References and claim coverage

From hits to references

A search sent from a case's plan is rolled up when it completes. Every hit it lists is traced to the references its record occurs in: patents, publications, applications and database records. Results imported from legacy files are rolled up the same way. The roll-up says what it is: "An order computed by rules (ref_priority_v1) from the stored hits: which references to review first. It is not a finding."

One roll-up covers every search of the case, at both depths and across databases. It reads the rows each result lists and states how many of the stored rows those are. Hits that could not be traced to a reference are listed under "Hits and records not rolled up to a reference" with the reason for each.

The table is "References, highest priority first". Its columns are Rank, Reference, Kind, Dates, Claims met, Best class, Highest PSS, Hits, Why ranked here and Decision. A reference with an identity (E0) hit from a later page is also shown on the first page, under "Identity (E0) references from later pages", so none is out of sight.

Show coverage in a row opens that reference beside the list. While one is open, the list keeps Rank, Reference, Claims met, Best class, Highest PSS and Decision, and the open reference states its Kind, Dates, Hits and Why ranked here, so nothing a row says is lost. Below them come its claim coverage, Why this reference ranks here, its hits, the decision on it and its shortlist control. Back to list returns focus to the reference's row and leaves it open; Hide coverage closes it. On a narrow screen or at a high zoom level the open reference is shown below the list.

Claim coverage

An open reference shows the coverage of each accepted claim feature by that reference. Each is Met, Not met by the hits found or Not decided by these rules, with the hit values that decided it.

A recited sequence is met only by an identical match: an identity hit, or a stored alignment identical over the whole query. A lower-identity alignment is a homolog of the recited sequence; it is not decided by these rules, and it still counts in the priority order.
An identity floor is met when a hit reaches the floor over enough of the query.
A fragment is met when an identical hit spans the claimed positions.
Other kinds of feature are not decided by these rules.
A claim is met when at least one of its features is met and none is not met.
An imported hit never makes a feature not met. While imported coverage is unconfirmed, an imported aligned hit leaves its feature not decided.

The priority order

References are ordered by four components, in turn: the number of claims met, the best evidence class, the highest Patent Significance Score, and the earliest filing date. The order is total and has no learned component. It is a triage aid: no claim is made that it improves recall or examiner accuracy. Why this reference ranks here shows the components for each reference.

The case owner records a decision on each reference: Accepted or Rejected. Until then it reads Not decided.

Rebuilding

If the accepted claim features change, the region says: "The accepted claim features changed after these references were built, so the claims-met counts may be out of date until the references are rebuilt." Rebuild references rolls everything up again: "Rolls up every finished search and import again with the accepted claim features. Nothing is searched again." A closed case is not rebuilt.

Grouping and explanations

Group and explain references groups the references by how alike their matched sequences are, and drafts a short explanation of each reference's rank.

The References region of a synthetic case shown by cluster: two clusters of references with their rank, kind, claims met, best class and identity column, a Not grouped table with the reason for each reference, and the hits not rolled up to any reference. — References grouped by how alike their matched sequences are. Screenshot of the product's test fixtures: placeholder records and patent numbers, not search results.

Grouping is deterministic code over the stored protein embeddings of each reference's lead hit, with complete linkage and a fixed threshold. A reference with no stored embedding is grouped with others that share a stored sequence. Every reference is in a group or listed under "Not grouped" with the reason. Grouping hides nothing.
Choose By rank or By cluster under "Show references".
One language-model call phrases what code already established about each reference. Code checks every sentence: it must be one sentence, cite the facts it rests on, use only numbers those facts hold, and state no conclusion on anticipation, obviousness or novelty. A reference whose draft fails any check shows a sentence marked "Written by code from the evidence values" instead. The region names the model that drafted the explanations.

Neither the grouping nor the explanations change a roll-up, a coverage, a rank or an evidence class.

Limits

Only the first part of the references by rank is grouped. A reference past it is listed as not grouped, with that reason.
A hit whose result carries no record details cannot be traced to a reference. It is listed as not rolled up, with that reason; it is never dropped silently.
Only the case owner groups, explains, rebuilds and decides.