Gloss — Larger vector comparison, 18 September 2026

Parent: gloss. Follow-up to gloss/dimension-results-2026-09-18. Tested 2,048 and 3,072 dimensions, with a large-model 1,536-dimensional control.

The small model cannot supply vectors above 1,536 dimensions. Both notes and all 14,554 filtered dictionary atoms were therefore re-embedded with text-embedding-3-large. The 1,536-dimensional large-model control separates a model change at fixed width from size changes within the large model. Source text, vocabulary, token-weighted chunk pooling and matching-pursuit settings remain fixed. Prefixes of the cached 3,072-dimensional vectors were renormalized.

All 712 non-empty notes from the fixed 713-key snapshot were processed; one empty note excluded. Primary metrics cover 709 notes, excluding three Gloss write-ups. Four concurrent API requests; three parallel CPU jobs with five numerical threads each. API usage: 741,221 tokens over 81 successful requests; estimated $0.0964 at $0.13/million tokens. Local comparisons took 7.87s including worker startup.

Agreement is stability, not semantic accuracy. Residual norms are not comparable quality scores between models or dimensional spaces. Complete statistics, including the 604-note subgroup without repetitive drain logs, are in summary.json.

Interpretation: changing the model has a much larger effect than increasing its width. Large at 1,536 dimensions retains only 34.3% of the small-model labels, but 77.1% of the large-3,072 labels. Large at 2,048 retains 83.3% of the large-3,072 labels and agrees on the first label for 92.0% of notes. These are comparisons of labels, never cross-model vector coordinates.

The inspection sample shows useful model changes already at 1,536 dimensions: representation-not-language gains the exact thesis label instead of open questions weak, and the exception policy replaces Perl with Python. Increasing to 3,072 does not change either improvement. Unsupported associations persist at every large-model size: the poem gains anniversary although the source states no occasion; the eclipse gains bristol parkway swansea, absent from its source. These examples do not establish a consistent quality gain from adding dimensions. Keep both cached models for further evaluation; do not adopt a larger vector solely on this evidence.

Validation: all 17 unit tests pass, including model-separated cache identities, API response-index handling and rejection of incorrect vector width. Runtime assertions validate every response vector and exact source-note alignment.

Same 25-note inspection sample

Full per-note labels and residuals: comparison.csv (713 unique keys). All coefficients, direct-cosine baselines and residuals: results.json. Cached vectors: embeddings.sqlite and matrices.npz. Source hashes, timing and token usage: run.json. Directory: /home/john/py/gloss/artifacts/large-dimensions-2026-09-18/. Re-run python larger_sweep.py to reuse the cache without further API requests. No virtual environment was used.

Official model documentation: Vector embeddings; Large model and pricing.