# AI Futures Project: Verified Compute Allocation Disclosures as a Foundational Governance Tool [Canonical reading](https://escapemeridian.com/readings/ai-futures-project-verified-compute-allocation-disclosures-as-a-a8ddc019) Eli Lifland argues that frontier AI companies should begin pacing: intervening to delay when AIs at a given capability level are developed. He proposes verified compute allocation disclosures (VCAD), records of how a company’s compute is used, checked by third-party auditors. Lifland’s central argument is that pacing rules need credible ways to verify both what compute is doing and whether any compute is hidden from auditors. That matters beyond rules delaying training runs: he argues that limiting software progress, such as algorithmic and data progress, can make pacing more effective by avoiding a buildup that could enable sudden progress later or diffuse to other actors. He also sees public disclosures as a way to make companies’ restraint more credible and give policymakers and the public a concrete benchmark for asking companies to reduce AI R&D compute. The proposal has limits: Lifland says shutting off or destroying GPUs could be easier to verify, though he considers such approaches less politically feasible. He assigns the arguments medium confidence and says important considerations may be missing. The post argues for developing and testing VCAD; it does not report results from such a process. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source AI Futures Project · Article Original source date: Nov 10, 2026 (2026-11-10T02:27:19+00:00). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-10T02:31:49.770Z. First observed: 2026-11-10T02:30:38.066Z. Retrieved: 2026-11-10T02:30:47.506Z. - [Verified Compute Allocation Disclosures as a Foundational Governance Tool](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Limits Lifland assigns the arguments medium confidence and says important considerations may be missing. The post argues for VCAD and describes how it could support pacing; it does not provide results from a deployed or tested VCAD process. Lifland notes that delaying training runs may be especially valuable near the point at which misaligned AIs could take over, qualifying the argument that slowing software progress is much more effective. ## Open question How could auditors establish that compute uses outside their visibility are not being missed? ## Source-stated Lifland proposes verified compute allocation disclosures, with compute-use records checked by third-party auditors, as a first step toward pacing. > Our proposed first step toward pacing is verified compute allocation disclosures (VCAD): companies should regularly report how their compute is used - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that VCAD is needed for verifying many pacing proposals because prohibited activity could take place on compute that is not effectively monitored. > The basic reason is that an attack on essentially any pacing regime is to do prohibited things with compute that isn’t effectively monitored. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that delaying training runs without slowing software progress could allow unused software progress to accumulate, enabling sudden progress later or diffusion to other actors. > This could lead to sudden bursts of progress when training runs are eventually approved, and furthermore the unused software progress could be stolen by China or more generally diffuse to other actors. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that public disclosures could give companies a trusted way to demonstrate restraint and provide a benchmark for public and policymaker requests to reduce AI R&D compute. > VCAD’s public disclosures (the “D”) offer a simple, commensurable benchmark for cross-company comparisons - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland says shutting off or destroying GPUs could make verification relatively easy, while judging such proposals less politically feasible than less extreme ones. > if the GPUs are all shut off, then it becomes relatively easy to verify via e.g. power monitoring. This is doubly true if the GPUs are destroyed. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source and contributor context Eli Lifland · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [image · AI Futures Project](https://substackcdn.com/image/fetch/$s_!RAfk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd9707784-7a6e-4526-811c-e1560613f2da_4200x2967.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Frontier model One of the most capable AI models available at a given time; the set shifts as models improve. - [Definition reference](https://www.gov.uk/government/publications/frontier-ai-capabilities-and-risks-discussion-paper) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## AI alignment Research and practice aimed at making AI systems pursue the goals their builders and users intend. - [Definition reference](https://en.wikipedia.org/wiki/AI_alignment) ## AI Futures Project AI Futures Project publishes AI scenarios, forecasting models and updates with distinct individual forecasts. - [Explore this context](/entities/ai-futures-project) - [Reference](https://blog.aifutures.org/) - [Reference](https://blog.aifutures.org/p/q1-2026-timelines-update) ## Editorial provenance Version 1. Publication ID: f130bd96-4da8-421b-80c7-28105e511876. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T02:31:49.770Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: a3b8204d4b6a8f20df1ea164934bacd88b0218569f0f97aed41f98176414dbbe. Source text hash: 708fa7f6fb506160e354e3c80eb860113d790124c73845dc562d373d219f56e3. ## Revision and source-check history ### Version 1 — 2026-11-10T02:31:49.770Z Eli Lifland argues that frontier AI companies should begin pacing: intervening to delay when AIs at a given capability level are developed. He proposes verified compute allocation disclosures (VCAD), records of how a company’s compute is used, checked by third-party auditors. Lifland’s central argument is that pacing rules need credible ways to verify both what compute is doing and whether any compute is hidden from auditors. That matters beyond rules delaying training runs: he argues that limiting software progress, such as algorithmic and data progress, can make pacing more effective by avoiding a buildup that could enable sudden progress later or diffuse to other actors. He also sees public disclosures as a way to make companies’ restraint more credible and give policymakers and the public a concrete benchmark for asking companies to reduce AI R&D compute. The proposal has limits: Lifland says shutting off or destroying GPUs could be easier to verify, though he considers such approaches less politically feasible. He assigns the arguments medium confidence and says important considerations may be missing. The post argues for developing and testing VCAD; it does not report results from such a process. ## Original source AI Futures Project · Article Original source date: Nov 10, 2026 (2026-11-10T02:27:19+00:00). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-10T02:31:49.770Z. First observed: 2026-11-10T02:30:38.066Z. Retrieved: 2026-11-10T02:30:47.506Z. - [Verified Compute Allocation Disclosures as a Foundational Governance Tool](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Limits Lifland assigns the arguments medium confidence and says important considerations may be missing. The post argues for VCAD and describes how it could support pacing; it does not provide results from a deployed or tested VCAD process. Lifland notes that delaying training runs may be especially valuable near the point at which misaligned AIs could take over, qualifying the argument that slowing software progress is much more effective. ## Open question How could auditors establish that compute uses outside their visibility are not being missed? ## Source-stated Lifland proposes verified compute allocation disclosures, with compute-use records checked by third-party auditors, as a first step toward pacing. > Our proposed first step toward pacing is verified compute allocation disclosures (VCAD): companies should regularly report how their compute is used - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that VCAD is needed for verifying many pacing proposals because prohibited activity could take place on compute that is not effectively monitored. > The basic reason is that an attack on essentially any pacing regime is to do prohibited things with compute that isn’t effectively monitored. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that delaying training runs without slowing software progress could allow unused software progress to accumulate, enabling sudden progress later or diffusion to other actors. > This could lead to sudden bursts of progress when training runs are eventually approved, and furthermore the unused software progress could be stolen by China or more generally diffuse to other actors. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland argues that public disclosures could give companies a trusted way to demonstrate restraint and provide a benchmark for public and policymaker requests to reduce AI R&D compute. > VCAD’s public disclosures (the “D”) offer a simple, commensurable benchmark for cross-company comparisons - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source-stated Lifland says shutting off or destroying GPUs could make verification relatively easy, while judging such proposals less politically feasible than less extreme ones. > if the GPUs are all shut off, then it becomes relatively easy to verify via e.g. power monitoring. This is doubly true if the GPUs are destroyed. - [Supporting source](https://blog.aifutures.org/p/verified-compute-allocation-disclosures) ## Source and contributor context Eli Lifland · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [image · AI Futures Project](https://substackcdn.com/image/fetch/$s_!RAfk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd9707784-7a6e-4526-811c-e1560613f2da_4200x2967.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Frontier model One of the most capable AI models available at a given time; the set shifts as models improve. - [Definition reference](https://www.gov.uk/government/publications/frontier-ai-capabilities-and-risks-discussion-paper) ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## AI alignment Research and practice aimed at making AI systems pursue the goals their builders and users intend. - [Definition reference](https://en.wikipedia.org/wiki/AI_alignment) ## AI Futures Project AI Futures Project publishes AI scenarios, forecasting models and updates with distinct individual forecasts. - [Explore this context](/entities/ai-futures-project) - [Reference](https://blog.aifutures.org/) - [Reference](https://blog.aifutures.org/p/q1-2026-timelines-update) ## Editorial provenance Version 1. Publication ID: f130bd96-4da8-421b-80c7-28105e511876. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T02:31:49.770Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: a3b8204d4b6a8f20df1ea164934bacd88b0218569f0f97aed41f98176414dbbe. Source text hash: 708fa7f6fb506160e354e3c80eb860113d790124c73845dc562d373d219f56e3. ### uncertain — 2026-11-10T02:31:49.770Z — anthropic/claude-opus-5.5 Publication: f130bd96-4da8-421b-80c7-28105e511876 attribution — pass: The byline names Eli Lifland, and the publisher is AI Futures Project. The draft attributes the arguments to Lifland and keeps them as his assertions. The post speaks as 'we', so 'Lifland and AI Futures Project' would be slightly more precise, but this is not a misattribution. > Verified Compute Allocation Disclosures as a Foundational Governance Tool A promising first step toward pacing the frontier Eli Lifland summary — uncertain: Mostly faithful, but it drops some of the source's force and stakes. The post says slowing software progress makes pacing 'much more effective'; the draft says 'more effective'. It omits the closing argument that VCAD is urgent, which ties pacing to avoiding a misaligned AI doing recursive self-improvement. It also omits that Lifland calls GPU shutoff 'promising', so labeling that point as a limit tilts it. 'Compute' is not explained for lay readers. Smallest repairs: restore 'much', add a sentence on the urgency argument, and note that he calls the extreme options promising. > Pacing might be incredibly urgent in order to avoid having a misaligned AI do recursive self-improvement. date — pass: originalPublished is 2026-11-10T02:27:19Z from the publisher page. This matches the post's Nov 10, 2026 byline and the feed entry. > Eli Lifland Nov 10, 2026 support — pass: Each claim is a direct quotation or a close paraphrase of the source. The limitations, including medium confidence, no reported results, and the footnote caveat, are grounded in the text. The question follows from the source's second verification component, checking for hidden compute. It introduces no unsupported benefits or costs. > Verifying there aren’t hidden compute uses outside of what they have visibility into. media — pass: The selected media is the verified direct link to the original article. It is the most direct way to reach the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/a8ddc019e16d99f69eef76283627c80ba3524652ee14467b190c2ea6b5813f9f/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/ai-futures-project-verified-compute-allocation-disclosures-as-a-a8ddc019.json)