What RedLetters is

RedLetters publishes one daily research observation from the red-letter words of Jesus. A finding is meant to show something textual, sourced, and interesting. It observes before it interprets.

The site is designed for human reading first. The underlying AI metadata is preserved for auditability, but public labels translate that metadata into clearer reader-facing language.


The corpus

RedLetters works from one fixed source: the spoken words of Jesus in Matthew, Mark, Luke, and John, King James Version. Nothing else enters the analysis — no epistles, no commentary, no narrator descriptions, no retrospective quotations by others.

The corpus never expands. That constraint is the product. As the obvious observations are exhausted, the findings get harder to find and more interesting to read.

Source: CrossWire SWORD KJV module, Words of Christ markup — tracing to Louis Klopsch's 1901 Red Letter KJV. Red-letter boundaries are editorial judgments, not manuscript features. Disputed passages (for example John 3:16–21) follow the Klopsch/CrossWire convention. 1,859 citable verses, drawn from a legacy full listing of 1,952 red-letter verse references. Public domain.

Live gate stats

The gate is fully autonomous: candidates are generated by AI, checked mechanically, reviewed by a separate AI pass, and only then allowed into the publishing runway. There is no human approval gate.

Published
112 findings have earned publication.
Rejected
141 candidates failed the gate.
Held in reserve
64 evaluated candidates are preserved as research reserve, not a human review queue.
Scheduled
7 gate-passing findings are currently in the runway.

What counts as red letters

On finding pages and in emails, quoted source text is rendered in red only when the words are part of the fixed red-letter corpus. Narrator text, commentary, and words spoken by other people are not styled as red-letter words.

This makes the visual rule simple for readers: if the quoted Bible words are red, RedLetters is treating them as words of Jesus under the fixed corpus convention. If they are not red, they are not part of that red-letter span.


Research lenses

Each finding has a lens. The lens is the reading angle: what kind of observation is this? Lenses are intentionally public-facing because they help readers understand why the finding is interesting.

Counting
Numerical/Statistical

This finding depends on counts, frequency, proportions, or measurable absence/presence.

Word Use
Linguistic/Wordplay

This finding is about a word, phrase, repeated term, translation choice, or language pattern.

Pattern / Structure
Pattern/Structural

This finding notices arrangement, repetition, sequence, symmetry, or form.

Contrast
Contrast/Tension

This finding turns on two sayings, themes, or Gospel presentations held in tension.

Historical Context
Historical/Cultural

This finding involves cultural setting, practices, geography, roles, or ancient-world context.

Scene / Drama
Narrative/Dramatic

This finding depends on the story moment: who is present, what just happened, or how the scene turns.

Conceptual Surprise
Philosophical/Unexpected

This finding notices a surprising idea, category shift, paradox, or unexpected framing.

Across Gospels
Cross-Gospel

This finding compares Jesus's words across Matthew, Mark, Luke, and John.


Sources inside each finding

Every finding lists the red-letter sources it depends on. When possible, the page also prints those verses directly below the source list, with only Jesus's words colored red.

Evidence notes and Greek notes are supporting context. They are not the finding itself; they explain why the observation is plausible, reproducible, or worth reading carefully.


How it works

Every finding is generated and reviewed entirely by AI. There is no human approval gate.

1
Generation. The research engine receives the fixed corpus, depth setting, and the list of already-published and rejected topics. It produces a structured candidate: title, hook, insight, source verses, evidence notes, lens, confidence, and safety level.
2
Mechanical validation. Deterministic checks run before AI review: every source verse must exist in the red-letter corpus, quantitative claims require verification output, and required fields must be present. A narrator verse in source_verses fails automatically.
3
Adversarial review. A second AI call evaluates novelty, clarity, conceptual overreach, and product fit. Mechanics are handled upstream; the reviewer judges whether the finding is genuinely interesting and honest.
4
Publication. Findings that pass the gate publish on schedule. Corrections are public and dated; the original is never silently overwritten.

AI review labels

Confidence and safety level are AI-given review metadata. They are preserved for auditability, filtering, and future machine analysis, but they are not the main way a human should read the finding.

On the public site, these fields appear quietly as an AI review line near the publication date. Support describes how directly the finding is backed by the corpus. Interpretive reach describes how much care is needed before turning the observation into a larger theological claim. It is not a warning label.

Reproducible support
Internal label: Verified. Countable and reproducible. A second reader with the same corpus arrives at the same result.
Strong textual support
Internal label: Probable. Strongly supported, with one minor interpretive step.
Plausible reading
Internal label: Interpretive. A plausible reading — not the only defensible one.
Exploratory hypothesis
Internal label: Speculative. A hypothesis clearly bounded by the text and labeled as such.
Safety levels: Safety 1–5 is internal AI review language for interpretive reach. Level 1 is purely textual. Level 5 is highly speculative and always labeled. Public pages translate this into phrases like "low interpretive reach" so readers do not mistake it for a danger score.

What this is not

RedLetters findings are not pastoral advice, theological conclusions, or the position of any church or denomination. They are research observations from a fixed text, generated and evaluated by AI, and disclosed with source verses, research lenses, and AI review labels.

AI-level detail

Internal fields are preserved as category, confidence, and safety_level. Public labels translate those fields without changing the underlying finding record, so historic and future findings remain compatible.