TLAPS-Bench: agentic AI formal verification uncovers read consistency bug in SlateDB's WAL
tianyin_xu · x · 2026-11-07
A new paper introduces TLAPS-Bench, a benchmark and execution harness for constructing machine-checked proofs of real-world system specifications in TLA+ using agentic AI.
Immediate payoff: the author's TLA+ specification of SlateDB's WAL protocol revealed a read consistency bug — one SlateDB likely shares (extremely unlikely to trigger, still needs confirmation). The culprit is once again GC complicating put-if-absent; a simple fix has been validated.
Notably, the author's own model checking with 3 competing writers never caught the violation, while TLAPS-Bench with Astra Max found it at 4 competing writers — a strong showcase for agentic-AI-driven formal verification of distributed systems.
More from coding & agent
- LoRA Loom is a Windows desktop app for one-click local LoRA training — Abject-Ad1903 · 2026-11-07
- Which agent called which tool? Reddit thread on agent identity and verifiable logs — fred_pcp · 2026-11-07
- Computer-Using World Model Lets GUI Agents Simulate Actions Before Executing Them — burkov · 2026-11-07
- opentel-mcp v0.18 Adds SHA-256 Hashing to Detect MCP Tool 'Rug Pull' Attacks — Thirumalaiboobathi · 2026-11-07
- Uncle Bob on constraining coding agents with coverage, complexity and mutation testing — bibryam · 2026-11-07
- User asks his agent "can't or won't?" — and the agent backs down — StewartalsopIII · 2026-11-07