The moving lens

Give a model a document and it can read it, up to a point. There is a limit to how much it can hold in mind at once, and even the largest limits fall far short of, say, the entire body of a country's legislation. The moving lens sidesteps the limit instead of fighting it. Rather than trying to hold a huge corpus in the model's attention all at once, it moves a small window across the corpus, changing what needs changing, and keeps the whole thing coherent as it goes.

Think of it as three motions. It pans across the corpus, one region at a time. It zooms in on the exact place a change belongs, down to the sentence. And it holds a wider focus on the rest of the document, so a change made in one place stays consistent with everywhere else it matters. The model only ever looks through a small window; the lens is what moves that window with intent.

The numbers that matter are about reach and effort, not the size of the machine. It has edited a corpus of more than five million tokens, far larger than any model's context window, in minutes, and the effort it spends per edit stays flat no matter how large the corpus grows. It never holds the whole corpus in memory; it carries a small, fixed footprint whether the corpus is ten pages or ten million tokens. We ran it over ten real Acts of the UK Parliament, made its changes, verified every one across the whole corpus, and left everything it did not touch exactly as it was.

The two usual ways to reach across a corpus this large both give something up. Loading it all into a model's context would demand orders of magnitude more memory to hold at once. Chopping it into pieces and passing over them again and again loses the consistency between the pieces, so a change in one chunk drifts from the same change in another. The moving lens keeps the whole corpus coherent at a fraction of the memory either of those would need.

What is solid is the reach, the speed, the flat effort, and the guarantee that everything outside an edit is left exactly as it was, proven now on clean, well-defined substitutions in text that defines its own terms, like law or reference material. Where we look next is the widening of that reach: changes that touch many places at once, insertions and deletions as well as substitutions, and eventually any large body of text, each step measured and shown the way this one was. The full method, the corpora, and the measurements are on the detail page.