What real prosecution shows

1. Reading an Office Action

On two real file histories, a 9-billion-parameter model on a laptop found every rejection, and no model named a reference the paper does not contain.

The app reads a Patent Center "download all documents" file, scanned pages included, and reports each rejection an Office Action states and the reasons a Notice of Allowance gives, with the quotation each came from. We scored that reading against the record, read by hand, on the file histories of 18/541,216 and 19/180,870: their Office Actions and their Notice of Allowance. One rejection is one of the examiner's own headings: a ground, the claims it covers and the references it applies, as the Office Action groups them.

The model that found every rejection ran on a laptop, the configuration that keeps a file on the machine. The same model reached through OpenRouter did worse on this run and some of its calls ran to the time limit: a result for that route rather than for the model, since OpenRouter chooses the host and hosts differ in how they serve it.

No model named a reference the paper does not contain, so the app's own check, which withholds any reference it cannot find in the paper, had nothing to withhold. The qwen3.5:9b run through OpenRouter listed 3 citations that are not references of record, a statute and a court decision among them; they are in the paper, so that check passed them, and the app now withholds statutes and case law as a separate step, a change made because of this run.

Scored against the record as read by hand; a patent practitioner is checking that reading, and this table will change if it does. One entry is open: the Final Rejection in the second file states its obviousness rejection over two references, then applies a third to later claims, and the hand reading lists all three. Read as two, the local model's one missed reference disappears; the rejection counts are the same either way. The last column counts the reasons a Notice of Allowance states, where it states them; the reasons for allowance are not always written out.

3. Where the app looks, and what it changed

The app reads the specification for potential subject matter beyond the parent's claims, as candidates for the practitioner, each tied to a passage of the specification; in the app that is the Embodiments step, and this measures the pass it builds on. We checked whether those passages sit near where each real continuation's new claim elements came from, and compared the app with two baselines at the same number of positions: random places in the specification, and the places the parent's own claims came from.

This one is a finding against the app. Its first step looks for matter the parent left unclaimed, and the stronger models do that most faithfully, which put them below the random baseline on most families. The continuations mostly stayed on the parent's own ground: 6 of the 17 we could read added nothing to the parent's claims at all, only leaving limitations out. That is the reason the first step is being revisited, and the reason the record of section 2 matters.

11 families could be scored: the rest had no machine-readable specification on the USPTO's system or added no new element. A new element is a limitation of the continuation's independent claims, of at least three significant words, that no allowed claim of the parent carries in its wording (the exact phrase or nine tenths of its words). The percentages are averaged family by family; the counts in brackets pool every new element across those families, which is why the two can differ. "Near" is within about a paragraph of where the element's words sit in the specification; the last column repeats all three at about 3 paragraphs, where the gaps narrow and the parent's own ground still leads in every row. Each baseline uses the same number of positions as the app located: random places are drawn uniformly across the specification, and the parent's positions are sampled from where its own claims came from, each averaged over 200 draws from a fixed seed.