Measured, not claimed

What real prosecution shows

Three measurements of the tool against public USPTO records: how well it reads an Office Action, what real continuations changed from their parents, and where the tool looks when it proposes directions. One of them is about a weakness, and it is here for the same reason the other two are.

Read this first

These are small studies: 2 file histories for the first, 20 families for the second, 11 for the third. Every application is public and named below so you can check it. Nothing on this page is a recommendation about any application, and none of it is a legal determination: what a record means for a particular continuation is the practitioner's reading.

1. Reading an Office Action

The tool 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.

Reading the Office Actions and the Notice of Allowance of the file histories above
Model	Run	Rejections found	Allowance reasons found
`Claude Sonnet 5.5`	remote	10 of 10	2 of 2
`kimi-k2.5`	remote	10 of 10	2 of 2
`qwen3.5:9b`	on a laptop	10 of 10	2 of 2
`qwen3.5:9b`	through OpenRouter	9 of 10	did not finish

A 9-billion-parameter model running on a laptop found every rejection, which is the configuration that keeps a file on the machine. The same model reached through OpenRouter did worse and some of its calls ran to the time limit. No model named a reference the paper does not contain: the tool checks every reference against the paper's own text and withholds one it cannot find.

Scored against the record as read by hand; a licensed practitioner's check of that reading is under way, and this table will change if it does. It counts rejections and invented references only, which hold however one disputed reference in the second file is read.

2. What real continuations changed

The tool marks every limitation of a parent's allowed claims as filed with the application or added in a round of prosecution. We asked whether that distinction lines up with what practitioners actually did in the continuations that followed. 20 families, sampled from the USPTO's public data: each a parent allowed after at least two Office Actions over prior art, and a continuation filed from it. For every limitation of the parent's allowed independent claims, we checked whether the continuation's independent claims still carry it.

What the continuations left out of their parent's allowed independent claims
Limitation in the parent's allowed claims	Left out of the continuation's independent claims
Filed with the application	25 of 197 (13%)
Added during prosecution	35 of 111 (32%)
All of them	60 of 308 (19%)

Limitations added during prosecution were left out about 2.5 times as often as filed ones: a difference of 19 points, with a 95% interval of 2 to 35 points when whole families are resampled. It holds when any one family is removed.

This is a description of what practitioners did in 20 families, not a recommendation. It does not say those limitations were unimportant, that leaving them out was wise, or anything about why each continuation took the shape it did; some pursued a different embodiment altogether. What it supports is narrower: the record of when each limitation entered the claims is worth having in front of you, which is why the tool shows it.

"Left out" means fewer than half of the limitation's significant words appear in the continuation's independent claims. A stricter test, the exact phrase or nine tenths of its words, gives 46% against 82%: the same direction, larger, and inflated by rewording. The earlier-versus-later distinction within prosecution did not hold in this sample and is not claimed.

Every family

Every family by application number, so any row can be looked up in Patent Center.

Each family's limitations, and how many the continuation left out
Parent	Continuation	Art unit	Filed, left out	Added, left out
14/893,689	17/681,899	1798	0 of 0	2 of 7
16/085,793	18/070,557	3752	0 of 3	0 of 7
16/485,486	18/157,310	2686	2 of 21	0 of 4
16/525,438	17/677,136	2613	2 of 7	2 of 5
16/589,636	18/239,553	2666	0 of 7	2 of 15
16/615,974	18/174,702	3676	2 of 7	1 of 10
16/645,022	18/465,810	3746	2 of 3	0 of 1
16/820,519	18/065,160	2497	3 of 8	3 of 5
16/853,828	18/075,487	2812	0 of 8	0 of 0
16/858,891	18/157,395	2812	0 of 10	3 of 5
16/885,445	17/709,763	2433	0 of 12	6 of 18
16/941,358	17/735,992	3711	2 of 34	0 of 2
16/951,296	18/465,924	2634	0 of 9	1 of 3
17/055,713	17/985,859	2478	1 of 17	0 of 0
17/095,189	18/065,161	3672	0 of 16	0 of 6
17/130,243	18/297,671	2139	0 of 10	5 of 7
17/167,459	17/709,680	3677	0 of 1	0 of 4
17/232,598	17/985,743	2818	8 of 9	0 of 0
17/536,510	18/093,647	3632	0 of 7	0 of 0
18/132,074	18/132,064	2637	3 of 8	10 of 12

3. Where the tool looks

The tool's first step proposes directions for a continuation, each tied to a passage of the specification. We checked whether those passages sit near where each real continuation's new claim elements came from, and compared the tool with two baselines at the same number of positions: random places in the specification, and the places the parent's own claims came from.

How often the tool's proposed directions sat near what each continuation went on to claim
Model	Run	Directions near what the continuation claimed	Random places	Where the parent's claims came from
`Claude Sonnet 5.5`	remote	16%	38%	68%
`kimi-k2.5`	remote	13%	35%	67%
`qwen3.5:9b`	on this machine	45%	36%	67%

This one is a finding against the tool. 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. "Near" is within about a paragraph, the only threshold measured.

What none of this shows

Whether any continuation was well drafted, what it did in its own prosecution, or what any application should claim. The tool drafts claim language and scores drafting quality; it does not assess patentability, and these measurements do not change that. The drafting benchmark, a separate study, is on the benchmark page, and one complete drafting run is printed on the sample page.