Cases and claim features
Create a case
Open Cases and choose New case. The form has two parts, "The case" and "The application documents":
- a case title and, optionally, the application number;
- the claims, as a .txt or .docx file, up to 1 MiB;
- optionally, the specification, as a .txt, .docx or .pdf file, up to 8 MiB;
- optionally, the sequence listing, as a WIPO ST.26 .xml file, up to 8 MiB.
All the files together may be up to 9 MiB. A sequence searcher can enter a case for an examiner, who then owns it. Choose Create case to send the files.
Each file's type is checked from its contents, not only its name, and a file whose contents do not match its type is refused. A sequence listing that is not a valid WIPO ST.26 listing is refused, and the case is not created.
What the case holds
A case page has eight regions, in the order of the work: Case, Features, Plan, Searches, References, Suggestions, Summary and Shortlist. They are tabs under the case's title row, which names the case, its application number and its status; on a large enough window the title row and the tabs stay in view as you scroll. The Plan, Searches, References and Suggestions tabs show a count only when the whole set is known: never a zero for a set that is still loading, could not be read, or has not been built or checked. The Case region shows "About this case", the documents and the sequence listing.
- Text is extracted from each document once and kept unchanged, with its SHA-256 fingerprint. Later steps cite character positions in that one text. Show text opens the stored text of a document.
- A PDF with no text layer is listed with the reason "the PDF has no text layer". Image-only PDFs are not read, and ST.25 listings are not read.
- In a Word file, only the body is read: headers, footers, footnotes and comments are not. In the body, an equation is kept as one line of text, for example A/B≥2.5 or (x^2)y. Superscript and subscript digits and signs are kept as superscript and subscript characters, for example 10⁶, 10⁻⁸ M, ¹²⁵I and H₂O; other superscript or subscript text is written as ^(…) or _(…), for example 1^(st). This holds whether the superscript is set on the text itself, through a Word character style, or by raising or lowering the text by at least 1 pt, as a PDF converted to Word often has it. Text raised or lowered by less than 1 pt, and a paragraph style that raises a whole paragraph, are read as ordinary text.
- A picture or chart, a symbol-font character, an embedded object, or an equation that cannot be written on one line, is marked in the text of a Word file as "[unreadable: image]", "[unreadable: symbol]", "[unreadable: object]" or "[unreadable: equation]", so none of these is left out without a mark.
- In a PDF, superscripts and subscripts are read as ordinary characters, so 10⁶ can read as 106 and 10⁻⁸ as 10-8; pictures in a PDF are not marked. Check the original file when a number matters.
- The sequence listing is parsed into a table of SEQ ID NOs with their molecule type, length and organism.
- No narrative synopsis of the specification is written.
A case moves forward through its statuses: Ingested, Features proposed, Features reviewed, Planned, Executing, Complete and Closed. A status never moves back.
Claim features
In the Features region, the case owner chooses Extract features. A language model reads the claims, the specification when it is short enough, and the type and length of each SEQ ID NO. It never receives sequence residues. It proposes claim features of these kinds:
| Kind | What it records |
|---|---|
| Recited sequence | A sequence the claim recites by SEQ ID NO. |
| Identity floor | The lowest sequence identity to a SEQ ID NO that the claim covers. |
| Fragment | A part of a listed sequence, by its positions. |
| Variant | Substitutions at stated positions of a listed sequence. |
| Fusion | A construct joined from components. |
| Wildcard position | A position the claim leaves open. |
| Transitional phrase | Comprising, Consisting of or Consisting essentially of. |
| Molecule constraint | A limit on the type of molecule. |
| Other limitation | Any other limitation. |
Code checks every proposal before it is shown. A proposal is refused when, for example, a quote it gives is not in the stored text, it cites a claim other than its own or its parent, it names a SEQ ID NO that is not in the listing, or its wording states a conclusion: "Its wording states a conclusion on anticipation, obviousness, novelty or patentability." A refused proposal is never shown as a feature; the region says how many failed and why.
A dependent claim's features are derived by code from its parent claim's, and are marked "Derived by a versioned rule."
Decide on each feature
Every proposed feature starts as "Proposed, not decided". For each one:
- Show cited text highlights the quoted span in the stored document.
- Why this feature explains the proposal: "Based on", "Rules applied", "Language model contribution", "Basis" and "Would change if". A model's proposal reads "Proposed by a language model and checked by code; not a measurement."
- Choose Accept, Reject or Revise, give a reason when you reject or revise, and choose Record decision.
Decisions are append-only and audited. A revision is stored beside the original, which stays. Under Add a feature you can add a feature the model did not propose; it passes the same checks.
Limits
- Only the case owner extracts and decides features: "Only the case's owner can decide features."
- If extraction fails, the region says why and features can still be added by hand.
- A closed case takes no more changes.
- The model's raw response and its SHA-256 are stored before it is read, so the proposals can be reproduced from the stored response.