The moving lens: method and measurements
This is the full record behind the moving lens: how it works, what it was run on, what was measured, and where the measurements stop.
The method. For each place a change is needed, the lens reads a fixed window of about two thousand tokens around the edit site, plus a smaller retrieval budget of about a thousand tokens that pulls in related passages from elsewhere in the corpus, so a change stays consistent with the rest. The text outside the window is frozen and spliced back byte-for-byte, so nothing drifts where no edit was intended. After each change, a mechanical pass over the whole corpus checks that the edit landed and that nothing else moved. The underlying engine is a diffusion model running on a single desktop GPU.
What it was run on. Three corpora. The platform's own research and planning material, concatenated to five million tokens. Ten real consolidated Acts of the UK Parliament, among them the Companies Act 2006, the Equality Act 2010, and the Financial Services and Markets Act 2000, at 4.47 million tokens, taken under the Open Government Licence. And a synthetic stand-in used before the real-legislation run.
What was measured. The effort per edit is flat: across a change in corpus size of more than four hundred fold, the time and memory per edit barely moved. Memory stayed at roughly seventeen gigabytes throughout. Every edit was verified across the whole corpus, and every region outside an edit was preserved exactly. As a matter of record, a full run of this kind costs about one and a half cents of electricity; the same task through a large frontier model, chunking a corpus this size across several passes, works out on published prices to between fifty and ninety US dollars and the better part of a day. That last figure is arithmetic from list pricing, not a measured run, and is set down here as context rather than as the point.
Where the research goes next. The economics and the preservation are solid, and what is proven now is a clean, verified substitution inside a body of text that defines its own terms, like law or reference material. From here the work broadens along clear lines: to changes that touch many places at once and stay consistent across all of them, to insertions and deletions rather than substitutions alone, and to prose that does not hand you its own vocabulary. We are building the piece that judges, up front, how far a given corpus suits the technique, so its range widens by design rather than by guess. What we are aiming at is a lens that can be pointed at any large body of text and trusted to change it as asked. Each of those steps is a thing we will measure and show, the same way we measured this one.
Status. The method was validated end to end in a research cycle on the twelfth of July.