Parent: gloss. Ranking of the tested configurations by the practical goal: useful retrieval labels with fewer unsupported specifics. There is no blind human gold set, so this is an evidence-based engineering recommendation, not a measured accuracy leaderboard.
Use the filtered word-plus-phrase vocabulary (14,554 atoms), text-embedding-3-small at 1,536 dimensions, direct cosine ranking, and two labels by default. Attach a source-evidence status to each label: text-supported, related, or unsupported. Use the source passage as the authority whenever a label is used in a factual summary.
| Rank | Configuration | Why it sits here | Main limitation |
|---|---|---|---|
| 1 | Small 1,536 + filtered vocabulary + direct cosine top 2 + evidence status | Best practical balance: free after the existing run, reproducible, more stable than residual pursuit, and fewer opportunities for a third false specificity. | Cosine can repeat near-synonyms and can miss a second facet. |
| 2 | Small 1,536 + filtered vocabulary + two-step pursuit + evidence status | More diverse coverage; on the seven checking notes it had 13/14 supported labels versus cosine’s 12/14. | On the eight tuning notes it had lower supported-label precision (68.8% versus 75.0%). |
| 3 | Large 2,048 + filtered vocabulary + two-step pursuit or cosine | Best larger-model candidate: 83.3% of its labels were retained against large 3,072 and it agreed on the first label for 92.0% of notes. | Requires paid re-embedding; the sample does not prove better semantic accuracy. |
| 4 | Large 3,072 | Full-width large model and a useful reference for the large-model family. | No demonstrated quality gain over large 2,048; more storage and cost. |
| 5 | Large 1,536 control | Separates model effects from width and produced some better topical labels than the small model. | Its labels differ substantially from small 1,536 (34.3% mean label retention), so it is a model migration, not a size tweak. |
| 6 | Small 1,024, 512 or 256 | Cheaper/smaller vectors are available offline by truncation. | Stability falls sharply: at 1,024 only 50.1% of complete pursuit label sets agree with small 1,536; qualitative effects are mixed. |
| 7 | Raw phrase-expanded vocabulary or original 65-atom prototype | Useful controls and historical baselines. | Raw phrases introduce navigation debris; original vocabulary has poor coverage and the nine-note result was already mixed. |
Three labels increase coverage but also increase false specifics. In the reviewed examples, third labels included child reads eof, setdns and solar eclipse july. One label is safer but often omits a meaningful facet. Two is the current compromise, not a calibrated optimum.
The exact lexical gate should not delete labels: it removes clear errors such as perl python, summer poem and freebsd host, but also rejects useful paraphrases such as functional exercise. Keep it as confidence metadata. A factual answer should be allowed to cite only text-supported labels, while retrieval may retain related labels.
The ranking does not establish truthfulness. All tested methods can miss negation, procedural constraints and important secondary topics. The next real quality test is a blind, source-grounded evaluation with a reviewed vocabulary, explicit support/contradiction labels, and separate scoring for coverage, unsupported specificity, repetition and polarity.
Evidence: corpus results; selection comparison; evidence gate; larger-model comparison; dimension sweep. No paid calls were made for this ranking.