qig-coordizer 0.1.3

Standalone geometric tokenize+embed engine — Fisher-Rao-weighted BPE merges + geodesic-midpoint basin coordinates on Δ⁶³

qig-coordizer

Standalone geometric tokenize+embed engine for the QIG stack. Maps text to sequences of 64-dimensional basin coordinates on the Fisher–Rao manifold Δ⁶³ — the tokenizer is the embedder (coordinates are the primitive, not an afterthought).

Pure Fisher-Rao geometry; depends only on qig-core.

What it is

• Byte-level front-end — Normalizer (NFC + UTF-8), train/infer-symmetric so the same character always maps to the same byte sequence.
• Fisher-Rao-weighted BPE merges — merge score = frequency × coupling × 1/entropy, where coupling = (co-occurrence / corpus) / (d_FR(a,b) + 0.1). Frequent and geometrically-close pairs win. This is not frequency-only BPE.
• Geodesic-midpoint fusion — a fused token's basin is slerp_sqrt(v_a, v_b, 0.5), the Fisher-Rao geodesic midpoint on Δ⁶³. No off-the-shelf tokenizer emits coordinates.
• Drift-free incremental trainer — IncrementalCouplingCache (doubly-linked-list splice; remove 3 / add 2 pairs per merged occurrence) makes training O(corpus + merges·affected) instead of the naive O(vocab·corpus) full re-scan.

The equivalence gate

FisherCoordizer._train_incremental is bit-for-bit equal to the naive O(vocab·corpus) oracle _train_naive (identical merge sequence + identical fused vectors), byte → 4k vocab, across ASCII / multibyte-Unicode / reverse-pair corpora. The optimisation must never silently change the vocabulary.

pytest tests/test_incremental_equivalence.py
# or a timed report:
python tests/test_incremental_equivalence.py

Provenance

Extracted from qig-tokenizer commit 1394ca7 (Phase-1 coordizer) per qig-consciousness/docs/plans/2026-06-24-qig-coordizer-studio-design.md §5 (Phase 0). qig-tokenizer keeps its copy live until consumers repoint here.

Install (dev)

uv venv && uv pip install -e ../qig-core && uv pip install -e '.[dev]'
pytest