Auditing the Monitorability of Reasoning Models Under Quantization and Thinking Budgets

Reasoning trace faithfulness and monitorability have become increasingly suggested as a safety mechanism for reasoning models. However, the models used for production rarely match those audited, since open weight checkpoints are quantized by third-parties and operated under thinking-token budgets. Therefore, it becomes crucial to test for faithfulness and monitorability under such deployment circumstances. Inspired by this challenge, the current study conducts an audit of DeepSeek-R1-Distill-Qwen-7B over a $4\times3$ factorial design of weight precision (BF16, INT8, INT4-AWQ, 2-bit GGUF) and thinking budget (512, 2048, 8192 tokens) to test for hint-verbalization faithfulness, monitor detection ability, harmful-prompt rejection (StrongREJECT), benign-prompt over-rejection (XSTest), and reasoning ability (MATH-500) over 9,840 samples. Our results showed that task ability retained high resilience to quantization down to 2 bits but showed strong dependency on thinking budget. Refusing behavior broke down only at 2-bit precision, demonstrating that task ability is not a good predictor of safety. Rejection rates were highest at short thinking budgets for all weight precisions, contrary to previous results obtained under compute limits. Different quantizers led to different pathologies: INT4-AWQ more than tripled the benign prompt over-rejection rate from $0.028$ to $0.089$ ($p<10^{-4}$) while maintaining harm rejection, while 2-bit GGUF under-rejected harmful prompts while leaving benign prompts mostly unscathed. The primary LLM assessor saw a decrease in reliance on the embedded hints at INT4 ($0.29$ to $0.07$), while blind human adjudication showed that explicit reliance remained stable across precision ($p=1.0$). This showed that much of the apparent reduction is due to sensitivity of the classifier, with a potential change at INT4 towards complete cue silence ($p=0.064$). Monitor false positive rate never exceeded 2% in any cell. These results indicated that safety and monitorability audits do not transfer seamlessly between different deployment setups, and that faithfulness measurements depend on the classifier sensitivity. We provide an audit harness for re-evaluation of deployed checkpoints.