docs(curriculum): document en core syntax pack
This commit is contained in:
Shay
parent
9866ee8d05
commit
74570757df
1 changed files with 158 additions and 0 deletions
158
docs/curriculum/en_core_syntax_v1.md
Normal file
158
docs/curriculum/en_core_syntax_v1.md
Normal file
|
|
@ -0,0 +1,158 @@
|
|||
# en_core_syntax_v1
|
||||
|
||||
Status: implemented candidate
|
||||
Scope: foundational language substrate pack
|
||||
Pack id: `en_core_syntax_v1`
|
||||
|
||||
---
|
||||
|
||||
## Purpose
|
||||
|
||||
`en_core_syntax_v1` seeds the first foundation-curriculum substrate needed for later claim parsing and relation binding.
|
||||
|
||||
The pack does not parse sentences. It establishes the lower-level English vocabulary for speaking about syntax, claim structure, evidence binding, reference, polarity, comparison, coordination, and conditional roles.
|
||||
|
||||
This is the first step toward the roadmap rule:
|
||||
|
||||
```text
|
||||
language -> relations -> quantity -> units -> logic -> evidence -> data -> algorithms -> systems -> domains
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Non-goals
|
||||
|
||||
This slice does not implement:
|
||||
|
||||
- a natural-language parser
|
||||
- dependency parsing
|
||||
- clause extraction
|
||||
- evidence-span extraction
|
||||
- claim graph construction
|
||||
- GSM8K statement classification changes
|
||||
- relation-binding execution
|
||||
- new runtime generation behavior
|
||||
|
||||
Those are follow-on slices. This pack only ratifies the vocabulary substrate and makes it loadable/resolvable through existing pack infrastructure.
|
||||
|
||||
---
|
||||
|
||||
## Pack contents
|
||||
|
||||
The pack contains 24 reviewed noun entries:
|
||||
|
||||
```text
|
||||
subject
|
||||
predicate
|
||||
agent
|
||||
patient
|
||||
object
|
||||
modifier
|
||||
clause
|
||||
sentence
|
||||
phrase
|
||||
antecedent
|
||||
consequent
|
||||
referent
|
||||
anaphor
|
||||
qualifier
|
||||
scope
|
||||
polarity
|
||||
coordination
|
||||
conjunction
|
||||
disjunction
|
||||
negation
|
||||
exception
|
||||
comparison
|
||||
attachment
|
||||
evidence_span
|
||||
```
|
||||
|
||||
Primary namespace families:
|
||||
|
||||
```text
|
||||
syntax.*
|
||||
claim.*
|
||||
provenance.*
|
||||
```
|
||||
|
||||
Each lexicon row uses provenance:
|
||||
|
||||
```text
|
||||
foundation_syntax_v1:reviewed:2026-05-30
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Files
|
||||
|
||||
```text
|
||||
language_packs/data/en_core_syntax_v1/manifest.json
|
||||
language_packs/data/en_core_syntax_v1/lexicon.jsonl
|
||||
language_packs/data/en_core_syntax_v1/glosses.jsonl
|
||||
tests/test_en_core_syntax_v1_pack.py
|
||||
```
|
||||
|
||||
The pack ships both a checksum-sealed `lexicon.jsonl` and a checksum-sealed `glosses.jsonl` companion overlay.
|
||||
|
||||
---
|
||||
|
||||
## Admission gates pinned by tests
|
||||
|
||||
`tests/test_en_core_syntax_v1_pack.py` verifies:
|
||||
|
||||
- `load_pack("en_core_syntax_v1")` succeeds;
|
||||
- manifest checksum matches `lexicon.jsonl` bytes;
|
||||
- manifest `glosses_checksum` matches `glosses.jsonl` bytes;
|
||||
- total entry count is 24;
|
||||
- POS distribution is exactly 24 `NOUN` rows;
|
||||
- expected lemmas are present in the compiled manifold;
|
||||
- entry ids are contiguous and zero-padded;
|
||||
- every lexicon row has reviewed foundation provenance;
|
||||
- every gloss lemma is resident in the same pack lexicon;
|
||||
- resolver registration routes syntax lemmas to this pack;
|
||||
- prior pack lemma routing remains unchanged.
|
||||
|
||||
---
|
||||
|
||||
## Resolver order
|
||||
|
||||
The pack is registered in `chat.pack_resolver.DEFAULT_RESOLVABLE_PACK_IDS` after the polarity pack and before relation/kinship packs:
|
||||
|
||||
```text
|
||||
en_core_polarity_v1
|
||||
en_core_syntax_v1
|
||||
en_core_relations_v1
|
||||
en_core_relations_v2
|
||||
```
|
||||
|
||||
This makes syntax vocabulary available to the same deterministic first-match resolver used by existing English content packs while preserving prior high-frequency lemma ownership.
|
||||
|
||||
---
|
||||
|
||||
## Follow-on work unlocked
|
||||
|
||||
The next slices should use this vocabulary to implement actual relation binding:
|
||||
|
||||
```text
|
||||
evals/language_claim_parsing
|
||||
evals/language_relation_binding
|
||||
tests/test_language_claim_parsing.py
|
||||
tests/test_relation_binding_replay.py
|
||||
```
|
||||
|
||||
Expected follow-on primitives:
|
||||
|
||||
- evidence-span extraction
|
||||
- subject/predicate/argument slot binding
|
||||
- modifier attachment
|
||||
- conditional antecedent/consequent binding
|
||||
- polarity and negation scope
|
||||
- reference/anaphora target binding
|
||||
- typed refusal when a required slot lacks evidence
|
||||
|
||||
---
|
||||
|
||||
## Boundary
|
||||
|
||||
This pack is a substrate, not a capability claim. It should not be marked as full language relation binding admission until parser/eval/runtime surfaces exist and pass their own tests.
|
||||
Loading…
Reference in a new issue