How Red Letter works

Flag methodology

Each church is scored on two independent 0–10 axes — MAGA Alignment and Apostasy — from public material (website, sermons, official socials, the lead pastor, the church-finder listing). The axes are reported as a pair and never averaged: a church can be high on one and zero on the other. Every nonzero criterion requires a quoted source; low-coverage scores are marked provisional. This page describes rubric v0.9; the signed-in map carries the current rubric.

Bands (per axis):

• None (0–1)
• Mild (2–3)
• Moderate (4–6)
• High (7–8)
• Severe (9–10)

How scores are produced

Scores come from an automated, human-audited pipeline that reads publicly available web content only — the church’s own website, published sermons (audio and video transcribed to text), official social accounts, and the church-finder listing. No private, leaked, or anonymously sourced material is used.

1. Extraction — candidate quotes are pulled from the captured material against the rubric’s criteria (claude-sonnet-4-6).
2. Independent verification — every candidate flag is re-checked by a separate model (claude-haiku-4-5-20251001) that confirms or refutes the quote against its source before it can influence any score.
3. Scoring — verified evidence is scored against the published rubric on this page (claude-sonnet-4-6); the arithmetic (weights, floors, trip-wires) is deterministic and auditable per church.
4. Deep dive (selected churches) — an opt-in adversarial multi-lens review (lenses: claude-sonnet-4-6, verified by claude-haiku-4-5-20251001, synthesized as prose by claude-opus-4-8 — synthesis never classifies or scores).

Two structural rules hold everywhere: every nonzero criterion requires a verbatim quoted source (no quote, no points), and the two axes are independent and never combined into any composite number.

Score formula

For each axis, the model assigns a 0–10 subscore to each criterion. The pipeline then derives the final axis score using a severity floor, so that one pervasive criterion can define the axis on its own:

weighted_sum   = clamp(0, 10, round( sum( weight_i * subscore_i ) ))
severity_floor = max non-affiliation subscore (the worst single criterion sets the minimum)
tripwire_floor = highest floor from any fired trip-wire (0 if none fired)
floor          = max(severity_floor, tripwire_floor)
score = max(weighted_sum, floor)   [then clamped 0-10]

For MAGA only, a flag bonus (0–2) for the christian_nation criterion is added to that subscore before computing the weighted sum (clamped to 10). Each church’s record in the map shows the weighted sum, the flag bonus, the floor and the reason for it on each axis, so the arithmetic is fully auditable. The floor is often the dominant term: a church with one criterion at 9 scores 9 even if the weighted average of all its criteria is only 6.

MAGA Alignment criteria

Apostasy criteria

Coded-language lexicon (internal recall aid)

The rubric includes a coded-language lexicon used during ingestion as an internal recall aid, to help recognize indirect phrasing. It is not a displayed flag and not a scorer input — it routes candidate phrases to criteria for the scorer to confirm or reject with a direct quote. The lexicon is not reproduced here, to avoid treating a recall tool as a public checklist.

Trip-wires (auto-floor the axis)

False-positive guards

• Patriotism != nationalism: praying for leaders, a July-4 mention, a small interior flag = low; sacralizing the nation = high.
• Traditional-but-welcoming = 0 on moral_revisionism (historic view while excluding no one). The flag is affirmation, not having a conviction.
• Policy view != contempt: score dehumanization of people, not a position on law.
• Silence != denial: score creedal only on explicit/functional denial, never on omission.
• Affiliation is a PRIOR: SBC/network points alone never push MAGA above the cap without a political signal.
• Do not score off the homepage alone.
• Each axis requires its own distinct evidentiary element — never the same clause counted twice. Score partisan_hostility (MAGA) on framing opponents as enemies/evil; score sanctified_violence (apostasy) ONLY when a statement endorses/glorifies/calls for actual physical harm against a present-day group. Demonizing rhetoric with no violence endorsement = MAGA-only. Violence against a non-political outsider with no partisan frame = apostasy-only. For partisan_media vs partisan_hostility: the hostile CONTENT is partisan_hostility; partisan_media fires ONLY on the separate act of treating a media source as authoritative. A church may score one axis, the other, or both — but each score must rest on its own evidence, not a shared sentence.
• partisan_hostility is DIRECTIONAL toward the MAGA-aligned side, and direction is inferred from CONTENT, not only from a named party: most real MAGA hostility is coded/issue/theological (woke, globalists, groomers, open borders, 'child sacrifice', 'the media', stolen election) and names no party — score it; do NOT zero a quote merely because it does not say 'Democrat'. Conversely, criticism aimed AT Trump / a MAGA-aligned administration / ICE / the MAGA movement, and advocacy for its targets, is the opposite direction = 0 however intense (including coded 'empire'/'the state'). Score the speaker's endorsed stance, never quoted material they rebut. 'The current administration' resolves by sermon date.

Source weighting

Limitations & accuracy

• Automated analysis can err. Extraction, transcription, and scoring are model-assisted; a model can mis-hear audio, misattribute a speaker, or miss context despite the independent verification pass.
• Coverage varies. Each church carries a coverage-confidence band (high / med / low) derived from how much of its public footprint was actually captured; a score whose source coverage is below 0.5 is marked provisional everywhere it appears.
• Point-in-time capture. Content reflects what was publicly available when it was retrieved and may be stale; churches change leadership, teaching, and websites.
• Scores are editorial opinion. The quoted material is reproduced verbatim from public sources; the scores, bands, and criteria assignments are this project’s published, criteria-based analysis and editorial opinion grounded in that quoted material — not assertions of fact beyond the quotes themselves.
• Unreviewed pins carry no score. Churches whose material was downloaded but not yet verified are explicitly labeled unreviewed and are never scored or shaded.

Corrections & right of reply

Anyone — including a church’s leadership or authorized representatives — may dispute the accuracy, attribution, or context of any quote or score. Demonstrable errors are corrected and the church is re-scored. Disputes that are not resolved can be noted alongside the record.

Request a correction

No account is needed. If you are signed in to the map, you can also use “request a correction” on a church’s record there.

Red Letter is an independent research project. It is not affiliated with any church, denomination, or political organization.