Claude's watermark does little to reconstruct work shared across several tools. I want to know how a proposal became an accepted change. An estimate of model participation answers a much smaller question, and I will not use a detector to fill in the rest of the history.
On August 14, 2026, I asked for a Codex agent to work alongside Claude. That same day, Anthropic explained the statistical watermark planned for text from future models. My request concerned collaboration between tools; it was not a completed experiment. The coincidence does make clear what information I would want to follow in that workflow.
According to the company, the watermark adds no hidden characters and identifies no person, organization, or conversation. It helps estimate the probability of Claude's participation in a text. Fine. When another tool reviews a proposal, though, I need to understand what it received and what it checked. Finding a pattern in the prose does not recover those decisions.
Imagine text started by a person, revised by an agent, then adjusted by somebody who knows the subject. Detected model participation can coexist with human contributions at several stages. Editing may also make participation harder to detect without erasing the fact that it happened. Treating the signal as an authorship certificate demands an answer the mechanism does not promise to provide.
With code, the gap is even easier to see. Knowing an explanation passed through Claude does not tell me why a line changed. It also says nothing about whether the reviewer checked the requested behavior or just improved the comment. I want the request connected to the diff and the evidence supporting acceptance. That lets me challenge a particular decision instead of guessing how it happened.
The record should serve the work. Turning every ordinary message into an investigation of who used which tool would be tedious and offer little obvious benefit. For a change I will take responsibility for, the origin of a decision can matter considerably. The amount recorded should follow that need, without inventing identification the watermark does not contain.
For collaboration between Claude and Codex, I choose a readable history of assignments and reviews. A statistical watermark can add its signal where relevant. Whoever accepts the result still needs to check the sources and expected behavior. A wrong claim remains wrong even when the detector accurately recognizes which model probably contributed it.