The Monitorability Tax: How Compression Mechanism Shapes Chain-of-Thought Oversight

Chain-of-thought (CoT) monitoring, in which a separate model reads a reasoning trace to catch reward hacking and misbehavior, has become a central oversight strategy for deployed reasoning models, even as the efficient-reasoning literature compresses CoT from thousands of tokens to hundreds. Whether this compression costs oversight, and whether the compression mechanism matters beyond the compression rate, remains unmeasured. We build a testbed that injects a side-channel hint into MATH-500 problems and scores a prompted LLM monitor by its rank-based detection AUC over the think trace, with the final answer hidden, then compress the CoT of two reasoner families through four mechanisms. The ordering persists both between protocols with similar aggregate behavioral hint-answer agreement rates and within every matched nominal token budget: the monitor's detection AUC differs sharply across mechanisms (0.89 vs. 0.62). Self-summarization stays highly monitorable, while concise prompting and hard truncation fall toward chance, and this ordering reproduces across reasoners, hint types, and monitor scales. Because realized trace lengths are not matched across mechanisms, we report this persistent ordering rather than an isolated mechanism effect. Masking every boxed value and planted-hint string in the monitor-visible text preserves the ordering, so it is not driven by verbatim answer visibility. Motivated by this, we build MonoSumm, a proof-of-concept one-pass distilled compressor that is more monitorable than same-scale full-CoT baselines at every scale while using 15–19× fewer tokens, though at somewhat lower task accuracy and without a matched-competence trained control. As a proof of concept, it provides early evidence that compression's monitorability cost is a design choice rather than an intrinsic price of compressing a chain of thought.