Project name chosen deliberately: to gloss a text is to annotate it with explanatory words alongside the original — exactly what the Amplified Bible does by hand, and exactly what this project tests doing mechanically. Status: prototyping.
A small, deliberately cheap feasibility check for the concept-extraction layer described in Concept Factorisation & Hierarchical Indices: can matching pursuit decompose a note's embedding into a few weighted words that actually describe what the note is about, rather than flattening it to one lossy point? If the labels look sensible on a handful of real notes, that's the case for building the concept index properly; if not, better to find out cheaply.
Grew out of a conversation (2026-09-18) connecting three previously separate threads: notes as legacy across sessions with no persistent memory otherwise; the Amplified Bible fork in Representation, Not Language (now resolved there — plurality-with-weights, not a centroid); and the cross-session continuity design in Structured Context Protocol and Shared Experience Architecture, which this would eventually serve as a retrieval layer for.
• Handover to Codex — specification for the initial nine-note feasibility prototype, completed 2026-09-18.
• Prototype results — 18 September 2026 — 72 candidates, 65 retained atoms, nine notes: four useful glosses, four partial, one poor. Plural merges materially change results; negation is lost in the labels. Worth a further bounded experiment, not a production index yet. Scripts and cached embeddings: /home/john/py/gloss.
• Corpus-wide results — 18 September 2026 — all 713 public-store keys accounted for (712 non-empty); six dictionary sizes up to 16,341 raw or 14,554 filtered candidates. Per-note tables and a 25-note spot review: 13 improved, 5 worse, 7 mixed versus the original 65-atom glosses. More vocabulary helps coverage, but residual subtraction still introduces unsupported specifics. Local interactive browser and CSV: /home/john/py/gloss/artifacts/corpus-2026-09-18/.
• Vector dimension sweep — 18 September 2026 — 256/512/1,024/1,536 dimensions across all 712 non-empty notes; four parallel workers × four threads, no API calls. Keep 1,536 for now: 1,024 retains 87% of first labels but only 50% of complete label sets; qualitative effects are mixed. Full CSV and provenance saved locally.
• Larger vector comparison — 18 September 2026 — 2,048/3,072 dimensions plus a 1,536 large-model control, all 712 non-empty notes. Changing model has a larger effect than increasing width; no consistent quality gain from added dimensions in the inspection sample. Estimated embedding cost $0.0964, cached.
• False-match selection experiment — 18 September 2026 — direct cosine versus pursuit at 1/2/3 labels, all 712 non-empty notes, no paid calls. Eight tuning and seven new checking notes reviewed: no stable winner; two-label pursuit is promising on the checking sample, but errors persist. Includes per-label evidence and missed-topic annotations.
• Source-evidence gate — 18 September 2026 — exact lexical support applied to cached labels across all 712 non-empty notes. It removes clear specifics such as Perl, summer and FreeBSD, but is too strict for paraphrases; use as confidence metadata, not hard deletion. No paid calls.
• Experiment ranking and recommended default — 18 September 2026 — ranked configurations and recommendation: filtered vocabulary + small 1,536-dimensional model + two direct-cosine labels + evidence status. No paid calls for this ranking.
• Implementation plan for the recommended configuration — 18 September 2026 — sidecar factors, canonical concepts, evidence status and concept-to-note index; source notes remain unchanged until review.
• Source-aware reranker — 18 September 2026 — widened the cached candidate pool to 50, promoted source-supported alternatives, and updated all 709 implicit !gloss companions. Unsupported factors fell from 88 to 1; canonical concepts and evidence rules are now draft control notes.
• Concept Factorisation & Hierarchical Indices — the design being tested.
• Representation, Not Language — the theory this grew out of.
• Structured Context Protocol and Shared Experience Architecture — why this matters beyond one session.