--- title: "How We Cross-Reference Scripture" category: "other" description: "Scripture links on this site are derived from the biblical texts themselves rather than copied from a cross-reference list — how that works, the five kinds of evidence it produces, and why they are never merged into a single score." tags: ["data", "method/textual-criticism", "cross-references", "septuagint", "transparency", "mermaid"] draft: false date_created: 2026-09-04 date_modified: 2026-10-20 ai_provider_models: - anthropic/claude-opus-5 - anthropic/claude-opus-5.5 --- # How We Cross-Reference Scripture Most cross-reference lists are inherited. A study Bible's centre column, the *Treasury of Scripture Knowledge* and a crowd-voted dataset each record what earlier readers saw as connected. This site uses two such lists. They have two limits: each reflects the interests of the people who compiled it, and none can show a connection nobody wrote down. So the links here are also **derived**: computed from the biblical texts in their original languages, by a method anyone can check. This page explains how the method works, what it can and cannot show, and why that matters when you study a passage. To skip the method and look up a verse: [Look up a verse in Scripture Links](https://the-way.lewy.au/references/){ .md-button } ## Study outline - [Why a list is not enough](#why-a-list-is-not-enough). Hebrews 10:5 and Psalm 40:6, and what a list cannot explain. - [Five kinds of link, kept separate](#five-kinds-of-link-kept-separate). The classes, and why they are never added into one score. - [How the matching works](#how-the-matching-works) and [where the thresholds come from](#where-the-thresholds-come-from). The measures, the cut-offs, and the rules added in October 2026. - [What each rule adds, and where it fails](#what-each-rule-adds-and-where-it-fails). How far to trust each rule, what only it finds, where the rules agree, and what all of them miss. - [What the method produces](#what-the-method-produces) and [what this gives you in a study](#what-this-gives-you-in-a-study). The counts, and the gaps the links reveal in a study. - [Compared with an inherited list](#compared-with-an-inherited-list), [tracing one verse](#tracing-one-verse) and [the Dead Sea Scrolls](#checking-against-the-dead-sea-scrolls). - [Limits](#limits) and [where the data lives](#where-the-data-lives). ## Why a list is not enough Hebrews 10:5 says *"a body you have prepared for me"* and is quoting Psalm 40. Turn to Psalm 40:6 in an English Bible and it reads *"you have opened my ears"*. Both translations are correct. The author of Hebrews is quoting the **Septuagint**, the Greek Old Testament, and at this verse the Septuagint reads differently from the Hebrew. A cross-reference list tells you the two verses are connected. It cannot explain why they read differently, and it cannot find other cases like it. Only the texts themselves can. ## Five kinds of link, kept separate The method produces five kinds of link. Each is stored on its own and they are **never added up into one relevance score**. They rest on different evidence, and the kind of evidence decides what a link can be used for. ```mermaid flowchart LR subgraph textual["Textual fact — the same language on both sides"] direction TB QG["quotation-greek
Greek New Testament quoting
the Greek Old Testament"] IB["inner-biblical
the Hebrew Old Testament
quoting itself"] XL["quotation-crosslang
the Greek New Testament read
against the Hebrew Old Testament"] AL["allusion-lemma
shared rare vocabulary,
no shared phrasing needed"] QG ~~~ IB ~~~ XL ~~~ AL end subgraph judged["Someone's judgement — useful, but not the same thing"] direction TB QH["quotation-hebrew
a 19th-century Hebrew New Testament
matching the Hebrew Old Testament"] XR["cross-references
crowd-assembled, inherited"] QH ~~~ XR end textual --> USE(["What a study may lean on"]) judged --> LEAD(["Leads worth chasing"]) ``` **Greek quotation.** The New Testament writers wrote in Greek and read their Old Testament in Greek. When Paul quotes Isaiah, both sides are in the same language, so a quotation shows up as the same words on the page. Anyone can check it. **Inner-biblical quotation.** The Hebrew Old Testament quotes itself often. The Hezekiah account appears in both 2 Kings and Isaiah, the Ten Commandments in both Exodus and Deuteronomy, and Chronicles retells Kings. Both sides are Hebrew, so these too are checkable on the page. **Allusion by rare words.** Sometimes one passage points to another without quoting it. Revelation 21:20 lists the jewels of the new Jerusalem, and Ezekiel 28:13 lists the jewels of the king of Tyre. They share no phrases, so a quotation search cannot connect them. They do share words that appear almost nowhere else in Scripture, and that is what this kind of link measures. **Hebrew New Testament matches.** Two 19th-century scholars each translated the Greek New Testament into Hebrew. Where their Hebrew for a quotation uses the Old Testament's own wording, that is a strong hint. It is still a translator's judgement about the text, so these links are labelled as candidates everywhere they appear. ## How the matching works The technique comes from text-reuse detection, the same field that produces plagiarism checkers. It is adjusted for Scripture, which quotes itself openly and often changes the wording as it does. ```mermaid flowchart TD subgraph prep["Prepare the texts"] direction TB A["Both texts, same language"] --> B["Normalise — accents and case removed"] --> C["Index every 4-word sequence"] end subgraph match["Find candidates"] direction TB D["Pairs sharing those sequences"] --> E["Score by local alignment, rarity and overlap"] end subgraph check["Corroborate"] direction TB F["Check against an unrelated reference list"] --> G(["Graded, typed link"]) end prep --> match --> check ``` Three steps in that process decide what it finds. **Accents are removed first.** Brenton's 1851 Septuagint and the modern SBL Greek New Testament place accents differently, so identical words would fail to match as printed. Removing accents before comparing lets them match. **Matches allow for edits.** Biblical writers adjust what they quote. Peter adds *"in the last days"* to his quotation of Joel, and Luke leaves a clause out of Isaiah 61. A measure that only counts unbroken runs of identical words fails at the first change. Such a measure misses **Matthew 1:23 quoting Isaiah 7:14**, one of the most important quotations in the New Testament, because Matthew's longest word-for-word stretch there is five words. Local alignment allows for added and missing words, and finds it. **Every link is checked against outside lists.** Each derived link is compared with two cross-reference lists that have no connection to this method or to each other: a modern crowd-voted set and the World English Bible translators' footnotes. When they agree, two independent sources point the same way. When a strong textual match appears in neither, the method has found a connection the tradition did not record. The footnote list is small, a few hundred notes against 830,000 crowd-voted links, but it concentrates on New Testament quotations of the Old, the same ground these links cover. It caught three quotations the larger list had missed. A third list, unfoldingWord's Translation Notes, is kept out of this check, because the rules below were tested against it. ### The measurements, by name The first pass is an index of every four-word sequence. A pair of verses must share at least two of them before it is scored. This keeps the work manageable: 7,939 Greek New Testament verses against 22,948 Septuagint verses makes 182 million possible pairs, far too many to align one by one. The index finds candidates cheaply, and the slower scoring runs only on those. The unique-phrase rule adds candidates from three-word phrases, and the Hebrew is indexed on dictionary forms. Each candidate is measured four ways. The four scores are stored side by side and never summed: | measure | what it says | |---|---| | `alignment` | **Smith-Waterman local alignment**, the main measure. It comes from biology, where the same problem appears in comparing DNA: find the best-matching stretch *inside* two longer sequences, allowing for added and missing pieces | | `longest_run` | the longest run of identical consecutive words. Kept because a reader can check it by hand (count the nine words), though it fails at the first edit | | `containment` | the share of the quoting verse's four-word sequences that also appear in the source, which shows how much of the verse *is* quotation | | `idf_overlap` | shared sequences weighted by how rare they are, so a common phrase like *"and it came to pass"* counts for little and a rare phrase counts for a lot | Allusions by rare words are scored differently, because there is no shared phrasing to measure. A word that appears in 30 verses or fewer across the whole Greek corpus counts as distinctive. Two passages are linked when the distinctive words they share add up past a set weight. ## Where the thresholds come from The outside lists show how often a link at each strength is confirmed by someone else, and each cut-off sits where that rate drops. The chart shows it for the Greek quotations: every pair the method scores, by alignment, and the share an outside list confirms.
Pairs aligning at 20 or more are confirmed 86% of the time, at 18-19 75%, at 15-17 49%, and at 10-14 15%. The line for a strong alignment is 18, the last band where three links in four are confirmed. For a baseline, a verse from each Testament picked at random is confirmed **0.08%** of the time, so even the weakest rule below is hundreds of times better than chance. ### Rules added in October 2026 Testing the links against two answer keys with no connection to the method, the WEB translators' quotation footnotes and unfoldingWord's Translation Notes, showed the strong links found only about one in six of the quotations those keys mark. The gaps had causes, and each cause now has its own rule. Every link records the rule that admitted it, and the page says so under each link ("Found by..."). - **Unique phrases.** A run of everyday words can be unique as a phrase. John 10:34 shares four common words with Psalm 82:6 (*"I said, you are gods"*), and the phrase occurs in one Septuagint verse. Three such phrases admit a link on their own. Phrases here are three words long, and a phrase found in twelve or more New Testament verses does not count. - **A citation formula.** A verse that says "it is written", "the Scripture", "that it might be fulfilled", in that verse or the one before, sharing two such phrases with one Old Testament verse. This is what finds the short, famous quotations: James 2:8, Romans 1:17, Matthew 4:7, Matthew 2:6 from Micah. Revelation is exempt, because its "it is written" is a name written in the book of life. - **Hebrew on dictionary forms.** Matching the Hebrew on MACULA's lemmas, with prefixes split and spelling normalised, and counting a weaker match whose neighbouring verses also match, because borrowed passages come in runs. A pair needs a rarity-weighted score of 36, or 22 with a matching neighbour on the same diagonal. 1 Chronicles 16:8 and Psalm 105:1 are the same words with one spelling difference; they scored 12 on the surface and now match. The sixteen best-known parallels went from 68 of 266 verses found to 211, at the same rate of confirmation as before. - **Reading the Greek against the Hebrew.** A vocabulary learned from the Septuagint's own translation choices lets a New Testament verse be compared with the Hebrew directly, for quotations that follow the Hebrew where the Septuagint reads differently: "Death is swallowed up in victory" (1 Corinthians 15:54; Isaiah 25:8), "I will strike the shepherd" (Matthew 26:31; Zechariah 13:7). Only in verses that cite Scripture, only the best match, and only with a clear margin over the second best. This is its own class, `quotation-crosslang`. [What the Scripture Links Reveal](https://the-way.lewy.au/scripture/what-the-links-reveal/) reads the results as a whole, including which gaps these rules closed and which remain. No single link is certain because of this. A link marked strong has passed one of these rules, and the Scripture Links page shows beside each link how often links found the same way are confirmed. ## What each rule adds, and where it fails Every rule answers to three questions: how often an outside list confirms what it finds, what it finds that no other rule does, and where it goes wrong. The first is trust; the second is the reason to keep it. A rule that is always confirmed and only ever finds what the others find adds nothing, and a rule with a lower rate can still be the only way to a real connection.
Read the chart from the right. The Hebrew New Testament is confirmed almost every time but adds little of its own, because a translator who knew the tradition rendered the quotations the tradition already knew. Unique phrases sit high and to the right: trusted, and the only route to 35 quotations in the answer keys. Alignment, the original measure, now finds just four quotations nothing else does. Shared rare words sit furthest left, confirmed 46% of the time, yet 89 of their links come from no other rule, which is why they are kept as a class of their own and labelled as echoes. In the Hebrew, chaining a run of verses is the least confirmed rule and the one that finds the most: 93 verses of the known parallels are reached by it alone. ### The New Testament rules
### The Old Testament rules
### Where the rules agree When several rules find the same connection, it rests on more than one kind of evidence. Each column below is one combination of rules, and a column with a single dot is what one rule contributes on its own.
Agreement tracks confirmation. Alignment with unique phrases is confirmed 91% of the time, and a connection found by three or more rules 64 times in 66; alignment alone, 47%. In the Hebrew, every link the printed wording finds is also found on dictionary forms, which is the rule working as designed: surface matches the lemmas did not confirm were formulas, and they were dropped.
### What every rule misses The answer keys name quotations independently of the method, so they also show what it cannot find. Each square is one quotation the WEB translators' footnotes or unfoldingWord's Translation Notes name.
The rules together find 176 of 477, about three in eight. The hollow squares fall into a few kinds: - **Allusions with no shared phrase.** The Beatitudes echo Isaiah and Psalm 37 (Matthew 5:3-5), and Hebrews 11:33-37 recalls Daniel's lions and Elijah's widow in a few words each. - **Quotations too short to carry a phrase.** "You shall not murder" is two Greek words (Matthew 5:21; Exodus 20:13). - **A phrase shared with a second verse.** Romans 4:3 introduces Genesis 15:6 as Scripture, but its closing words in Greek, "credited to him for righteousness", also stand in Psalm 106:31, so the phrase is not unique and the formula rule cannot use it. - **Quotations that follow the Hebrew in few words**, such as Matthew 2:15 and John 19:37. - **What leaves no Greek to match**: the cry from the cross (Matthew 27:46) and Daniel's "abomination of desolation" (Matthew 24:15).
For the Old Testament, the rules find 211 of the 266 verses in sixteen well-known parallels. The misses are short commandments ("You shall not murder" is two Hebrew words), the family lists of Ezra 2 and Nehemiah 7, whose numbers differ (775 sons of Arah in Ezra 2:5, 652 in Nehemiah 7:10), and the single line Joel and Amos share. Each chart lists its misses, and each opens in Scripture Links. ## What the method produces Every link is stored in one table, `scripture_links`, in the site's reference database. The table is rebuilt from scratch on each build. Nobody edits it by hand, and running the build again produces the same rows. The five kinds of link, the texts compared, and the current counts of strong links: | class | from | to | links | |---|---|---|---| | `inner-biblical` | Westminster Leningrad Codex, matched on MACULA lemmas | itself | 2,054 | | `quotation-greek` | SBL Greek New Testament | Brenton's 1851 Septuagint | 246 | | `allusion-lemma` | MACULA Greek lemmas | Septuagint lemmas | 103 | | `quotation-hebrew` | Delitzsch Hebrew New Testament | Westminster Leningrad Codex | 49 | | `quotation-crosslang` | MACULA Greek (Strong's numbers) | Westminster Leningrad Codex, through a lexicon learned from the Septuagint | 48 | All of these source texts are openly licensed, so anyone can rebuild the table. The [source catalogue](https://the-way.lewy.au/about/about-our-datasets/) gives each one's licence and origin. The `inner-biblical` count is the largest because the Hebrew Bible quotes itself so often. The method finds the well-known parallels without being told about them: 2 Kings 19 ↔ Isaiah 37, Deuteronomy 5 ↔ Exodus 20, 2 Samuel 22 ↔ Psalm 18, Psalm 14 ↔ Psalm 53. Looking up a verse returns the evidence for each link: ```bash uv run python query.py trace Heb 10 5 ``` For each connection it prints how it was found, how strong it is, the linked verse in its original language, an English translation, and the words the two verses share. The site's authoring tools use the same lookups. A second script works the other way: given a study, it lists passages connected to the study's verses that the study never cites. ## What this gives you in a study **It finds connections the lists missed.** In the method's first version, of the 140 strong Greek quotations, ten (about one in fourteen) appear in neither outside list. That count is cautious. Twenty-three of the 140 are missing *at that exact verse*, but in 13 of those a list records the same quotation one or two verses earlier, so only ten are true gaps. **Matthew 11:10** is the clearest example. *"Behold, I send my messenger before your face"* is usually linked to Malachi 3:1, correctly. Its first nine Greek words are also word for word Exodus 23:20: *Ἰδοὺ ἐγὼ ἀποστέλλω τὸν ἄγγελόν μου πρὸ προσώπου σου*. Matthew has combined two passages. Both lists record the Malachi half and neither records the Exodus half, even though the nine-word match is easy to see once you look. Luke 1:31 is similar. Every list links it to Isaiah 7:14, and none links it to Genesis 16:11, where the angel uses the same words to Hagar about naming Ishmael. **It reaches the deuterocanon.** Our Septuagint data covers Wisdom, Sirach and Maccabees, so the method finds connections Protestant reference lists leave out: Paul at the Areopagus (Acts 17:29) echoing Wisdom 13:10, and Hebrews 11:5 on Enoch echoing Sirach 44:16. Whatever their authority, the New Testament writers knew these books, and seeing where their language appears is part of reading the text carefully. **It shows what a study left out.** Every study on this site lists the passages it covers. The same data can be run the other way to ask *what connects to these passages that this study never mentions?* That check found [The Way](https://the-way.lewy.au/jesus/the-way/) citing Hebrews 3 without naming Psalm 95, which Hebrews 3 quotes three times at its three most important points. **It shows which Old Testament text a writer was quoting.** Hebrews 10:5 follows the Greek; Matthew 21:5 is closer to the Hebrew. That tells you how the New Testament writers read their Scriptures, and the links record it for each quotation. ## Compared with an inherited list The usual comparison is the *Treasury of Scripture Knowledge* (TSK), published in 1836. It is in the public domain and is still the most complete chain-reference set in English. It is not in our database: the outside check described above uses a modern crowd-voted set, with the World English Bible's footnotes as a second, much smaller source. TSK and the derived links differ in four ways. TSK is better on one of them; the derived links are better on the other three. **TSK finds shared meaning.** Its compilers linked passages on the same theme in different words, such as a mercy in Exodus and a mercy in Luke with no words in common. Our method matches words, so it cannot see those links. For thematic study, use TSK. **Derived links are graded.** TSK gives a verse forty references side by side, with no way to tell the quotation from the faint echo. Here each link carries a score, so `alignment 40` and `alignment 18` say different things, and the cut-off between them was set by measurement. **Derived links show which text the author was reading.** TSK is keyed to an English Bible. It points Hebrews 10:5 to Psalm 40:6 and stops. It cannot show that the two verses read differently because Hebrews quotes the Greek, since the list contains no Greek. The derived link records it (see the Hebrews 10:5 example at the top of this page). **Derived links can be extended and run in reverse.** TSK covers the Protestant canon and was finished in 1836. It cannot reach Wisdom or Sirach, cannot tell you what a study failed to cite, and cannot be rebuilt when a better text becomes available. Links computed from the texts themselves can do all three. ## Tracing one verse Ask where Hebrews 10:5 comes from, and the answer shows the Greek of both verses, the words they share, and how an English Old Testament renders the same verse: > **Hebrews 10:5** — "…but you prepared a body for me." > **quotes Septuagint Psalm 39:7 = Psalm 40:6**, sharing > *Θυσίαν καὶ προσφορὰν οὐκ ἠθέλησας, σῶμα δὲ κατηρτίσω μοι* > **English Psalm 40:6** — "Sacrifice and offering you didn't desire. You have opened my ears." The two Greek texts share the same words. The English Psalm says something else because it is translated from the Hebrew. Because links are grouped by how they were found, two more things show up. When a verse draws on **two** sources, both appear: Matthew 21:5 quotes Isaiah 62:11 and echoes Zechariah 9:9, and the trace returns both. And when a quoted Old Testament verse is also found in the Dead Sea Scrolls, any scroll reading that the Masoretic Hebrew lacks is attached to it, so you can see straight away whether the quoted text is disputed. ## Checking against the Dead Sea Scrolls The same database holds the biblical Dead Sea Scrolls and can show where a scroll reads differently from the Masoretic Hebrew. Deuteronomy 32:8 is the standard example: 4Q37 reads *"sons of God"* where the Masoretic has *"sons of Israel"*, and both the Septuagint and the New Testament follow the scroll. The scrolls are fragmentary, so two safeguards are built into the reports. **Forty-six per cent of the letters in this corpus are a modern editor's reconstruction**, and only a third of words survive with every letter intact. A reading therefore counts only when the differing word itself is whole. Every reading also reports how much of its verse survives, so you can tell a well-preserved variant from one legible word on a torn line. Deuteronomy 32:8's reading is clear, but only two words of that verse survive, and a study citing it should say so. The comparison runs in one direction. A word the scroll has and the Masoretic lacks is a variant reading. A word the Masoretic has and the scroll lacks is almost always damage to the scroll, and counting it would turn holes into omissions. ## Limits [What the Scripture Links Reveal](https://the-way.lewy.au/scripture/what-the-links-reveal/) reads these results as a whole: the most-quoted verse and chapter, the book that quotes most densely, and the well-known connections that fall under the threshold or outside the data. Some connections no word method reaches. Matthew 24:15's "abomination of desolation" shares three or four words with Daniel in forms that vary between Daniel's verses. Matthew 27:46 gives Jesus' cry in Hebrew and Aramaic written out in Greek letters, so there is no Greek to match. Jude 14-15 quotes 1 Enoch, which is not among the texts compared. And no links are computed from one New Testament book to another. The links do not interpret. A link says two passages share wording or rare vocabulary. Whether the second fulfils the first, whether the connection is typological, and whether the author meant it are questions for the study, and the tools leave them there. The links do not replace reading the passage. Like a concordance, they tell you where to look. The links only go as far as we can verify. Where the Septuagint's chapter numbers cannot be reliably matched to an English Bible's (parts of Jeremiah, most of Proverbs), the link is recorded with no English reference, so it never points you to the wrong verse. ## Where the data lives The reference database is built from openly licensed sources and rebuilt from scratch on every build, so anyone can reproduce it. The [source catalogue](https://the-way.lewy.au/about/about-our-datasets/) lists those sources and their licences. The [translations page](https://the-way.lewy.au/scripture/translations/) covers the texts themselves and where their verse numbering differs.