Link the paper and add analysis quickstart entry points

README.md

@@ -11,6 +11,7 @@ tags:
 - llm-as-a-judge
 - benchmark
 - cyclomatic-complexity
+- arxiv:2609.19616
 task_categories:
 - text-generation
 size_categories:
@@ -103,6 +104,11 @@ This dataset accompanies the NeurIPS 2026 Evaluations & Datasets track paper
 Generation Reliability* (Michael Hernandez and Tian Zhao, University of
 Wisconsin-Milwaukee). Code and analysis: <https://github.com/uwm-se/ComplexityKink>.
 
+**Paper:** [The Complexity Kink (NeurIPS 2026)](https://arxiv.org/abs/2609.19616)  
+**Get started:** [Open the analysis notebook](https://huggingface.co/datasets/uwm-software-engineering/complexity-kink-research/blob/main/examples/Complexity_Kink_Quickstart.ipynb) | [Command-line example](https://huggingface.co/datasets/uwm-software-engineering/complexity-kink-research/blob/main/examples/quickstart.py)
+
+Use the released results to compare prompt-side complexity measures, investigate disagreements between unit tests and LLM auditors, or examine how task and source composition affect reported performance. The starter examples explore these questions using the published data, without new model calls. See the intended uses and limitations below when interpreting pooled results.
+
 It scores the structural complexity of 5,000 Python programming tasks from the
 prompt alone, never from a model's code or test results, and pairs that score with
 105,000 generated solutions (21 LLMs x 5,000 prompts), their unit-test outcomes,