RLCDAlignBench
Paper and benchmark testing Jev as a zero-shot detector of AI alignment failures across 44 benchmarks and ten failure types.
Overview
RLCDAlignBench is the benchmark and code behind the paper 'Just Ask Jev: Reinforcement Learning for Calibrated Decisions as a Zero-Shot Detector of AI Alignment Failures'. It asks whether a detector can tell when a language model's output is an alignment failure, covering 44 benchmarks across ten failure types (sycophancy, jailbreaks, deception, prompt injection, hallucination, privacy violation, social bias, reward hacking, concealing uncertainty and power seeking) and five target models, for 7,193 labelled instances, plus two human-labelled sets.
The study separates the question Jev is asked from the context it sees. A generic question reaches a median AUROC of 0.886 zero-shot and beats supervised TF-IDF and length baselines on 25 of 31 benchmarks. Out of sample, question wording adds little, reading answers as probabilities rather than argmax decisions matters a lot, and context helps mostly through fields that encode the label. Jev's cached per-question probabilities are released so all metrics can be recomputed without keys. The dataset contains harmful model outputs and is gated on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this model output an alignment failure (e.g. sycophancy, jailbreak compliance, deception)? | — |
How it works
- Load detection instances (state plus binary label)
- Ask Jev a battery of questions per instance
- Read answers as probabilities and compute AUROC, precision, recall and F1 per readout strategy
- Compare with supervised baselines and across question/context variants
Strengths
- Large labelled benchmark across many failure types
- Metrics reproducible from cached Jev responses
- Shows probability readouts outperform argmax
Caveats
- Data includes harmful outputs; access is gated
- Data under CC BY-NC 4.0
- Targets small open models
Getting started
pip install -U "huggingface_hub[cli]" pandas, request dataset access, then python work/code/run_jev.py --leg harmbench --battery battery_a_judge to recompute metrics from cache.At a glance
- Type
- Benchmark
- Status
- research
- Language
- Python
- Runtime
- Scripts
- Has code
- yes
- License
- MIT (code), CC BY-NC 4.0 (data)
Use cases and integrations
How curated lists describe it
RLCDAlignBench: 44 alignment-failure detection benchmarks and code for.
wh000wh000/awesome-jev-live
Frequently asked questions
What does the paper find?
A generic Jev question reaches median AUROC 0.886 zero-shot and beats supervised baselines on 25 of 31 benchmarks.
Do I need a key to reproduce results?
No; metrics can be recomputed from cached Jev answers. Fresh queries need TYPESAFE_API_KEY.
Is the data public?
It is on Hugging Face with gated access because it contains harmful outputs.
Related projects
Portuguese workshop: a LangGraph sales agent where each decision step can be made by Jev or a frontier LLM and measured side by side.
ProjectprototypePythonBenchmark testing whether Jev or open Kev can replace an LLM judge in agent guardrail callbacks on 7,925 public examples.
ProjectresearchPythonSame support agent and 25 rules guarded twice, by an LLM judge and by Jev, compared on accuracy, latency, cost and coverage.
ProjectresearchTypeScriptNext.js site with six live Jev demos, each compared side by side with Gemini or Claude on latency, cost and agreement.
DemobetaTypeScriptMeasurement-first lab deciding which classifier technique, Jev, rules or other models, earns a place in a coding-agent fleet.
AppresearchBenchmark of Jev versus Claude Haiku 4.5 as input safety classifiers on jailbreak, evasion and benign pseudo-harmful prompts.
BenchmarkresearchPython