Proposed architecture, Machine Intelligence Research Institute
• R1 SASH confidential network logger Research prototype, Singapore AI Safety Hub (SASH)
• R3 TOPLOC Open-source project, Prime Intellect

1. Verification Plan, R. Dean (2026). Original
2. Get Involved in Verification, AI Futures Project (2026). Original
3. Verifying international AI deals: Plan A, the state-of-play, and what you can do to help, T. Milton et al. (2026). Original
4. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). Original
5. Fitting a Network TAP to our Inference Verification Prototype, Amodo Design (2026). Original
6. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). Original
7. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). Original
8. DiLoCo: Distributed Low-Communication Training of Language Models, A. Douillard et al. (2024). Original
9. Does Distributed Training Undermine Compute Governance?, R. Rahman (2026). Original
10. Traffic Shaping for Workload Classification, Lucid Computing (2026). Original
11. The Tray as a Bandwidth Boundary, Amodo Design (2026). Original
12. DiFR: Inference Verification Despite Nondeterminism, A. Karvonen et al. (2025). Original
13. TOPLOC: A Locality Sensitive Hashing Scheme for Trustless Verifiable Inference, J. M. Ong et al. (2025). Original
14. PrimeIntellect-ai/toploc (GitHub repository), Prime Intellect (2025). Original
15. INTELLECT-2: A Reasoning Model Trained Through Globally Decentralized Reinforcement Learning, Prime Intellect Team et al. (2025). Original
16. adamkarvonen/difr (GitHub repository), A. Karvonen (2025). Original
17. Scaling Recomputation Inference Verification, Amodo Design (2026). Original
18. SYNTHETIC-2 Release: Four Million Collaboratively Generated Reasoning Traces, Prime Intellect (2025). Original
19. An Inference Verification Prototype — Stage 1, Amodo Design (2026). Original
20. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). Original
21. Example Schemes for Verifying High-Stakes AI Agreements, Amodo Design (2026). Original
22. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). Original