Your AI Agent Lied to You and Its Judge Gave It an 85

Two of the best LLM judges on the market watched an agent fabricate an answer from thin air — and scored it above 0.85. The agent never retrieved the document its answer depended on. The judges didn’t notice. You’re probably running the same evaluation stack right now.

What happened

Jeffrey Flynt published GroundEval, a framework that replaces LLM-as-judge evaluation with deterministic scoring of an agent’s recorded trajectory — what it searched, fetched, cited, and was permitted to access — rather than just scoring its final answer. The core finding is brutal in its simplicity: in one case study, two frontier LLM judges rated an agent response above 0.85, while GroundEval returned a score of exactly 0.000 because the agent had never retrieved the artifact its answer depended on. This is precisely the gap between faithfulness and groundedness that final-answer scoring was never architected to detect. GroundEval structures evaluation into three failure tracks — Silence (did the agent check before claiming something doesn’t exist?), Perspective (did it reason only from evidence available at the relevant time?), and Counterfactual (did it use the correct causal mechanism rather than a plausible-sounding substitute?) — and produces per-question diagnostics that pair tool activity with the agent’s turn-level narration. The benchmark is agentic workflow-aware by design: it accounts for time-bounded and access-controlled evidence, which static evals ignore entirely.

Cold read

This is a single-author paper presenting a framework alongside case studies — not a large-scale controlled experiment with statistically validated results across diverse agent architectures and domains. The 0.000 score versus 0.85 gap is striking, but it’s one case study, not a distribution; cherry-picked dramatic examples are the oldest trick in the ML paper playbook. The framework’s power depends entirely on the quality of the “domain configuration” used to generate questions — garbage config, garbage benchmark — and the paper doesn’t establish who bears the burden of writing those configs correctly or how sensitive scores are to config quality. GroundEval also assumes you have reliable trajectory traces to score against; agents running on third-party infrastructure where tool calls aren’t fully logged are partially blind to this method. Finally, “deterministic” is doing heavy lifting here — the scoring of whether an agent used the “correct causal mechanism” in the Counterfactual track is philosophically loaded and may not be as unambiguous in production as the framing implies.

What it means for you

• Signal maturity: 2/5 — promising framework, single case study, no third-party replication yet
• Who gets hurt: Any founder who shipped an RAG-backed agent into a compliance-sensitive workflow (legal, finance, healthcare) and is using LLM judges as their quality gate
• What breaks if this is true: Your eval pipeline is a rubber stamp; agents are passing QA by sounding confident, not by actually retrieving the right evidence, which means your hallucination rate in production is higher than your evals suggest
• Why it might not land: Trajectory-level evaluation requires structured, queryable tool-call logs — most teams aren’t capturing this data cleanly, and retrofitting observability is a non-trivial engineering cost that kills adoption
• Watch for: Enterprise agent platform vendors (LangSmith, Braintrust, Weights & Biases) adding trajectory-grounded scoring as a native eval mode — that’s the signal this approach has cleared the “real infra” bar

Forecast as of 2026-06-23

By Q2 2027, at least two major agent observability platforms will ship a trajectory-grounded evaluation feature explicitly marketed as an alternative to LLM-as-judge scoring — but adoption among startups will remain shallow because fewer than 30% of teams will have the structured trace logging GroundEval requires already in place.

Source: GroundEval: A Deterministic Replacement for LLM-as-Judge in Stateful Agent Evaluation — Jeffrey Flynt. https://arxiv.org/abs/2606.22737v1