Proposal Explorer

Add claims to see which ones the mechanisms address.

In production

• R3 Deterministic and bit-exact inference for reproducing open-model inference from receipts in Gensyn's information-market service

Built for an adversarial prover

Zero-knowledge proofs of inference, Bounding unexplained information in outputs, Deterministic and bit-exact inference and Bandwidth limits and compartmentalization

No new hardware needed

Zero-knowledge proofs of inference and Deterministic and bit-exact inference

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

• Zero-knowledge proofs of inference: Independent red-team
• Bounding unexplained information in outputs: Independent red-team
• Deterministic and bit-exact inference: Analysis
• Bandwidth limits and compartmentalization: Analysis

Open significant flaws

13 flaws in 4 mechanisms

• The proof covers a fixed-point approximation, not the floating-point model in Zero-knowledge proofs of inference

  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 calls floating-point emulation in ZKPs an open problem. ^{1 5 7 11}

  Open question · Significant · Open. On the record

• A proof speaks only for the computations that were proven in Zero-knowledge proofs of inference

  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. ¹²

  Theoretical argument · Significant · Open. On the record

  Related mechanism R1 Proofs of useful work for capacity accounting: The record names proof-of-work accounting as the kind of compute accounting needed to show that proven inference was the only work done. A pointer, not evidence that this flaw is mitigated. Add

• Proofs do not bind computational effort (Hollow-LLM) in Zero-knowledge proofs of inference

  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. ¹⁰

  Demonstrated attack · Significant · Open. On the record

• Prompt-controlled entropy inflation widens the covert channel in Bounding unexplained information in outputs

  Gumbel-based inference verification tolerates token choices that honest GPU nondeterminism could produce, and the size of that tolerated set grows with the model's output entropy. Kezins, an independent researcher, showed that an adversary who controls the prompt distribution can raise output entropy and roughly double the bits leaked per token. Across six models of 1 to 32 billion parameters, this cut the slowdown from 146–254 times under benign prompts to 60–118 times. Kezins argues that architectures built on the same unexplained-information bound inherit this attack surface, and recommends calibrating tolerances against local token entropy rather than benign traffic. ^{14 15}

  Demonstrated attack · Significant · Open. On the record

  Related mechanism R3 Deterministic and bit-exact inference: Bit-exact replay would remove the tolerance for numerical noise that sets the size of this channel. The record notes that it needs full hardware and software metadata. A pointer, not evidence that this flaw is mitigated. In the proposal.

• Information the declared computation explains is not bounded in Bounding unexplained information in outputs

  The bound limits unexplained bits only. Outputs that the declared computation fully explains can still carry valuable information: a compression study notes that an adversary with inference access can extract more proprietary information per bit than naive transmission allows. ^{13 16}

  Theoretical argument · Significant · Open. On the record

• Channels other than checked outputs are outside the bound in Bounding unexplained information in outputs

  The inference-verification scheme treats side channels as out of scope. A low-trust system design argues that suppressing physical covert bandwidth below kilobits per second is much more achievable than aiming for zero, and that a malicious device can leak one bit of information by deliberately outputting a wrong result. ^{11 14}

  Theoretical argument · Significant · Open. On the record

  Related mechanism R1 Side-channel suppression for isolated facilities: Physical side channels need separate suppression, which is this mechanism's purpose. A pointer, not evidence that this flaw is mitigated. Add

• The facility-level design is untested in Bounding unexplained information in outputs

  The compute-verification architecture is described with protocol details, potential attacks and prototyping plans, but no prototype results have been published. ¹³

  Open question · Significant · Open. On the record

• Cross-hardware replay relies on reverse-engineered, closed behaviour in Deterministic and bit-exact inference

  Emulating one GPU's rounding on another requires reverse-engineering tensor-core arithmetic and modelling proprietary kernel choices. Hawkeye covers a subset of NVIDIA architectures and states that attention and other higher-level operations need further reverse engineering. For the bit-exact emulator, a proprietary Hopper kernel family is an open edge case. ^{17 19}

  Open question · Significant · Open. On the record

• 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. ^{28 31 32}

  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. ²⁸

  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. ²⁸

  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. ²⁸

  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. ³⁰

  Theoretical argument · Significant · Open. On the record

Open minor flaws

2 mechanisms with minor findings

• The model architecture is disclosed in Zero-knowledge proofs of inference

  ZKML "requires that the model architecture (but not weights) is revealed", and zkLLM assumes a publicly known model structure. Architecture can be commercially sensitive. ^{1 3}

  Theoretical argument · Minor · Open. On the record

• Some kernels remain genuinely nondeterministic in Deterministic and bit-exact inference

  The bit-exact work separates kernels that are deterministic but not batch-invariant from truly nondeterministic ones that use atomic functions. Some integer de-quantization kernels use atomic additions and remain nondeterministic, so exact replay needs backends that avoid them. ¹⁷

  Open question · Minor · Open. On the record

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

• R1 Proofs of useful work for capacity accounting Add
  • Bears on the open significant flaw “A proof speaks only for the computations that were proven” in Zero-knowledge proofs of inference. The record names proof-of-work accounting as the kind of compute accounting needed to show that proven inference was the only work done.
  • 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.
• R1 Side-channel suppression for isolated facilities Add
  • Bears on the open significant flaw “Channels other than checked outputs are outside the bound” in Bounding unexplained information in outputs. Physical side channels need separate suppression, which is this mechanism's purpose.
  • Bounding unexplained information in outputs waits on it. Physical side channels need separate suppression, and one design treats a low residual bandwidth, rather than zero, as the realistic target.
• R2 Tamper evidence for verifier devices Add
  • 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.
• R1 Network taps and certifiers Add
  • Bandwidth limits and compartmentalization waits on it. The verifier must know that all traffic leaving a pod crosses the capped, monitored links.
  • Bounding unexplained information in outputs depends on it.
• R3 Sampled inference recomputation Add
  • Bounding unexplained information in outputs depends on it.

Missing prerequisites

• R3 Sampled inference recomputation needed by Bounding unexplained information in outputs Add
• R1 Side-channel suppression for isolated facilities needed by Bounding unexplained information in outputs Add
• R1 Network taps and certifiers needed by Bounding unexplained information in outputs Add
• R2 Tamper evidence for verifier devices needed by Bandwidth limits and compartmentalization Add

Shared foundations

• Network taps and certifiers relied on by Bounding unexplained information in outputs and Bandwidth limits and compartmentalization

Blockers

17 blockers recorded

• 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}
  • ZKML and zkLLM prove fixed-point arithmetic, and floating-point emulation in ZKPs is described as an open problem. Performance & compatibility ^{1 3 11}
  • 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}
  • 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
• Bounding unexplained information in outputs
  • The prover's compute must be isolated so that all traffic passes through the verifier's interlock; any unmonitored path voids the bound. Coverage & hidden compute. Waits on Bandwidth limits and compartmentalization ¹³
  • Physical side channels need separate suppression, and one design treats a low residual bandwidth, rather than zero, as the realistic target. Coverage & hidden compute. Waits on Side-channel suppression for isolated facilities ^{11 14}
  • Tolerance for numerical nondeterminism sets the size of the residual channel; bit-exact replay would remove it but needs full hardware and software metadata. Protocol soundness. Waits on Deterministic and bit-exact inference ^{11 15}
  • Recomputation over confidential weights and inputs needs a protected setting: prover recomputation in a verifier-controlled enclosure, verifier recomputation in a prover-controlled enclosure, or zero-knowledge proofs. Privacy & leakage ¹³
  • No prototype of the facility-level architecture exists to red-team. Adversarial validation ¹³
• Deterministic and bit-exact inference
  • Batch-invariant kernels cost throughput: in Thinking Machines' Qwen3-8B test, an improved deterministic build took 42 s against 26 s for vLLM's default, and SGLang reports an average 34.35% slowdown on its FlashInfer and FlashAttention 3 backends. Performance & compatibility ^{18 20}
  • Coverage is incomplete: the bit-exact emulator targets dense blocks on NVIDIA GPUs and excludes mixture-of-experts inference and training, and vLLM's batch-invariant mode is in beta, with open work on AMD hardware and speculative decoding. Performance & compatibility ^{17 21 33}
  • Amodo's status page for the AI 2040 verification plan rates a reproducible inference stack for that plan as 'not started'. Performance & compatibility ³⁴
  • Exact replay requires the prover to disclose weights, software versions, parallelism and batch sizes to whoever recomputes. Privacy & leakage ^{11 17}
• Bandwidth limits and compartmentalization
  • No cap that a verifier can check has been implemented or red-teamed. Adversarial validation ²⁸
  • The verifier must know that all traffic leaving a pod crosses the capped, monitored links. Coverage & hidden compute. Waits on Network taps and certifiers ¹¹
  • 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 ^{28 30}
  • Advances in low-communication training could shrink the margin that the cap enforces. Capacity bounds ^{28 31 32}

From the family or selected implementation's record. ?

Model weights

Depends on the design for Bounding unexplained information in outputs and Deterministic and bit-exact inference.

Hidden by Zero-knowledge proofs of inference.

Not involved: Bandwidth limits and compartmentalization.

Inputs and outputs

Shown by Zero-knowledge proofs of inference.

Depends on the design for Bounding unexplained information in outputs and Deterministic and bit-exact inference.

Not involved: Bandwidth limits and compartmentalization.

Training data

Not involved: Zero-knowledge proofs of inference, Bounding unexplained information in outputs, Deterministic and bit-exact inference and Bandwidth limits and compartmentalization.

Zero-knowledge proofs of inference

• R1 Attestable zero-knowledge inference prover Product, Attestable
• R2 EZKL Product, Zkonduit
• R1 Low-trust AI compute verification system overview Proposed architecture, Machine Intelligence Research Institute
• R2 zkLLM Research prototype, University of Waterloo

Bounding unexplained information in outputs

None on the map

Deterministic and bit-exact inference

• R2 Batch-invariant inference kernels (Thinking Machines) Open-source project, Thinking Machines Lab
• R3 Verde and RepOps (Gensyn) Product, Gensyn
• R1 Low-trust AI compute verification system overview Proposed architecture, Machine Intelligence Research Institute

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

1. zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun et al. (2024). Original
2. zkllm-ccs2024: code for zkLLM: Zero Knowledge Proofs for Large Language Models, H. Sun (2024). Original
3. ZKML: An Optimizing System for ML Inference in Zero-Knowledge Proofs, B.-J. Chen et al. (2024). Original
4. NanoZK: Privacy-Preserving Verifiable Inference for Large Language Models via Layerwise Zero-Knowledge Proofs, Z. Wang (2026). Original
5. Proving LLMs at Scale, Attestable (2026). Original
6. Verifiable evaluations of machine learning models using zkSNARKs, T. South et al. (2024). Original
7. Zkonduit EZKL Security Assessment, F. Casal et al. (2025). Original
8. DeepProve-1: The First zkML System to Prove a Full LLM Inference, Lagrange Labs (2025). Original
9. Lagrange-Labs/deep-prove (GitHub repository), Lagrange Labs (2026). Original
10. Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference, C. Gong et al. (2026). Original
11. A System Overview for Near-Term, Low-Trust AI Compute Verification, N. Cankaya (2026). Original
12. Pacing AI Requires Proof, Attestable (2026). Original
13. Verifying AI Compute by Bounding Unexplained Information Exfiltration, J. Petrie & Y. Mühlhäuser (2026). Original
14. Verifying LLM Inference to Detect Model Weight Exfiltration, R. Rinberg et al. (2025). Original
15. Adversarial Entropy Inflation Against Gumbel-Based Inference Verification, N. Kezins (2026). Original
16. Haiku to Opus in Just 10 bits: LLMs Unlock Large Compression Gains, R. Rinberg et al. (2026). Original
17. Bit-Exact AI Inference Verification Without Performance Tradeoffs, N. Cankaya (2026). Original
18. Defeating Nondeterminism in LLM Inference, H. He & Thinking Machines Lab (2025). Original
19. Hawkeye: Reproducing GPU-Level Non-Determinism, E. Badash et al. (2026). Original
20. Towards Deterministic Inference in SGLang and Reproducible RL Training, The SGLang Team (2025). Original
21. Batch Invariance (vLLM documentation), vLLM project (2026). Original
22. gensyn-ai/ree: Gensyn Reproducible Execution Environment (GitHub repository), Gensyn (2026). Original
23. EigenCloud Brings Verifiable AI to Mass Market with EigenAI and EigenCompute Launches, EigenCloud (2025). Original
24. Building Delphi: Pricing, Settlement, and Agentic Trading, D. Jedamski (2026). Original
25. Reproducible Execution Environment (REE) (Gensyn documentation), Gensyn (2026). Original
26. What is Delphi? (Delphi documentation), Gensyn (2026). Original
27. Verification Plan, R. Dean (2026). Original
28. Traffic Shaping for Workload Classification, Lucid Computing (2026). Original
29. De-risking Interconnect Limits for AI Verification, A. Scher et al. (2026). Original
30. The Tray as a Bandwidth Boundary, Amodo Design (2026). Original
31. DiLoCo: Distributed Low-Communication Training of Language Models, A. Douillard et al. (2024). Original
32. Does Distributed Training Undermine Compute Governance?, R. Rahman (2026). Original
33. [Feature]: Batch Invariant Feature and Performance Optimization (vLLM issue #27433), vLLM project contributors (2025). Original
34. AI 2040 Plan A — Verification SITREP, Amodo Design (2026). Original