diff --git a/chrome/content/zotero/xpcom/bestMatch.js b/chrome/content/zotero/xpcom/bestMatch.js
index 88a7eadb8f..d895f90dd0 100644
--- a/chrome/content/zotero/xpcom/bestMatch.js
+++ b/chrome/content/zotero/xpcom/bestMatch.js
@@ -43,8 +43,9 @@ Zotero.BestMatch = new function () {
 	// model's reading leads: it is the calibrated signal, and it chose which
 	// passages are worth showing. Saying the query's own words lifts a
 	// passage above an equally similar one that only paraphrases them.
-	const SEMANTIC_WEIGHT = 0.7;
-	const LEXICAL_WEIGHT = 0.3;
+	// Passages of an item weaker than this fraction below its strongest
+	// aren't shown
+	const PASSAGE_MARGIN = 0.5;
 	// About a line: what a passage is quoted down to for a one-line preview
 	const SNIPPET_CHARS = 150;
 	// A sentence longer than this is quoted as a line of its words rather
@@ -726,9 +727,11 @@ Zotero.BestMatch = new function () {
 		/**
 		 * Every passage explaining why an item matched this session's query,
 		 * strongest first: the item's passages the model matched and the ones
-		 * that say the query's words. Only the engines that matched the item
-		 * are asked. The strongest MAX_QUOTED_PASSAGES also carry a `snippet`,
-		 * the one line that best shows the query.
+		 * that say the query's words, each as strong as its better
+		 * explanation, and none far weaker than the strongest. Only the
+		 * engines that matched the item are asked. The strongest
+		 * MAX_QUOTED_PASSAGES also carry a `snippet`, the one line that best
+		 * shows the query.
 		 *
 		 * @param {Number} itemID
 		 * @return {Promise<Object[]>} - Entries with `text`, `ranges` and
@@ -757,27 +760,22 @@ Zotero.BestMatch = new function () {
 			for (let i = 0; i < passages.length; i++) {
 				let passage = passages[i];
 				let share = lexical ? lexical[i] : 0;
-				// A passage the model didn't match has only its words to
-				// recommend it, so one that says nothing of the query isn't a
-				// match at all
-				let weighed = passage.score != null;
-				if (!weighed && !share) {
+				// A passage is as strong as its better explanation: the
+				// model's score on its band, unclamped so the strongest keep
+				// their order, or the share of the query the passage says
+				let fraction = typeof passage.score == 'number'
+					? Zotero.Embeddings.getScoreFraction(passage.score, { clamped: false })
+					: 0;
+				let strength = Math.max(fraction, share);
+				if (strength <= 0) {
 					continue;
 				}
-				// Over the engines that spoke for this item, so a strength is
-				// the same 0-1 fraction whether one weighed the passage or
-				// both did
-				let semanticWeight = weighed ? SEMANTIC_WEIGHT : 0;
-				let lexicalWeight = lexical ? LEXICAL_WEIGHT : 0;
-				let fraction = weighed ? Zotero.Embeddings.getScoreFraction(passage.score) : 0;
-				entries.push({
-					...passage,
-					ranges: ranges[i],
-					strength: (semanticWeight * fraction + lexicalWeight * share)
-						/ (semanticWeight + lexicalWeight)
-				});
+				entries.push({ ...passage, ranges: ranges[i], strength });
 			}
 			entries.sort((a, b) => b.strength - a.strength);
+			// Passages far below the item's best don't explain its match
+			let cutoff = (entries[0]?.strength ?? 0) * (1 - PASSAGE_MARGIN);
+			entries = entries.filter(entry => entry.strength >= cutoff);
 			// Only the strongest few passages are shown as rows in the tree,
 			// so only they get a line chosen to quote
 			await this._pickSnippets(entries.slice(0, MAX_QUOTED_PASSAGES));
diff --git a/test/tests/bestMatchTest.js b/test/tests/bestMatchTest.js
index 233fa581b2..be9703029f 100644
--- a/test/tests/bestMatchTest.js
+++ b/test/tests/bestMatchTest.js
@@ -622,10 +622,25 @@ describe("Zotero.BestMatch", function () {
 			let session = await sessionFor();
 			let excerpts = await session.getMatchingExcerpts(attachment.id);
 			assert.lengthOf(excerpts, 2);
-			// 0.7 * 0.5 + 0.3 * 1 beats 0.7 * 0.9 + 0.3 * 0
+			// Each passage is as strong as its better explanation: saying
+			// the whole query beats resembling it closely
 			assert.include(excerpts[0].text, 'the owl itself');
-			assert.closeTo(excerpts[0].strength, 0.7 * 0.5 + 0.3, 1e-9);
-			assert.closeTo(excerpts[1].strength, 0.7 * 0.9, 1e-9);
+			assert.closeTo(excerpts[0].strength, 1, 1e-9);
+			assert.closeTo(excerpts[1].strength, 0.9, 1e-9);
+		});
+
+		it("should drop passages far below the item's strongest", async function () {
+			stubs.push(sinon.stub(Zotero.Embeddings, 'isEnabled').returns(true));
+			stubs.push(sinon.stub(Zotero.Embeddings, 'getScoreFraction').callsFake(score => score));
+			stubs.push(sinon.stub(Zotero.Embeddings, 'getIndexedChunks').resolves([
+				{ text: 'the strongest passage about the owl', score: 0.9 },
+				{ text: 'a passage nearly as strong', score: 0.5 },
+				{ text: 'a passage far weaker', score: 0.4 }
+			]));
+
+			let session = await sessionFor({ lexical: false });
+			let excerpts = await session.getMatchingExcerpts(attachment.id);
+			assert.deepEqual(excerpts.map(excerpt => excerpt.strength), [0.9, 0.5]);
 		});
 
 		it("should skip the lexical engine for an item that didn't match it", async function () {
@@ -685,9 +700,9 @@ describe("Zotero.BestMatch", function () {
 			assert.isTrue(chunksStub.alwaysCalledWith(attachment.id, { queryText: 'owl' }));
 			assert.isFalse(cutStub.called);
 			let byText = Object.fromEntries(excerpts.map(excerpt => [excerpt.text, excerpt]));
-			// Matched by the model alone: weighed by the model alone
+			// Matched by the model alone: as strong as the model's score
 			let liked = byText['a passage the model likes'];
-			assert.closeTo(liked.strength, 0.7 * 0.6, 1e-9);
+			assert.closeTo(liked.strength, 0.6, 1e-9);
 			assert.equal(liked.pageLabel, '3');
 			// Below the model's floor but saying the query: kept on the
 			// lexical engine's word alone, where the index places it