Methodology

A contradiction record is a claim: someone — a historical collection, a contributor, or CodexTrace’s own research — has asserted that two or more passages conflict. Creating the record asserts nothing about whether the conflict is real.

The platform enforces this structurally:

• Any imported claim defaults to claimed_contradiction, never to a formal verdict.
• Default review status is unreviewed.
• A stronger classification requires a written formal analysis on the record.
• The UI never renders a claim as a verdict the record does not hold.

These are three different objects and are never merged.

Formal logical contradiction

A proposition and its denial are asserted of the same subject, referent, event, time, sense, scope and textual tradition. Failure on any dimension defeats the formal claim — it does not thereby resolve the passage.

Tension

The passages resist joint reading without meeting the formal test. Calling something a tension is not a euphemism for “not a contradiction”; it is a statement that the formal test is not the right instrument for it.

Textual variant

The difference lies between manuscript witnesses at a defined variant unit. This is a transmission phenomenon. A variant may explain an apparent contradiction, but it is never itself one.

A record may carry two propositions and an assessment on seven dimensions. Each takes yes, no, uncertain or not_applicable.

Propositions are written in a form that makes the conflict checkable, for example kills(chief, 800, E) against kills(chief, 300, E). If subject, event, action and counting criterion are held fixed, those two conflict. A harmonisation that separates the events does not dissolve the analysis — it adds a premise, and that premise is recorded explicitly. Premises are never injected silently.

An uncertain on any dimension is a legitimate terminal state. It is recorded, not resolved by preference.

A harmonisation is stored with the reconciling claim, the premises it must supply that the text does not state, the type of each premise (textual, contextual, historical, theological), and its sources where curated. Critical responses are stored symmetrically. Neither is rendered as the answer.

For narrative passages CodexTrace records who is speaking, at what quotation depth, and on whose authority the content is presented. Given “And YHWH said unto Moses…” the platform distinguishes:

• Textually established: the narrator reports that YHWH said X.
• Not established by grammar alone: that the reported speech historically occurred.

This is a literary distinction between narration and independently established occurrence. It is not an argument that the speech did not occur, and not an argument that it did.

Variant records define a variant unit, its readings, and — only where curated — its witnesses. An uncurated witness list renders as “Research pending”. Approximating one is prohibited: a fabricated siglum is indistinguishable from a real one to most readers, which makes it the most damaging failure mode available to a platform of this kind. Every witness row requires a source; the database enforces it with a NOT NULL constraint.

Token, lemma, morphology and gloss data are ingested only from openly licensed datasets recorded in the source registry, and always carry dataset attribution. Nothing in this layer is generated. A gloss is labelled a gloss, never a definition.

1. Primary text and manuscript witnesses.
2. Critical editions and peer-reviewed scholarship.
3. Institutional databases and catalogues.
4. Interpretive works: commentary, apologetic, skeptical, rabbinic, patristic.
5. Community contributions.

Position in this hierarchy describes evidence type, not truth. A rabbinic or patristic source is a primary witness to interpretation history even where it is a secondary witness to the text.

• primary — Primary
• critical — Critical / scholarly
• interpretive — Interpretive
• community — Community

Every page renders evidence in this order, and each block declares its layer:

1. Source & provenance — Who said it, where, and under what licence.
2. Textual evidence — Primary text and translation wording.
3. Original-language evidence — Tokens, lemmas, morphology and glosses from licensed datasets.
4. Manuscript & variant evidence — Witnesses, readings and repository records.
5. Interpretation — Attributed arguments: commentary, apologetic, skeptical, rabbinic, patristic.
6. CodexTrace analysis — Editorial classification, formal analysis, harmonisation cost.
7. AI synthesis — Model output over retrieved evidence. Never a source.
8. Community — User discussion and submissions. Untrusted by default.

A page must not render a lower layer above a higher one. A harmonisation never appears above the passages it reconciles, and an AI summary never appears above the manuscript evidence it summarises.

Review status moves through:

Unreviewed → In review → Reviewed → Disputed → Retired

A reviewer may not decide their own submission — enforced both in the application and by a database trigger. Editorial changes to research conclusions write to revision_history; nothing is rewritten silently. Duplicate detection surfaces candidates and never merges them.

There is no single objective count of contradictions in the Bible. Counts depend on the source collection, canon, translation, manuscript tradition, grouping rules, duplication rules, definition and interpretation. CodexTrace therefore never says “there are X contradictions”; it says “X contradiction claims in collection Y”.

AI may summarise, compare and route across retrieved records. AI may not:

• serve as a source;
• supply evidence that retrieval did not return;
• assert historical dependence from literary similarity;
• generate morphology, sigla, witnesses, dates, identifiers or citations;
• override the classification of a record.

Every factual claim in an answer carries a citation to a retrieved record id, and citations that do not resolve are stripped before display. When retrieval returns nothing, no answer is generated at all — the platform does not answer from model memory. Retrieved text, comments and submissions are supplied to the model as delimited data, never as instructions.

Signed-in contributors may discuss records, cite evidence inline, and submit structured evidence with sources. Submissions enter moderation and never become canonical research automatically. Community content is an untrusted layer and does not enter the AI retrieval corpus by default.

Contribute evidence →

Uncertainty is displayed, not smoothed. unreviewed, uncertain, insufficient_evidence and “not yet curated” are first-class outputs. A page with thin evidence says so.

Known limits of the current corpus:

• Witness lists for the seeded variant records are not curated, and are shown as research pending rather than approximated.
• The original-language layer ships as an editorial stub with no morphology and no Strong’s numbers until a licensed tagged dataset is imported.
• No commentary corpus is licensed for local reproduction, so the commentary layer links outward rather than reproducing text.
• Passage text coverage is partial: verses cited by seeded records are present, and others state that text has not been seeded.

Source catalogue · Licences & attribution · Deployment status