How We Read These Filings

What Was Sampled

What was excluded, and why it matters. The provisional and application samples are software filings in a limited set of technology areas, so nothing in them speaks to chemistry, mechanics or biotech. The continuation sample spans several areas but was drawn from a list of granted patents, which is not a random sample of continuations. Documents whose text could not be retrieved are recorded as unobserved, never as absent and never as zero.

A separate set of 53 filings from named firms was read for comparison. Those comparisons are not published here. The comparator for them is the same 32 calibration documents rather than a technology-matched control, so a difference between the two groups cannot be attributed to the firms, and presenting it as though it could would be an assessment of identifiable practitioners' work. Worked examples on these pages are drawn only from the reference-numbered corpus families.

What Was Compared

Documents were read as filed, from the published record, in the versions the file history contains. For the provisional comparison, each element of the later application was recorded against the provisional as present, partial, additional, or not aligned. For continuations, the child's description was compared with its direct parent, not with a named ancestor further up the chain.

What the Words Mean

Three terms carry most of the weight, and each is narrower than it sounds.

• Additional. A reading compared two documents and recorded this element of the later one as not present in the earlier one. It does not mean the earlier document was deficient, incomplete or insufficient, and it does not mean the content was required. Necessity was never measured.
• Element. A passage the reader treated as one unit of content. How finely a document divides into elements is a judgment, which is why element shares are reported as properties of a reading rather than of a document.
• Status, on an office-action ground. What the reply did: amended, argued, both, cancelled the claims, or contained no responsive text. It records the reply's content and nothing about what the examiner did next.

Content was classified against two fixed vocabularies, frozen before the readings that use them: 84 kinds for specification content and 74 for office-action content. They are separate sets in separate namespaces, so an identifier from one does not mean anything in the other.

How the Readings Were Checked

• Every recorded passage had to be found at the locator given for it, in the source document, and every output had to match its schema. A reading that failed was retried once and then recorded as failed rather than repaired.
• Kind identifiers had to come from the frozen set. A reader could not invent one.
• Passages the reader left unclassified had to carry a reason. Checking those reasons afterwards found metadata, headings, page furniture and scanning artefacts, and none citing a missing kind, which is the evidence that substantive content was not quietly routed out of the counts.
• Four office-action pairs were re-read by a different reader with no access to the first reader's answers. The two agreed on the status of all seven grounds in those pairs. Seven grounds is far too small a base to support a reliability claim, and it was not treated as one.
• Figures in these pages were recomputed from the raw reading outputs rather than taken from a summary.

Isolation was by instruction and staging rather than enforced by the operating system. Readers were given only their own task inputs, but nothing prevented a determined process from reading more, and the record says so rather than describing it as confinement.

What This Cannot Establish

Three limits are worth stating plainly, because each one bounds a figure that appears on another page.

• Element counts depend on the reader. An independent reader segmented the same documents into 5% to 38% more elements than the primary reader. Shares computed over elements therefore describe a reading. Counts of whole documents do not have this problem, which is why they are preferred wherever both were available.
• Figures from different readers are never combined. The provisional comparison, the archetype placement and the continuation alignments were produced by different readers, and in one case by two readers in two waves. They are reported separately and never averaged.
• No holdout test was run. Nothing here was validated by applying it to a reserved set of filings and checking whether it held. These are descriptions of the documents that were read.
• Small groups stay small. Some figures rest on two, three or five items. Where that is true the page gives the count beside the share, and several breakdowns in the underlying material were left out of these pages entirely because their denominators were too small to mean anything.
• None of it is a legal conclusion. Sufficiency, written description, enablement, eligibility, novelty, obviousness, new matter, inventorship and whether any rejection was overcome are all outside what a count of document contents can reach.

Why Publish the Margins

A method page that reports only the checks that passed is worth less than no method page, because a reader cannot tell which figures to lean on. The zero-margin result above weakens one claim on one page. Leaving it out would have made that page read as stronger than the evidence behind it, which is the failure this page exists to prevent.

Educational, not legal advice. These pages describe documents. Decisions about filings belong with a registered patent practitioner.