P086 Mutually trusted verification cluster
Can a jointly trusted neutral cluster be built, costed and red-teamed that replays training/inference with a measured leak bound, trusted by both parties?
Building, trusting and operating a jointly trusted (neutral) cluster that replays or verifies training and inference.
Status
Analyzed A design sketch of a neutral jointly-trusted replay cluster (no external communication, no persistent storage, one cluster per chip type), with further designs in Harack et al. (2025) section 2.5.3 and the Proofworks mutually-trusted-compute document. Nothing built: no costed design, no two-party red team, no measured leak bound. Harack et al.: the barriers are security issues with no publicly-known robust solutions. The workshop note that one ~100-chip neutral datacenter project exists is unattributed and does not count.
"Instead, the Prover and Verifier can agree on a neutral jointly-trusted cluster, which they will use to execute the verification protocol. The Verifier needs to be able to trust in the integrity of the cluster’s computation. The Prover, conversely, needs to trust that the cluster will not reveal its private data, by verifying that it cannot communicate with the outside world and has no persistent storage."
Posed by
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections:
building sufficiently trustworthy neutral clusters
(§5.2.2, p.29, L1552–1553) (source states the problem but not the quantity the question asks for)
Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Harack (2026), AI Verification: Infrastructure for Prosperity, Governance, and Peace (Lawfare); Hausenloy & Li (2026), The Compute Verification Post; AI Security Forum, Requests for Discussion; Dean, Verification Plan (AI 2040); Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P086.1 | Has a mutually trusted, jointly operated or overseen verification cluster been built, or given a costed design, and red-teamed by both parties? | open |
| P086.2 | Has any civilian AI data centre or verification cluster been assessed as meeting RAND SL4/SL5 (robust to a competent cyber state)? | open |
| P086.3 | Can one jointly trusted cluster replay training segments produced on several distinct accelerator types within a stated tolerance? | open |
| P086.4 | Is there a cluster design in which both parties can verify absence of external communication and persistent storage, red-teamed with a measured leak bound (bits/day)? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P086.5 | What minimal set of hardware and software in a verification cluster must both parties trust, and how is trust in those components bootstrapped? |
| P086.6 | Which architecture should a trusted cluster use (FPGAs, copies of prover pods), and how is untrusted prover code debugged inside it? |
| P086.7 | How could enclosure-based verification be scaled to meet political needs? |
| P086.8 | Can a mutually trusted cluster verify training (not just inference), with verifier-controlled checkpointing? |
| P086.9 | Does fault-tolerant AllReduce sustain sealed-training throughput to a stated bound under a stated number of node failures? |
Possibly relevant work found by automated search
Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.
For P086.4: Is there a cluster design in which both parties can verify absence of external communication and persistent storage, red-teamed with a measured leak bound (bits/day)?
For P086.5: What minimal set of hardware and software in a verification cluster must both parties trust, and how is trust in those components bootstrapped?
For P086.9: Does fault-tolerant AllReduce sustain sealed-training throughput to a stated bound under a stated number of node failures?
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question