Datasets:

---

v13s

/

golden-fto-layer-a

like 0

Follow

Vox 1

v1.2 lands two structural additions over v1.0.x — looking for
schema feedback before we lock these into the LTS contract:

1. categories[] on each prior_art row. A multi-category
   EP citation like XY (examiner saying "this doc challenges
   both novelty AND inventive step") used to collapse into the
   single severity enum, silently dropping the Y dimension.
   v1.2 preserves the raw signal: categories=["X", "Y"] for XY,
   ["X", "Y", "I"] for XYI. severity stays as the dominant-
   component summary for backwards compat.

2. prior_art_index/<jurisdiction>/index.parquet sibling
   partition. When 2 EP search reports cite the same prior-art
   doc (e.g. EP3290023 + EP4621049 both citing US10721059), the
   ref_id appears as 2 rows in the cases partition. The new index
   pre-computes per-(ref_id, citing_jurisdiction) aggregates with
   citation_count + severity_distribution + first/last_cited_date
   so consumers can ask "how often was doc X cited as prior art
   across EP examinations" without walking the cases partition.

Open questions where outside perspective would help:

• Order in categories[]: alphabetical (current) or
  examiner-cited-order? We have no signal on dominance order
  beyond convention.
• Index partition jurisdiction key: (ref_id, citing_juris)
  preserves "how often did EP examiners cite X" but requires a
  consumer-side join to get global counts. Should we add a global
  partition too?
• JP slice landing (κ-5 INPIT bulk): JP native codes 31/32/33
  → translate to ST.14 X/Y/A in the extractor for cross-juris
  comparability, OR keep raw codes in categories with a
  separate translation table?

Schema bumps are SemVer-MAJOR so they're deliberate, not casual —
v1.2 design doc:
https://github.com/masterleopold/parallax/blob/main/docs/06-evaluation/schema-v1-2-design.md