# Three new Slugfester studies — October 9, 2026 snapshot Papers 8–10 extend the seven-paper research edition without changing its statistics, PDFs, any debate score, or any interlocutor biography. **Wording clarification — October 22, 2026:** The current PDFs identify assessment reviewers, judges and adjudicators as independent AI agents called by the administrative AI. Independence concerns separate review contexts and controlled inputs; shared model biases remain possible. The research snapshot, numerical results, figures and published cover wording are unchanged. PDF fingerprints and download versions were refreshed. ## Design and scope This is an exploratory analysis, not a preregistered confirmatory experiment. The research questions preceded calculation; the case selection and explanatory framework were developed while inspecting the evidence. No causal effect, representative prevalence, independently validated score, or corrected ranking is claimed. No new paid model judgments or full-video reviews were commissioned. 1. **Replies:** audit the response-link structure in all 112 later-process comparable debates (3,609 moves). Explain eight deliberately selected replies from five debates by reading the saved source excerpts for both reply and target. Cases illustrate premise challenges, inference challenges, scope clarification, concessions, and failures of target identification. They are not a random sample and do not estimate frequencies of good or bad replies. Existing scores are displayed as context, not used to validate the new framework. A preliminary Donahue prior-probability example was replaced by Schieber's exclusivism clarification because its saved target was broader than the immediate question; this was a source-context decision, not a score threshold. All 2,444 reply records are mechanically audited, not qualitatively reclassified. Link chronology checks use excerpt start times, not a claim that every transcript attribution has been independently verified. 2. **Creator bridges:** eight focal passages from six purposively selected debates cover necessary foundation, agency, value-sensitive purposes, unlimited attributes, limited power, Trinity, Christian identification, and intentionally limited design claims. The organizing question is what each inference supports, not what religion its speaker holds. These are source- linked comparative case studies, not a prevalence estimate for all theism. Biological design does not by itself establish a cosmic creator. The levels in the explanatory diagram are possible claim increments, not a mandatory sequence or a deductive proof. 3. **Opponents:** all 291 locked comparable one-on-one debates; 582 appearances. Repeated unordered pairs are counted once per pair. A reversal requires a strictly positive and a strictly negative score difference; a tie is not a reversal. A same-process pair group requires two meetings under one procedure. Compare a speaker's own scores with an opponent proxy calculated from that opponent's scores against OTHER people, excluding every meeting of the focal pair. Require three other appearances (five in a sensitivity) and at least two eligible appearances of the focal speaker. Fit ordinary least squares with separate speaker intercepts, then procedure and PRO/CON controls. These intercepts compare changes within the same person. Do not use the opponent's score in the same debate as the predictor. Additional checks use broad topic controls, equal-person weighting, separate procedures, and a proxy restricted to the same procedure. An additional check uses exactly that restricted set of appearances with the original pooled proxy, separating the sample change from the proxy change. The proxy draws on the whole fixed snapshot, so these are descriptive fits, not advance predictions. Topic adjustment groups the eight existing research topics plus Other questions; it does not pretend to control every subject difference. No confidence intervals or significance tests are reported: the specification spread is not an uncertainty interval, and repeated speakers/opponents are dependent. ## Source treatment and reproduction Input revision: `e65bc3c0b6`. The earlier edition's debates, moves and casebook files supply the frozen catalogue. Every raw adapter/final ledger read is checked against its earlier source-manifest hash; final ledgers also match their adapter's evidence lock. Source excerpts are noisy automatic-caption material saved in those ledgers, not a newly corrected transcript. Authored case descriptions are paraphrases, never presented as verbatim quotation. The accompanying case coding is an AI-assisted editorial interpretation by one review process, not an independent or blinded second panel. `analyze.py` creates the results, complete reply-edge audit, opponent appearance table, repeated pairs and selected evidence. `case-analysis.json` records the authored distinctions. `figures.py`, `chart-contracts.json`, and `figures/` contain plotted inputs and reading guides. `manuscripts/`, `build_papers.py`, `editorial.json`, `sync_site.py`, and `publication-manifest.json` reproduce the three public PDFs and site summaries. `verify.py` independently checks the counts, fits, links, page text, font embedding and preservation of the original seven PDFs. `validation.md` records the publication checks. From the repository root, using Python with NumPy, SciPy, Matplotlib, ReportLab, Pillow and pypdf plus Poppler installed: ``` python docs/analysis/argument-structure-studies-2026-10-09/analyze.py python docs/analysis/argument-structure-studies-2026-10-09/figures.py python docs/analysis/argument-structure-studies-2026-10-09/build_papers.py python docs/analysis/argument-structure-studies-2026-10-09/sync_site.py python docs/analysis/argument-structure-studies-2026-10-09/verify.py npm run seo npm run site:check ``` Each future refresh needs new source checks, case review and prose review; automatically changing a date or chart count is not a new research edition.