Earlier and later semantic neighbors

Three different numbers

• Query match is the similarity between your search text and a result's first independent claim. It only orders search results.
• Claim similarity is the cosine similarity of two patents' first independent claims, as stored embeddings, shown ×100. It selects and ranks neighbors. It is not a probability.
• Neighbor counts are how many external first-independent-claim neighbors clear the 75% baseline on each side of the source's filing date. They are raw observations. No percentile or peer-cohort rating is computed yet.

The signals

• Semantic novelty: earlier external crowding. This counts neighbors ≥75% filed strictly before the source application by a different applicant. The source applicant's own earlier filings are shown separately and not counted. Fewer earlier neighbors means less crowding in the indexed corpus.
• Semantic adoption: later external reach. This counts neighbors ≥75% filed strictly after the source and within 5 calendar years of it, by a different applicant. Same-applicant later filings are shown separately as internal follow-on.
• "External" compares applicant names as filed. A neighbor with no applicant name is unresolved and never assumed external, so it enters neither count.
• Neighbors filed on the same day, or with no filing date, enter neither count and are listed separately.
• Earlier neighbors are candidate evidence, not automatically qualifying prior art. The split uses the application filing date. The patent page's "Search Prior Art" uses the earliest priority date the specification claims, which is a different cutoff.

Completeness

• Neighbors come from the top-200 approximate (HNSW) neighbors of the first independent claim. When even the last retrieved neighbor clears the threshold, more may exist past the cap, so the counts are shown as lower bounds (for example "≥200").
• The search is approximate. A count is what the index returned, not an exhaustive scan of the corpus.
• When the 5-year window after filing extends past the newest grant in the corpus, the later count is marked immature. It is what has been observed so far, not a complete window.
• Only granted US patents are indexed. Applications that are pending, abandoned, or published but not yet granted are absent, so recent years are under-represented.

Not supported by the current corpus

• Family deduplication. There is no family data, so a continuation and its parent each count once. The source patent itself is excluded.
• Percentile ratings and peer cohorts. These need CPC classes and a validated reference population, which are not indexed. No percentile is ever shown.
• Other claims. Only the first independent claim is embedded; later independent and dependent claims are not compared.
• Corporate ownership. The applicant relation compares names as filed. Subsidiaries and renamed companies count as different applicants, so they count as external.

Exploring versus the baseline

Evidence-page filters (threshold, applicant relation, window, grant-date lens) only change the list you are reading. The counts on the patent page, the result grid, and portfolios always use the baseline described here. Changing a rule changes the method version, so results from different versions are never mixed.