WELLyes1/almiraah_coordinate_db

AL-MIRʾĀH Coordinate Database

Accumulated query-coordinate records for the AL-MIRʾĀH 99-Names coordinate system: classical Arabic terms projected into a fitted Poincaré-disk basis built on CAMeLBERT-ca contextual embeddings. Each row is the coordinate profile of one Arabic term measured against the fixed 99-Name basis.

This is the accumulated query set (790 records at the current commit). It is distinct from the fixed 99-entry basis source and the fitted 99-node Poincaré coordinates, which are released separately in the project repository under data/paper_b/.

Schema

Each record (13 fields):

Field	Description
term_ar	The Arabic query term (surface form)
term_undiacritized	The term with short vowels/tatweel stripped (embedding input)
top_name_attractors	Basis Names of highest cosine similarity, with scores
bottom_names	Most-repelled basis Names (most-negative cosine), with scores
estimated_position	Poincaré-disk placement: px, py, r, plus level (0–2) and level_label (Dhāt / Ṣifāt / Afʿāl)
dominant_wazn	Prevailing morphological pattern in the term's neighbourhood
query_wazn	The query term's own parsed pattern (miqyās positional method)
query_wazn_status	Parse status; unvocalised forms may yield candidate sets rather than a unique pattern
cluster_wazn	The attractor cluster's prevailing pattern — reported separately and never substituted for query_wazn
abjad_value	Mashriqi Abjad total — provisional (see note below)
poincare_dist_to_primary	Distance to the primary attractor; a confidence signal (larger = lower confidence)
timestamp	Record creation time
pipeline	Engine/pipeline version that produced the record

Using this with an LLM

This dataset is the measured output of the AL-MIRʾĀH coordinate system; the data itself is documentation. An LLM does not use it by reading these records directly. The intended runtime path is the live MCP server, which exposes the coordinate engine as callable tools:

• philological_lookup(term) — full coordinate record for one term
• root_analysis(root) — basis Names sharing a triliteral root, with field geometry
• semantic_neighbors(term, k, …) — nearest accumulated queries by Poincaré distance
• compare_terms(term1, term2) — shared/divergent attractors, distances, hierarchy load
• semantic_project(candidates, context) — project candidate forms against a context The role this plays for an LLM is a constraining prior for classical Arabic terms with clean triliteral structure: rather than generating an interpretation freely (and risking confabulation), the model can query a term's attractor profile, tier, and repelled Names against a fixed basis, and condition its output on that structure. The coordinates do not carry meaning on their own — they constrain where meaning may sit. Confidence should track poincare_dist_to_primary: a large distance to the primary attractor signals a weak placement that should not be over-read.

This dataset (the accumulated query set) is useful for inspecting what terms have already been placed and their neighbourhoods; for live projection of new terms, call the MCP server.

MCP endpoint: https://wellyes1-almiraah-transformer.hf.space/mcp

Embedding protocol (v3)

• Model: CAMeL-Lab/bert-base-arabic-camelbert-ca
• Representation: mean of the target word's subword span, layer 8, averaged across three fixed Arabic carrier sentences (not bare-term last-layer)
• Placement: similarity-weighted Karcher mean (Riemannian barycenter) of the top-5 attractor positions in the fitted disk
• Random seed: 42 The v3 protocol corrects an earlier deployed version that embedded bare terms with last-layer pooling; see the companion paper for the protocol-change robustness analysis.

Important provenance notes

• Abjad values are provisional. abjad_value is stored but not finalized: an audit flagged several stored totals, and the numbers depend on orthographic conventions (article-inclusive vs article-free, madda/hamza/tāʾ-marbūṭa valuation) that are not yet adjudicated. Treat abjad_value as convention-labelled and provisional, not as a settled figure.
• Coordinates are basis-specific. They are properties of the 99-Names basis; a different basis would yield a different coordinate system on the same model.
• Tier signals can disagree. level (radial) and the attractor vote are correlated but not identical; both are reported.
• Wazn coverage. Roughly 17% of accumulated terms fall outside the current template inventory and carry a not_parsed-type status.
• query_wazn vs cluster_wazn. These are deliberately separate: the first is the term's own parse, the second is the neighbourhood's prevailing pattern. Do not conflate them.

Citation

• Mislati, A. (2026). The 99-Names Coordinate System: A Fitted Relational Basis for Classical Arabic in CAMeLBERT-ca. Zenodo. https://doi.org/10.5281/zenodo.20739416
• Mislati, A. (2026). Templatic Morphology as Decodable Geometry, and Abjad Letter-Values as an Attention Probe, in Classical Arabic Transformer Models. Zenodo. https://doi.org/10.5281/zenodo.20735409

Links

• Code and basis source: https://github.com/ahmedmest81-ctrl/ALmiraah-project
• Live MCP server: https://wellyes1-almiraah-transformer.hf.space/mcp

License

Released under CC BY 4.0. Share and adapt with attribution.