Never retype: if the model can't retype it, it can't mistype it
BigDocsBigDocs: part 6 of 10
The benchmark behind this series: see the results
The most natural request in the world is “give me the contract back with your changes”. It is also where the quietest errors come from.
To return a document, a model writes it out again. Every word it copies is a word it could change. Most of the time it doesn’t. Occasionally 30 days becomes 60, “must” becomes “should”, a sub-clause merges into its neighbour, or the sentence before a bulleted list disappears. Nothing flags it. The output reads beautifully, and the error sits in a part of the document nobody thought had been touched.
Point, don’t copy
The principle is simple. The model never retypes text that already exists. It points at it, by clause number, by library position or by a short exact span, and code copies it. Whatever the model writes is new text only, and code checks that every pointer resolves to something real.
The same applies to the Wordhouse. When a position is used, the model names it and code fetches the wording. A tender answer drawn from the library is copied in word for word by a script, not paraphrased into the response.
Edits as operations
Changes are expressed the same way: replace this phrase, delete that one, insert this after that anchor. Code applies those operations to the exact pieces of the source and writes them as real tracked changes in the customer’s own Word file.
Three useful properties follow.
Untouched text is identical by construction. Rejecting every tracked change gives you back the original, word for word. That can be tested with a plain text comparison rather than taken on trust.
Every change is visible. If the model does alter wording it shouldn’t have, it shows up as a tracked change rather than hiding inside a rewritten paragraph.
Conflicts are flagged, not guessed. If two edits overlap, code stops and asks rather than picking one.
A lighter hand
Working this way also changes the size of the edits. In the benchmark the pipeline deleted about 6.5 words per changed paragraph. The single prompt deleted about 10 when working paragraph by paragraph, and about 22 when asked only for its changes. All three of its full runs removed a whole section without marking it “[Intentionally omitted]”, leaving a gap in the numbering, and none repaired the cross-references. BigDocs: the benchmark
Smaller edits matter more than they look. A counterparty reads a light redline as a negotiation and a heavy one as a rewrite.
Beyond contracts
The pattern travels. An HR team updating forty employment contracts for a new award rate doesn’t need forty regenerated contracts; it needs one precise change placed forty times. A support team updating help articles after a release needs the changed sentences, not new articles. Marketing needs the claim corrected, not the brochure rewritten.
What your team can do on Monday
Stop asking for the whole document back. Ask for a list of changes, each with the exact current text and the replacement, and apply them yourself or with a simple script. You will find the edits get smaller, and you will be able to see every one.