# Direct slogan study: streamlined protocol

September 4, 2026. This replaces the initial two-reader pilot plan at the user's
request to use Light and avoid unnecessary rigor or time. The original pilot
records remain intact but are not pooled with the new uniform review.

## Fixed question, population, and evidence

Review all 187 religious-versus-skeptical debates in the frozen 253-assessment
snapshot. The source fingerprints have all been authenticated. Read each whole
retained transcript, not merely the scored argument cards. Original scores,
assessment records, rankings, and the other six papers remain unchanged.

One fresh, source-only Astra review at low (Light) reasoning effort is used for
every debate. The review receives no old scores, critiques, tags, slogan results,
or expected direction of difference. A/B speaker labels have an outcome-independent
random order; source identities remain recognizable, so this is not identity-blind.
The five completed medium-effort pilot pairs are excluded from the final estimates.

## The same substantive distinctions

- **Slogan form:** a compact or repeatable affirmation presents a general conclusion
  as settled. Brevity, confidence, repetition, humor, metaphor, or emotion alone
  does not make a defective argument.
- **Support substitution:** in its actual use, the affirmation replaces a reason
  the speaker has not supplied. A brief summary of reasons given elsewhere is not
  such a substitute. A disputed premise or weak ordinary argument is not automatically
  a slogan. Check the surrounding exchange and later qualifications.
- **Protection from criticism:** positive textual evidence shows how objections or
  contrary outcomes are neutralized regardless of their merits. Examples include
  treating every outcome as confirmation, treating disagreement itself as proof of
  a critic's defect, making a promised test impossible to fail, and invoking
  unrestricted unseen intentions or compensation so no contrary pattern can matter.
  Merely not stating a falsifier, expressing faith, making a moral/metaphysical
  claim, or offering weak evidence is insufficient. Philosophical and moral claims
  can be open to reasons without a decisive laboratory test.
- **Emotional language:** observable comfort, fear, shame, identity, or ridicule.
  Record this separately. It establishes neither motives nor audience effects.

An unsupported slogan requires form=yes and substitution=yes. A criticism-resistant
slogan additionally requires protection=yes. Preserve unclear findings instead of
forcing a verdict. Keep protected non-slogan arguments as a separate category.

## Scope, counting, and checks

Attribute substantive debate speech to A/B, excluded material to O, and ambiguous
speaker passages to U. Exclude advertisements, introductions, housekeeping,
audience/moderator speech, unrelated biography, and edited promotional teasers that
replay later speech. Genuine renewed argument after a reply may count again;
an editor's duplicate clip does not. Include substantive answers to audience questions.
Do not attribute speech just by which ideology it resembles.

Count each distinct claim-use episode once in an uninterrupted argument. Immediate
repetition is one episode. Later renewed use after an intervening reply can be a
new episode. The count is of detected uses in these transcripts, not every belief
held by the speakers. Record confirmed and genuinely uncertain candidates, with
short exact quotations, event boundaries, findings, and concrete explanations.
Also preserve one informative non-slogan contrast per side where available.

Programmatic checks verify full event coverage, allowed labels, source fingerprints,
quotation location, and arithmetic. Targeted editorial checks resolve unclear
passages relevant to the findings and inspect every quotation used in the paper.
Keep the original model output and separately record any source-based correction.
There is no full second-reading or third-adjudicator stage in this streamlined study.
Do not describe the final corpus as independently double-coded.

Primary comparisons: episodes per 10,000 attributed substantive caption words,
first within each debate and then averaged with equal debate weight. Also report
pooled rates, counts, side denominators, and how many debates contain an episode.
Treat criticism-resistant slogans separately from unsupported slogans. Show how
borderline cases and unknown speech affect the interpretation. Check earlier/later
groups, the narrower topic set, and removal of Lennox's debates. Use complete-debate
resampling for uncertainty ranges; explain that this does not capture model bias,
caption errors, or all dependence caused by repeated speakers.

The paper must distinguish a direct transcript review from the previous score-based
risk proxy. It must not infer emotional enforcement of religion, motives, or universal
traits of believers or nonbelievers from these selected debates. Use plain language,
clear graph keys, examples and counterexamples, and premise-to-conclusion summaries.
Embed every PDF font and inspect the rendered result before replacing the site copy.

Expected incremental cash cost remains $0 through existing ChatGPT subscription
access. No paid API, paid transcription, credit purchase, or reset redemption is
part of this plan. Check the shared allowance between bounded batches.