VQV Signal

AI AT WORK WATCH TECHNICAL

Hao AI Lab proposes formal verification for agent-built LLM serving code

According to a post on X by @haoailab, Hao AI Lab introduced vstack, a five-stage formal verification workflow for coding agents, and veriserve, a verifiable disaggregated serving engine. The lab says the work uncovered three classes of real RDMA and cache races in production serving engines. It also says veriserve's KV-page lifecycle design is backed by a TLAPS safety proof, with a reported +0.18% mean p50 time-per-output-token overhead on GB200.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-11T20:45:00+00:00 Detected 2026-11-13T02:19:35+00:00
View original source

According to a post on X by @haoailab, Hao AI Lab introduced vstack, a five-stage formal verification workflow for coding agents, and veriserve, a verifiable disaggregated serving engine. The lab says the work uncovered three classes of real RDMA and cache races in production serving engines. It also says veriserve's KV-page lifecycle design is backed by a TLAPS safety proof, with a reported +0.18% mean p50 time-per-output-token overhead on GB200.

AI-assisted summary based on the listed source.

Hao AI Lab introduced vstack, a five-stage formal verification workflow for coding agents, and veriserve, a verifiable disaggregated serving engine. The lab says the work uncovered three classes of real RDMA and cache races in production serving engines and backs veriserve's KV-page lifecycle design with a TLAPS...

For teams building LLM serving infrastructure, the claim is that formal verification could catch RDMA and cache race bugs before they reach production. For developers using coding agents, vstack points to a structured verification step for agent-written code. The 0.18% overhead figure comes from the lab itself, so independent testing would be worth watching.

Teams using AI at work may want to compare this against current productivity and review workflows.

Signal Strength 81% Technical label WATCH Public Interest 27 Category AI AT WORK Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 28 Novelty Interest Score 48 Consequence Score 38 Curiosity Score 16 Shareability Score 42

VQV surfaced this signal because it is recent, relevant to AI Coding Tools, connected to Scobleizer News Pulse (X).