From ac3b139a8c7d514991fe77d6c7ac3d5c39b4ce78 Mon Sep 17 00:00:00 2001 From: Shay Date: Sat, 6 Jun 2026 08:03:34 -0700 Subject: [PATCH] docs(language-packs): document relational predicate proposal --- .../README.md | 48 +++++++++++++++++++ 1 file changed, 48 insertions(+) create mode 100644 language_packs/data/en_core_relational_predicates_v1/README.md diff --git a/language_packs/data/en_core_relational_predicates_v1/README.md b/language_packs/data/en_core_relational_predicates_v1/README.md new file mode 100644 index 00000000..68db9ad0 --- /dev/null +++ b/language_packs/data/en_core_relational_predicates_v1/README.md @@ -0,0 +1,48 @@ +# en_core_relational_predicates_v1 + +Proposal-only curated English relational-predicate pack for finite-entity grounding. + +This pack is intentionally compact. It is not a bulk corpus and is not mounted by +default. It contributes binary predicate surfaces only, leaving existing noun, +adverb, and adposition packs untouched. + +## Scope + +- Kinship predicates: `parent_of`, `child_of`, `sibling_of`, `spouse_of` +- Ordering/comparison predicates: `less_than`, `greater_than`, `equal_to`, `distinct_from` +- Spatial predicates: `left_of`, `right_of`, `inside_of`, `adjacent_to` +- Temporal predicates: `before_event`, `after_event`, `during_event`, `overlaps_event` + +## Curation rules + +- Deterministic lexical order: kinship, ordering/comparison, spatial, temporal. +- One row per lemma. +- Every lemma spans at least two semantic domains through `semantic_domains`. +- Surfaces are predicate-shaped to avoid collisions with prior mounted lexical + surfaces such as `parent`, `child`, `left`, `before`, and `after`. +- `gate_engaged` is `false`; this remains pending review and runtime admission. + +## Collision audit + +Audited candidate lemmas against the explicitly requested packs: + +- `en_core_relations_v1` +- `en_core_cognition_v1` + +Additional adjacent mounted/content packs checked while drafting: + +- `en_core_relations_v2` +- `en_core_spatial_v1` +- `en_core_temporal_v1` + +No proposed lemma duplicates a reviewed lemma observed in those packs. The pack +therefore avoids the reverted-pack failure mode where natural-language lemmas +collided with existing mounted entries. + +## Checksum + +`lexicon.jsonl` SHA-256: + +```text +5f3615c69e583b9e1891afd73b90786e5ec563400b24706dbbccb1b4dccef63a +```