Methodology

The Corpus

RedLetters analyzes only the spoken words attributed to Jesus in the four canonical gospels: Matthew, Mark, Luke, and John. The source text is the King James Version, using public-domain red-letter markup. Parables and extended discourses spoken by Jesus are included. Narrator descriptions, retrospective quotations by others, and all epistolary, prophetic, and apocryphal texts are excluded.

The corpus is fixed. It does not expand. That constraint is the product.

What Is a Finding

A finding is a specific, sourced, defensible observation about the text. It is not a devotional thought, sermon illustration, pastoral instruction, or theological conclusion. It observes; it does not conclude. A finding may be numerical, linguistic, structural, historical, narrative, cross-gospel, or unexpectedly connected, but it must remain traceable to the corpus and honest about certainty.

Autonomous Pipeline

Every finding is generated and evaluated entirely by AI. There is no human review gate. The pipeline runs in two passes:

1. Generation pass. The research engine receives the complete system prompt, the runtime context (depth setting, published titles, rejected topics), and produces a candidate finding as structured JSON.
2. Adversarial review pass. A second AI call with a critic prompt evaluates the candidate: checks source verse accuracy, flags overconfident claims, verifies that quantitative assertions are traceable, and applies the gate thresholds.

Gate: a finding publishes if confidence is Verified or Probable and safety level is 1–3. Interpretive and Speculative findings, or findings at safety levels 4–5, go to a holding queue and publish on a rotating basis — one per week.

If the pipeline fails to produce a valid finding, the system logs the failure and no finding is published that day.

Confidence Levels

Verified

Countable, reproducible directly from the text. A second reader with the same corpus would arrive at the same count.

Probable

Strongly supported by the text, with one minor interpretive step.

Interpretive

A plausible reading, but not the only defensible one.

Speculative

A hypothesis clearly labeled as speculative and bounded by the text.

Safety Levels

1

Purely textual observation — no doctrinal implication.

2

Cross-gospel or structural observation.

3

Historical or cultural context, well-sourced.

4

Interpretive finding that touches theology or doctrine.

5

Highly speculative; must be labeled speculative.

Counting and Verification Rules

Every quantitative or pattern-based finding must have a reproducible verification run stored with method, query, result, and timestamp before it may be published. Verified confidence requires a reproducible count. The system never makes unsourced quantitative claims.

Correction Policy

Correction notes are visible and dated. No finding is silently overwritten after publish. If the underlying text analysis changes, the original finding remains with a dated correction note appended.

Excluded Sources

Modern commentary, sermons, devotionals, and secondary scholarship are excluded from AI generation context. The system works from the text alone.