133 lines
4.7 KiB
Markdown
133 lines
4.7 KiB
Markdown
# ADR-0015 — Language Packs as Compiled Linguistic Manifolds
|
|
|
|
**Status:** Accepted
|
|
**Date:** 2026-05-13
|
|
|
|
---
|
|
|
|
## Context
|
|
|
|
CORE's language philosophy is not localization. English, Hebrew, and Koine Greek
|
|
serve distinct architectural roles in CORE-Logos:
|
|
|
|
| Language | Role |
|
|
|---|---|
|
|
| English | Operational base and articulation surface |
|
|
| Hebrew | Depth-root language: root morphology, semantic compression, creation-word density |
|
|
| Koine Greek | Depth-relation language: Logos precision, case/aspect/voice/mood, clause relation |
|
|
|
|
The existing `sensorium/adapters/text.py` scaffold mounted `en`, `he`, and `grc`,
|
|
but the packs were only token lookup wrappers. That is insufficient for CORE.
|
|
A language pack must not be a dataset or a translation table. It must be a
|
|
compiled linguistic manifold.
|
|
|
|
---
|
|
|
|
## Decision
|
|
|
|
A CORE language pack is a deterministic, checksummed, compiled linguistic
|
|
manifold containing:
|
|
|
|
- a manifest with language role, script, normalization policy, source manifest,
|
|
determinism class, checksum, gate state, and OOV policy;
|
|
- lexical entries and morphology entries;
|
|
- grammar attractors;
|
|
- cross-language resonance edges;
|
|
- holonomy alignment cases proving that aligned clauses produce coherent field
|
|
path resonance.
|
|
|
|
## Terminology Boundary
|
|
|
|
This distinction is mandatory:
|
|
|
|
| Term | Meaning |
|
|
|---|---|
|
|
| Vocabulary point / manifold point | Position in the field; surface token entry |
|
|
| Transition rotor | Operator between points, constructed by algebra |
|
|
| Persona motor | Field-bias operator |
|
|
| Grammar attractor | Structural pressure seeded from recurring linguistic form |
|
|
|
|
Vocabulary entries are not transition rotors. Conflating point and operator is
|
|
an algebraic category error. The vocabulary may store multivectors/null points;
|
|
rotor construction belongs to the algebra layer.
|
|
|
|
## OOV Policy
|
|
|
|
Unknown surfaces must not silently collapse to a shared point.
|
|
|
|
| Pack role | OOV behavior |
|
|
|---|---|
|
|
| English operational/articulation | Tagged fallback may be used during early operation |
|
|
| Hebrew depth-root | Fail closed during and after seeding unless explicit expansion path is active |
|
|
| Koine Greek depth-relation | Fail closed during and after seeding unless explicit expansion path is active |
|
|
| Post-seeding expansion | OOV creates a vocab-expansion proposal; it is not projected silently |
|
|
|
|
Returning the same `e1` point for every unknown Hebrew or Greek form erases the
|
|
distinctions those languages exist to preserve; it is anti-Logos.
|
|
|
|
## Morphology / Semantics / Alignment
|
|
|
|
Morphology is operator composition. Semantic domain is attractor geometry.
|
|
Alignment is resonance. These must not be collapsed into one multiplication.
|
|
|
|
For Hebrew, composition order is load-bearing:
|
|
|
|
```text
|
|
V_surface = (((V_root · M_stem) · M_inflection) · M_affix_chain)
|
|
```
|
|
|
|
For Koine Greek, the pack should compose lemma anchors with case/aspect/voice/
|
|
mood/clause-role operators. The grammar relation is structural, not metadata.
|
|
|
|
Semantic domains seed attractors rather than becoming opaque morphology factors.
|
|
Cross-language alignment is a weighted graph, not a translation table.
|
|
|
|
## Crown Proof: Holonomy Resonance
|
|
|
|
Token-level alignment is necessary but insufficient. The decisive proof of the
|
|
three-language design is dynamic:
|
|
|
|
```text
|
|
holonomy(hebrew canonical clause)
|
|
resonates with
|
|
holonomy(koine greek canonical clause)
|
|
and maps coherently to
|
|
holonomy(english articulation clause)
|
|
```
|
|
|
|
Aligned clauses should produce nearby/coherent holonomies without flattening
|
|
their distinctions. Unrelated clauses should remain geometrically distinct.
|
|
Word-order changes should change holonomy. This is the CORE-Logos proof that
|
|
language packs preserve ordered field paths rather than merely mapping tokens.
|
|
|
|
The first holonomy alignment cases should be small and exact: Logos, beginning,
|
|
light, life, spirit, truth, covenant, grace, kingdom, creation.
|
|
|
|
---
|
|
|
|
## Consequences
|
|
|
|
**Positive:**
|
|
- Prevents future agents from treating language packs as datasets.
|
|
- Gives Hebrew and Koine Greek concrete architectural roles.
|
|
- Prevents silent OOV collapse in depth languages.
|
|
- Establishes holonomy-level resonance as the validation gate.
|
|
|
|
**Negative:**
|
|
- Requires a real Supervised Seeding Epoch before Hebrew/Greek gates engage.
|
|
- Requires deterministic morphology and grammar scaffolds before depth packs are
|
|
operational.
|
|
- Requires carefully pinned canonical texts and checksums for D0 ingestion.
|
|
|
|
---
|
|
|
|
## Implementation Order
|
|
|
|
1. Terminology and schema foundation (`language_packs/schema.py`).
|
|
2. Pack roles and OOV policy in `sensorium`.
|
|
3. Split text adapters into English, Hebrew, Koine Greek specializations.
|
|
4. Add grammar scaffold artifacts.
|
|
5. Add tri-language resonance graph.
|
|
6. Add holonomy resonance proof cases.
|
|
|
|
No LLM extraction may feed the gate. Structural segmentation only.
|