Building the Westminster Proof-Text Dataset

A System Cannot Be Tested by a List of Favorite Variants

The next article begins the audit of Westminster’s doctrine of Holy Scripture. From that point forward, the series will move proposition by proposition through the Confession, Larger Catechism, Shorter Catechism, and later coherence dossiers.

The project therefore needs more than articles. It needs a ledger.

Without a dataset, famous variants will receive disproportionate attention, duplicated proof texts will be counted several times or not at all, a weakened proof may be confused with a lost doctrine, and the final verdict will depend upon memory. The problem becomes worse when one passage supports several confessional clauses or one proposition is supported by many passages.

The Westminster reconstruction will use a versioned relational dataset in which the fundamental observation is not “a verse” and not “a paragraph,” but a proposition–passage relationship.

That unit allows the project to ask the right question:

What precise work was this passage assigned to perform for this precise confessional proposition, and what happens to that work under each identified textual corpus?

Article 12 established the reproducible research protocol and its minimum evidence fields. This article turns that protocol into the operational Westminster proof-text ledger.

The Confessional Witnesses Must Be Versioned

“The Westminster Standards” refers to related documents with their own textual and ecclesiastical history.

The Westminster Assembly Project identifies a 1646 printed Confession without Scripture proofs and a 1647 edition with Scripture quotations and references annexed.¹ The distinction matters. The wording of the confession and the appended proof apparatus are historically related but not identical objects.

The Orthodox Presbyterian Church publishes the Confession, Larger Catechism, and Shorter Catechism in their American constitutional forms and also supplies editions with full Scripture proofs and a Scripture index.² Those files are excellent normalized working witnesses, but they cannot be cited as though every punctuation mark, reference expansion, or American revision stood in the first Westminster printing.

The dataset will therefore distinguish:

• Assembly doctrinal text: the original Westminster form;
• Assembly proof witness: the 1647 proof-bearing editions where available;
• American revision status: places where later Presbyterian adoption altered the doctrinal text;
• working normalized text: the identified OPC edition used for searchable clause and reference alignment;
• proof-source status: original Assembly proof, later inherited proof, cross-reference, or editorial “see” reference; and
• document: WCF, WLC, or WSC.

This prevents a later editorial proof from being attributed to the Assembly without evidence. It also prevents a historical proof from being dismissed merely because a modern edition formats it differently.

Proof Texts Are Evidence of Intended Warrant, Not Inspired Marginalia

The proof references are the best historical starting point for asking how Westminster doctrine was grounded in Scripture. They are not themselves inspired. Their presence does not end exegesis, and their absence does not prove that the Assembly knew no other relevant passage.

Three principles follow.

First, the proof list is presumptively relevant. A cited passage should be mapped to the clause for which its marker was supplied before modern alternatives are substituted.

Second, the proof list is not self-interpreting. Several passages under one marker may contribute different premises. A long reference may contain one decisive clause. A “see” reference may be confirmatory rather than principal.

Third, the proof list is not exhaustive. Westminster Confession 1.6 expressly recognizes doctrine deduced by good and necessary consequence. A system-level reconstruction must use the whole designated corpus, not only the historical marginal references.

The dataset accordingly preserves both historical proofs and canonical alternatives. They are never merged into one unlabeled list.

Atomic Propositions Come Before Proof Rows

A Westminster paragraph may contain a dozen claims joined by semicolons, subordinate clauses, qualifications, and denials. Assigning one grade to the whole paragraph would hide where the corpus succeeds or fails. Before atomization, however, Safeguard 19 requires a historical-meaning record so that modern analytical language does not replace the Assembly’s proposition.

The doctrinal text must first be divided into atomic propositions: the smallest assertions that can receive a meaningful proof and verdict. Each proposition record contains:

• a stable proposition identifier;
• document and unit, such as WCF 1.8 or WLC 3;
• original witness wording and its source-artifact form;
• controlling terms in their contemporary sense, scope, and polemical setting;
• primary historical evidence and remaining interpretive uncertainty;
• exact confessional clause;
• normalized proposition in subject–predicate form;
• qualifiers of time, agency, extent, and respect;
• explicit negations or excluded errors;
• relation to adjacent propositions;
• parallel WCF/WLC/WSC units;
• confessional reach: C, P, R, or W; and
• whether the claim is express or requires good and necessary consequence.

Each proposition is then divided into indispensable premises. Identity, deity, personality, distinction, number, exclusivity, unity, equality, relation, universality, modality, and necessity receive separate premise records whenever the exact doctrine requires them. A shortened approximation is never allowed to inherit the grade of the complete proposition.

For example, WCF 1.8 cannot responsibly be treated as one indivisible sentence. It asserts, among other things, the original languages of the two testaments, immediate inspiration, providential preservation, authenticity, final appeal in controversies, the people’s right and interest in Scripture, the duty of translation, and purposes of vernacular Scripture.

The textual question affecting “kept pure in all ages” may leave the duty of translation untouched. Atomic records allow that distinction.

One Row per Proposition–Passage Relationship

After propositions are registered, each proof passage receives a relation row. The same biblical passage may appear in several rows because it performs work for several propositions. That is not double counting if the relation identifiers remain distinct.

Each relation row records:

• proposition identifier;
• proof-source witness and marker;
• normalized biblical reference;
• exact biblical clause doing the work;
• immediate context range;
• proof function;
• express or inferential status;
• premises supplied by the passage;
• premises not supplied by the passage;
• dependency on another cited passage;
• strongest alternative proof under the critical corpus; and
• result under each textual control.

The approved proof functions are:

The function assignment is provisional until the passage is exegeted. The dataset records both the preliminary classification and the adjudicated classification, with reasons for any change.

Unique Textual Events Must Be Separated from Proof Relationships

One Greek variant can affect five Westminster citations. Five different textual variants can affect one proposition. If the series counts only citations, one variant may be inflated. If it counts only variants, the extent of confessional dependence may disappear.

The ledger therefore uses two primary identifiers:

• textual event ID: one variation unit in one textual history;
• proof relation ID: one proposition–passage relationship.

Suppose a disputed clause appears once in Scripture but is cited under WCF 2.3, WLC 9, WLC 10, and WSC 6. The system records one textual event linked to four proof relations. Cumulative reporting can then state both numbers:

One unique variant affected four historical proof relationships across three standards.

---

1. Westminster Assembly Project, “Principal Documents: The Confession of Faith”, including the 1646 text without Scripture proofs and the 1647 edition with proofs, accessed July 17, 2026. ↩
2. Orthodox Presbyterian Church, “Confession and Catechisms”, with linked editions of the Confession with proofs, Larger Catechism with proofs, Shorter Catechism with proofs, and Scripture index, accessed July 17, 2026. ↩

That sentence is more informative than either “four doctrines were changed” or “only one verse is involved.”

The Core Relational Tables

The living dataset is organized into linked tables rather than one impossibly wide spreadsheet.

1. standard_witnesses

Identifies the confessional edition, document, jurisdictional form, date, source, and revision status.

2. standard_units

Stores WCF chapters and paragraphs, WLC questions and answers, WSC questions and answers, and their normalized parallel links.

3. propositions

Stores atomic doctrinal claims, logical qualifications, exclusions, reach labels, and good-and-necessary-consequence status.

4. historical_meanings

Stores the original confessional wording, controlling term, contemporary sense, scope, excluded alternatives, polemical setting, primary evidence, source form, verification status, and uncertainty note.

5. premises

Stores every indispensable premise, its relation to the exact proposition, and its status as explicitly stated, necessarily entailed, cumulatively probable, merely compatible, not established, or contradicted.

6. proof_relations

Links each proposition to each historical or proposed biblical proof and records function, exact clause, context, and dependency.

7. textual_events

Stores one Greek or Hebrew variation unit, its scope, edition locations, and relation to documentary witnesses.

8. readings

Stores the received reading, NA28, UBS6, BHS, published BHQ status, apparatus alternatives, punctuation, morphology, and literal gloss.

9. translations

Stores ESV 2025 body, ESV footnote, E-A through E-F relationship code, KJV/AV witness, literal rendering, and departure classification.

10. arguments

Stores local exegetical premises, canonical premises, alternative constructions, hidden or inherited premises, objections, harmonizations, and repairs.

11. standard_units0

Stores the strongest materially different rival model, propositions shared with the preferred construction, the premise needed to exclude the rival, closure status, and the greatest conclusion actually warranted.

12. standard_units1

Stores claim-level certainty, derivation grade A–F, coherence code C0–C4, C/P/R/W reach, demonstrated scope, premise-status summary, closure result, no-borrowing result, negative-control result, greatest warranted conclusion, and prose rationale.

13. standard_units2

Stores creator, title, edition, date, locator, stable link, and artifact form: facsimile, scan, OCR, diplomatic transcription, corrected text, modernized edition, translation, or excerpt. It also records access and verification status and the authoritative artifact against which a quotation was checked.

14. standard_units3

Stores reviewer, article number, response decision for a materially criticized living person, jurisdiction and pastoral-scope note, correction status, revision trigger, prior finding, supersession relation, downstream recalculation flag, and change log.

A denormalized export will combine these fields for readers who prefer CSV. The relational form remains authoritative because it prevents the same textual fact from being copied inconsistently across many rows.

Required Fields for Every Published Case

No case can receive a final doctrinal grade until the following minimum record is complete.

Blank fields remain visibly blank. “Not applicable,” “not published,” “not examined,” and “undetermined” are different values.

The Locked Judgment Axes

Article 11 established four independent judgment axes. The dataset preserves them separately.

Certainty

Each textual, translational, exegetical, proof-function, and doctrinal claim receives one of five labels: established, strongly supported, probable, contested, or undetermined.

Derivation grade

The A–F grade records what the designated corpus can bear:

• A: exact proposition explicitly established;
• B: exact proposition necessarily entailed by a sound and closed cumulative argument;
• C: substantial but incomplete support; exact reconstruction fails;
• D: completed by an indispensable external premise;
• E: merely compatible or underdetermined; or
• F: contradicted or excluded.

Only A and B are successful exact reconstructions. The grade may not outrun the weakest indispensable premise.

Coherence code

The C0–C4 code records the relation among operative readings:

• C0: no relevant tension;
• C1: resolvable tension;
• C2: repair-dependent;
• C3: systemically unstable; or
• C4: strict contradiction.

Confessional reach

• C: catholic or ecumenical;
• P: magisterial Protestant;
• R: broadly Reformed; and
• W: Westminster-specific precision.

No arithmetic average will collapse these axes into one score. A proposition can be grade C, coherence C0, reach C/P, with a strongly supported judgment. Each answer means something different.

A Registered Example Without a Premature Verdict

The method can be illustrated from WCF 1.8 without deciding the later textual case.

Psalm 119:89 receives a separate proof-relation row under the same proposition. The two passages are not assumed to perform identical work. Matthew 5:18 may concern the abiding validity and verbal particularity of the law; Psalm 119:89 may concern the settled character of God’s Word. The later article must establish, not presume, the exact relation.

This example shows why the dataset is necessary. “Westminster cites Matthew 5:18 for preservation” is a useful beginning, not a completed argument.

The “Many Passages” Objection Is Built into the Test

Lane Keister has objected that interpreting one passage is not the same as formulating a doctrine, which is ordinarily based on many passages.³ That principle is correct and is now operationalized rather than used as a slogan.

When a received reading is absent from the critical main text, the dataset asks:

1. What exact premise did the affected reading supply?
2. Which other critical-text passages supply the same premise?
3. Are those passages independent, or do they presuppose the disputed reading?
4. Is the proposition stated directly or reconstructed cumulatively?
5. Does the alternative retain the Westminster precision or only a broader orthodox core?
6. Does any needed qualification come from an ESV departure or apparatus reading rather than the controlling main text?

This allows three different outcomes to remain distinct:

• the local proof weakens while the doctrine remains direct elsewhere;
• the doctrine survives cumulatively with reduced explicitness; or
• an indispensable premise disappears from the selected corpus.

The first does not justify “the doctrine is destroyed.” The third is not answered by “there are many verses.” The dataset forces both claims to display their evidence.

Negative Controls Prevent Thesis-Driven Counting

Every alleged critical-text problem receives a received-text control. The record asks whether the same difficulty appears in the received corpus, whether the received wording changes the proof function, and whether either side requires a repair.

The allowed control results are:

• shared;
• critical-aggravated;
• critical-alleviated;
• critical-specific;
• received-specific; or
• undetermined.

A case marked “shared” cannot be counted as evidence that the critical text created the problem. A received advantage must be demonstrated at the proposition and proof-function level, not inferred from the presence of extra words.

The dataset also contains a disconfirming-evidence field. If the critical corpus supplies an equally direct alternative proof or resolves an apparent contradiction more naturally, that evidence must be entered even when it cuts against the series’ thesis.

Repairs and Hidden Premises Must Be Visible

A reconstruction may use a reading from the apparatus rather than the main text, an ESV departure, a conjectural emendation, a qualification imported from another textual tradition, or a theological premise inherited from the received-text synthesis.

Some of these moves may be defensible. The methodological issue is disclosure.

The argument table therefore identifies the source of each premise:

• direct main-text statement;
• lexical or grammatical inference;
• canonical inference;
• historical information;
• apparatus reading;
• ESV departure;
• received reading absent from the critical main text;
• conjecture or emendation; or
• inherited confessional synthesis.

Grade D is possible only when an indispensable premise comes from outside the selected corpus and that dependence is demonstrated. It is never assigned merely because the interpreter already knows Reformed theology.

The closure table prevents a second form of over-crediting. A list of named examples cannot be treated as exhaustive merely because the preferred theology has long understood it that way. The record must identify a rival model, state whether that model can affirm every cited proposition, and name the textual premise that excludes it. If no such premise is present, the dataset records the strongest positive conclusion the corpus does warrant and refuses the exact grade.

Quality Control and Review

Before a case moves from draft to publishable finding, it must pass eight reviews.

Reference review

The confessional marker, biblical range, versification, and source witness are verified. Apparent differences caused by reference numbering are resolved.

Transcription review

Greek, Hebrew, editorial signs, ESV wording, footnotes, and AV wording are checked against the cited editions. Search snippets and uncited interlinears cannot bear a disputed reading.

Exegetical review

The strongest grammatical and contextual alternatives are stated. The project’s preferred theology cannot convert a possible interpretation into a necessary one.

Logical review

Premises and conclusion are tested. Contradiction claims identify subject, time, respect, and sense. “Doctrine elsewhere” claims produce the inferential chain.

Historical-meaning review

The exact confessional wording, contemporary sense, scope, and exclusions are checked against primary historical evidence before a modern paraphrase controls the audit.

Entailment and closure review

Every indispensable premise has one of the six statuses. The strongest rival model is recorded, closure is demonstrated or denied, the no-borrowing deletion test is run, and the greatest warranted conclusion does not exceed the evidence.

Source-stewardship review

Every material source discloses its artifact form, access status, and verification status. Named-person response decisions, correction obligations, lay jurisdiction, and pastoral purpose are recorded where relevant.

Verdict review

Certainty, grade, code, reach, scope, and negative-control result agree with the evidence record. A reviewer may disagree with the judgment, but no field may conceal the basis of the result.

Each review receives a name or role, status, date, and note. Unreviewed records remain provisional.

Versioning and Edition Drift

The dataset is living but not silently mutable. A new NA edition, UBS correction, BHQ fascicle, ESV revision, manuscript transcription, or stronger exegetical argument can trigger reassessment.

Each revision preserves:

• the former reading and finding;
• the new evidence;
• the new reading or judgment;
• changed certainty, grade, or code;
• affected proof relations;
• downstream articles requiring correction; and
• cumulative totals requiring recalculation.

This is especially important for Articles 148–156. Final counts will be generated from current records while retaining prior versions. They will not be adjusted impressionistically to fit a conclusion.

What the Final Dataset Must Be Able to Answer

At completion, the ledger must answer questions such as:

• How many atomic Westminster propositions were tested?
• How many historical proof relationships were registered?
• How many unique textual events affected them?
• How many effects were shared by the received control?
• How many ESV differences were translation choices rather than textual differences?
• How many proofs moved from principal to cumulative?
• How many propositions remained grade A or B?
• How many required an apparatus reading, ESV departure, or inherited premise?
• How many C4 contradictions, if any, survived full logical review?
• Which results affected catholic, Protestant, Reformed, or Westminster-specific precision?
• Which conclusions would change if NA29 or a new BHQ fascicle were adopted?

If the dataset cannot answer these questions, the final verdict is not a reconstruction test. It is a collection of essays.

Verdict

The Westminster proof-text dataset design is adopted. Its fundamental unit is one atomic proposition–passage relationship, linked to a separately counted textual event. Original and later confessional witnesses, historical and alternative proofs, source editions, translations, premises, controls, verdict axes, and revisions receive separate fields.

The “many passages” objection is incorporated as a required canonical reconstruction, while local proof loss remains measurable. No later article may infer doctrinal survival from an unspecified “elsewhere,” and no later article may infer doctrinal collapse from one changed proof without auditing the remaining corpus.

No A–F doctrinal grade is assigned to this methodological article. Grades begin when an identified proposition is actually tested.

What This Does Not Prove

Building the dataset does not prove the series’ thesis, establish that the Assembly’s proof assignments are all equally strong, or turn theological judgment into mechanical counting. A database can preserve a bad argument as efficiently as a good one.

The design does not make proof texts inspired, limit doctrine to historical marginal references, or deny good and necessary consequence. It ensures that every conclusion can be traced from confessional proposition through textual evidence and inference to a bounded verdict.

Text and Edition Record

• Primary confessional historical witness: Westminster Assembly proof-bearing editions, beginning with the 1647 Confession edition identified by the Westminster Assembly Project.
• Normalized working witnesses: OPC Confession, Larger Catechism, and Shorter Catechism editions with full Scripture proofs and Scripture index.
• Confessional version fields: Original Westminster, American revisions, normalized working text, and proof-source status.
• Textual controls: Scrivener 1894, Stephanus 1550, Beza 1598, Ben Hayyim/Bomberg 1524–25, NA28, UBS6, BHS 1997, and published BHQ fascicles.
• English controls: ESV 2025 and an identified Authorized Version witness; NKJV where a received/critical footnote comparison is useful.
• Judgment axes: Five certainty labels; A–F derivation; C0–C4 coherence; C/P/R/W reach; demonstrated scope.
• Edition-date lock: July 17, 2026.

Bibliography

• Keister, Lane. “An Answer to Chris Thomas.” Green Baggins. November 26, 2024.
• Orthodox Presbyterian Church. “Confession and Catechisms.”
• Westminster Assembly. The Confession of Faith, with the Quotations and Texts of Scripture Annexed. London, 1647.
• Westminster Assembly. The Larger Catechism. 1648.
• Westminster Assembly. The Shorter Catechism. 1648.
• Westminster Assembly Project. “Principal Documents: The Confession of Faith.”

Revision History

• Version 2.0 — July 18, 2026: Rebuilt the data architecture under Safeguards 1–19; added historical-meaning, indispensable-premise, countermodel/closure, source-form, stewardship, response, correction, and supersession tables; adopted the six premise statuses and revised A–F scale; required greatest-warranted-conclusion and no-borrowing fields; and expanded review gates from five to eight.
• Version 1.0 — July 17, 2026: Adopted the relational Westminster proof-text dataset, proposition–passage and textual-event identifiers, proof functions, required fields, review gates, and versioning rules.