Citations for the agentic era

Citation was designed for humans. Footnotes, bibliographies, “see page 12” — all of it assumes a person did the reading, formed a judgment, and pointed back at their sources afterward. It works because we trust the person. It was never verifiable; it couldn't be. You can't inspect what a human paid attention to.

With language models, you can. When a model generates an answer, the connection between output and source isn't a claim it makes about itself — it's a measurable property of the computation. Every generated token attended to specific spans of the input, and that attention can be read out directly.

TokenPath does exactly that. We resolve every span of generated text back to the source that produced it — token-level, for any model, in one API call after the answer. No prompting the model to cite, no trusting its self-report. Attribution as measurement, not testimony.

As agents write more of the words people act on, this is the citation system the agentic world needs — native to how models actually work, not borrowed from how humans wrote papers. If that's a problem you have too, we'd love to hear from you.