P073 On-device compute-graph logging and guarantee checking

Can on-device compute-graph logging run on a multi-node job with verified completeness and low overhead, and how many falsified operations go undetected?

On-device compute-graph logging and its falsification.

Status

Analyzed Design only: compute-graph declaration to a guarantee processor, static or dynamic, with a proposed description language and kernel translator, and a back-of-envelope log rate (about 10 KB/s per accelerator, which the authors say 'does not scale well with cluster size' when stored locally). Nothing run on a multi-node job; completeness and overhead unmeasured; falsified-operation detection posed as research questions, not tested. Related built work that is not this mechanism: ShadowScope (Almusaddar et al. 2025) validates single-GPU kernel execution against golden performance-counter traces with a non-adaptive attacker; the headline names multi-node logging, so that single-GPU result is a cross-reference.

Alternatively, a description language specifically for flexHEG compute graphs could be developed along with tools to translate accelerator kernels into this language.
Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2) (2025-04 (April 2025 report; arXiv v1 2025-06-03)), L1395–1396

Posed by

Also raised in: Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Peigné-Lefebvre, Nguyen & Wang (2026), Zero Knowledge Verification for Frontier AI Training Is Possible, arXiv 2606.05433; AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.

Sub-questions

IDSub-questionWhat the sources report
P073.1Has dynamic compute-graph / operation logging run on a multi-node job with verified completeness and overhead below a stated bound (incl. MoE dynamic declaration)?open
P073.2What is the compressed log rate per accelerator, and can graphs be assembled and analysed in a distributed, failure-tolerant way at a stated (frontier) accelerator count?open
P073.3Is there a published trace of kernel launches and HBM write timings for a frontier job on multiple accelerator types?open
P073.4Does a translator map production kernels (and host CPU ops) into a checkable description language with stated coverage, securely?open
P073.5Does a compute-graph checker report FLOP contributing to an output to a stated precision (incl. streamed outputs) and map receipts to high-level claims automatically?partly answered
P073.6In red-team tests, how many falsified operations (spoofed DMA, dynamically modified or compliant-looking graphs) go undetected, and can checks make attacker cost grow per operation?open
P073.7Does adding power and timing side data raise detection of falsified compute-graph claims by a measured amount?partly answered

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 P073.1: Has dynamic compute-graph / operation logging run on a multi-node job with verified completeness and overhead below a stated bound (incl. MoE dynamic declaration)?

For P073.2: What is the compressed log rate per accelerator, and can graphs be assembled and analysed in a distributed, failure-tolerant way at a stated (frontier) accelerator count?

Data