The mark on your text becomes checkable. The 85% fidelity figure is not in the record.

The mechanism is narrower than the word watermark suggests. When Claude weighs near-equivalent next words, that choice is normally settled by a random number; Anthropic says watermarking keeps the choice random but draws the randomness from a key and the few preceding words, so a keyholder can later score whether a passage matches the choices that key implies ^[3]. That yields a probability and, by design, nothing else: Anthropic says the mark carries no identifying information and cannot be traced to a specific person, organisation or chat ^[5]. Files get the stronger treatment, signed C2PA metadata that indicates Claude processing and reveals whether the file has been tampered with since ^[7].

So the origin of text becomes checkable in a thin, one-bit sense. The figure on the card is a different matter. A human retest fidelity number, 85% or any other, appears nowhere in the material supplied here ^[16]. What is on the record is the input side of the synthetic-respondent business: CVS Health's agents were built on 2.9 million consented responses from more than 400,000 participants across more than 200 behavioural scenarios ^[12], CVS Health has used those simulations to guide decision-making with Simile's support ^[19], and Simile says it trained a confidence model that predicts the accuracy of every simulation ^[13] while raising over $200 million at a $2 billion post-money valuation ^[14]. TechCrunch's verdict on the company's stated mission to simulate all eight billion people is 'preposterous' ^[15]. Sample counts and a valuation are not a fidelity measure. What arrives in August covers where text passed through; how well a simulated person stands in for a real one is still a number only the vendor holds.