TLAPS-proof LLM serving engine: 516,973 obligations at just 0.18% latency overhead
tianyin_xu · x · 2026-11-13
UCSD's Hao AI Lab introduces vstack, a 5-stage formal verification workflow for coding agents, and veriserve, a prototype verifiable LLM serving engine with PD disaggregation.
- Why: as one engineer orchestrates thousands of parallel coding agents, human review and agent-to-agent review hit a wall; the bottleneck shifts from capability to verification.
- Findings: applying vstack to production Prefill-Decode serving uncovered three classes of real RDMA and cache concurrency bugs in SGLang and vLLM, leading to 6 upstream PRs.
- Scale: veriserve's PdLedger KV-page lifecycle model carries a 226-module, 516,973-obligation TLAPS safety proof — reportedly the largest TLAPS use case to date — adding only 0.18% mean p50 TPOT overhead on GB200.
The post argues verifiable serving defines the "second half" of LLM infra as agents write production systems' deepest layers.
More from coding & agent
- Armin Ronacher: I started 141 projects in a year, but only maintain 10-15 — TheTuringPost · 2026-11-13
- One prompt spawns a 1M-nanobot GPU swarm that morphs into any shape at 60fps — bennash · 2026-11-13
- Dev shares personal agent stack: Hermes + open-source Hermex iOS client — _ScottCondron · 2026-11-13
- Anthropic launches self-hosted Claude Managed Agents in private beta for on-prem deployment — ClaudeDevs · 2026-11-13
- DeepSeek V4 Pro agent sets up a full Hive lab in VMware in about half an hour — huangyun_122 · 2026-11-13
- Benchmark Router launches to run public LLM benchmarks on any model at 1/10th the cost — ycombinator · 2026-11-13