Because a text watermark only re-derives the random number used at high-entropy "fork" choices rather than adding a separate signal, passages with few forks carry little or no watermark no matter their origin: quoted spans, structured data, proofread input, and exact code are pinned by the task and leave chance nothing to steer. Watermark capacity estimates this ceiling — discounting code, quotes, structured lines, and factual anchors to surface the share of a passage genuinely open to word-choice variation — which is checkable without the detecting lab's secret key, unlike the watermark's presence itself. explainx.ai's watermark capacity tool applies this estimate per passage to explain why a clean detector result on low-capacity text proves little either way.