Gloss — Source-aware reranker results, 18 September 2026

Parent: gloss. Replaced the first-generation companion factors with a source-aware reranker over the cached embeddings. Source notes and keys are unchanged.

The reranker searches the top 50 direct-cosine candidates rather than only the original top three. It assigns a small bonus to labels whose content words all occur in the source, a penalty to partially supported labels, and a larger penalty to labels with no lexical evidence. It retains two distinct labels and records both raw and adjusted scores. This is a conservative heuristic, not a language-understanding proof.

Factor statusCountShare of 1,418 factors
text-supported140198.8%
related161.1%
unsupported10.1%

The previous top-two companions contained 824 text-supported, 506 related and 88 unsupported factors. The reranker produces 1,401 text-supported, 16 related and 1 unsupported factors across the same 709 source notes. This is a large reduction in unsupported specificity, but exact lexical support can reject useful paraphrases; the statuses are therefore visible and the canonical vocabulary remains draft.

Source notePrevious problematic selectionReranked selection
PROGRAMMING_RULES/exceptionsexception handling; perl pythonexception handling; error handling
poems/first-to-Kathykathy; summer poemkathy; heart
hosts/gravlaxwebdav notes; gravlax and kelp; freebsd hostwebdav notes; webdav vhost
TFG/exercise-demo-protocolfunctional exercise; review and handover; sample demonstrationsdemonstrating; exercise

The reranked companions are published under the same implicit keys, such as hosts/gravlax!gloss. They carry selection_method: source-rerank-v1. The 709 generated notes exclude the three generated Gloss notes and the one empty source note.

Control notes: canonical concepts, evidence rules, and evaluation set. Next work is curation and blind source-grounded review, not another embedding-size sweep.

Local artifacts: /home/john/py/gloss/artifacts/reranked-companions-2026-09-18/ and /home/john/py/gloss/artifacts/gloss-companions-reranked-2026-09-18/.

version 1  ·  created 2026-09-18  ·  updated 2026-09-18  ·  tags gloss, corpus, embeddings, results