Results and evidence classes
What this result covers
A completed search's page has two panes. The left pane holds the query, What this result covers and "On this result", which links to each table on the page; its first link, Skip to hits, takes the keyboard straight to the hits. The right pane holds the view controls and the tables. On a narrow screen or at a high zoom level the left pane comes first, above the first hit. It is never collapsed or hidden.
What this result covers says which depth ran, against which corpus snapshot and databases, what the identity lane covers for this query, how many records were masked, which filters or exclusions applied, and any known limit that applied to the query. Read it before the hits: it says what an empty or short list does and does not show.
Evidence classes
Every hit of a sequence search carries one evidence class. The class says how the hit was found and how strongly it is supported.
| Class | What it means |
|---|---|
| E0 Exact identity | The record's whole sequence is identical to the whole query. The identity lane finds it before any AI tier, and it is never cut from a result. |
| E0 Reverse-complement identity | The record's whole sequence is identical to the reverse complement of a nucleotide query. |
| E1 Aligned | The record aligns with the query, and the alignment is stored. |
| E2 Profile or domain | The record matches a profile hidden Markov model or a protein domain. |
| E3 Remote homology, found by AI retrieval and verified by alignment | AI similarity retrieval found the record and alignment then confirmed it. |
| E4 Structure | Reserved for structure evidence. No hit is given this class today. |
| E7 Unverified, AI similarity only, not aligned | AI similarity retrieval found the record, but it has not been aligned. It is not evidence. |
E5 and E6 are reserved and are never shown. A pattern match carries no evidence class: it is shown as a pattern match, with its positions. A candidate found by AI retrieval counts as evidence only after alignment confirms it.
Each hit also says how it was found: exact identity, AI similarity retrieval, near-identity pre-screen, profile search, full-corpus alignment, mmseqs2 nucleotide search or combined ranking. An aligned nucleotide hit that is not an identity hit was found by alignment search, not by AI retrieval, so its class is E1 Aligned. A nucleotide candidate that mmseqs2 nucleotide search found but for which no alignment was returned is unverified too, and is not evidence; its label names that search and says the alignment was not returned, instead of AI similarity.
Hits, profile hits and unverified candidates
- Hits lists the aligned and identity hits. Each names its patent or publication number, its SEQ ID NO, the part of the query it matches, and, for a nucleotide hit, the strand. Every hit can be reached by paging.
- Profile hits lists profile hits apart: "Found by profile search. Profile E-values are on a different scale from the pairwise scores under Hits and are not compared with them."
- Unverified candidates: not evidence lists the candidates that were found but not aligned. When AI similarity retrieval alone found them, it says: "AI similarity retrieval found these candidates, but none has been aligned, so none carries alignment evidence. They are not part of this result’s evidence and are not used in the examiner report. Their order is AI retrieval order, not a score." When a deterministic search, such as mmseqs2 nucleotide search, found one of them, the explanation names the searches that found them instead and says that no alignment was returned for any of them, and the first column is "Retrieval order" rather than "AI retrieval order".
Hits with an identical sequence can be grouped, and records from the same patent family can be masked. Masking hides a record; it deletes nothing, and the result says how many records are masked. If grouping could not check every hit, the result says: "Grouping is incomplete: some hits could not be checked for identical sequences, so duplicates may still appear as separate hits."
Lanes, order and limits
Lane view chooses what the hit tables show:
- All lanes; excluded candidates listed apart
- Deterministic evidence only
- AI-discovered only
Order sorts hits by stored order, E-value, alignment score, PQM or Score-over-Length, and Limits set ranges on PQM and Score-over-Length. Choose Apply view to apply them, or "Return to the default: all lanes, stored order, no limits or filters". Hits that lack the sort key are listed last, and the page says how many. Identity hits are never cut from the page by a view.
Show stored results opens the complete result set kept for the search, "Stored results, in rank order".
Deep searches and incomplete results
A deep search states how much of its snapshot it scanned. When it scanned only part of the snapshot, or part of the scan reached the most matches it can keep, the result is stated as incomplete and says that records that match may be missing. When part of the scan could not be shown to have returned every match, the result says so and does not claim that every match was returned. A deep search whose coverage was not recorded says that, and makes no claim about coverage.
A deep search of fewer than all databases says which it searched. A deep search has no out-of-distribution check, no Patent Significance Scores and no profile search.
Running and failed searches
While a search runs, the page reads "Running. This search has not completed, so it has no hits yet." A search past its estimate or its deadline says so, and that it has not failed.
A search that did not complete reads "This search did not complete, so it has no result. This is not a refusal of access and not an empty result." When the failed stage is known, the sentence names it.
Notes and relevance judgments
Each hit has Note on this hit and Relevance judgment: Relevant, Not relevant or Uncertain. "A recorded judgment is permanent: it cannot be deleted. Recording another adds a new judgment; earlier ones stay, with their dates." Notes are amended, never overwritten.
The examiner report
Each search writes an examiner report. Its determination tier is computed by code. Its narrative is written by a language model, labeled as AI-generated advisory text with the model id, and not checked by code. Unverified candidates are not used in the report.
Limits
- There is no free-text search within a result set.
- The false discovery rate is not computed for nucleotide searches, and the result says why. When its computation fails, the result says that too.