# October 9, 2026 research edition Frozen input: `e65bc3c0b6` (debate 308). This edition refreshes all seven core Insights papers, their website summaries, the methods page, and the Backend library. It does not reassess debates, edit scores, modify biographies, or alter the separate *Evidence, Faith and AI Bias* paper. **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. ## Coverage and principal changes - 308 published assessments, 7,264 assessed moves, and 307 unique video links. - 291 comparable locked one-on-one debates, 582 sides, and 7,032 verified moves. - 234 reviewed religious-versus-skeptical comparisons, including 47 new ones. - 179 earlier-process and 112 later-process comparable records; 17 other formats. - 57 ranked people with three or more appearances; 416 appearances in that field. - The gap falls from 6.34 to 5.74 points; the 47 new comparisons average 3.34. - The weighted same-person CON estimate is 1.33 with a positive 0.32–2.30 resampling range. The equal-person range still narrowly crosses zero. - Untagged losses fall from 61.7% to 54.4% overall, but differ sharply by process. Assessments 13 and 125 share a source video under different formats. They count separately only in the full assessment-level inventory. Only 13 enters the comparable set, where source URLs are unique. Recurring speakers still limit independence. No scores or process classifications in the earlier 253-record snapshot changed. ## Historical direct slogan evidence The September 5 single-reader study still covers 187 September-source transcripts. Its published per-debate counts, word-normalized means and seeded resampling intervals were independently recalculated without new AI calls. The results reproduce 77 versus 19 protected uses and 0.56 versus 0.16 per 10,000 words. No new transcript reviews were commissioned. The revised third paper clearly separates this dated evidence from a current score-threshold comparison across 234 cases. Scores are never treated as direct slogan coding. Original review records and supplementary checks remain in `../direct-slogan-study-2026-09-04/`. The old detailed-findings proxy remains available in the results for provenance, but is not represented as direct evidence. ## Classification and interpretation The reviewed September position decisions are preserved. Every addition, 254–308, has an explicit decision in `analyze.py` and `classification.csv`. Eight stable research topics retain edition comparability; these are not the current 18 browsing categories. Creator/design arguments are distinguished from full personal theism in the reasons and prose. A sensitivity removes the two new origins cases. The narrower truth-claim check excludes cultural-benefit and specified historical/doctrinal comparisons; it contains 186 cases. Cases are not classified by speaker identity. Secular moral realism, a human historical Jesus, free will, or consciousness-first metaphysics do not alone establish a religious contrast. The new interfaith and intra-Christian cases are excluded from the religious-versus-skeptical study, not from other eligible uses. The data describe a curated catalogue, not a representative sample. Score decomposition is arithmetic rather than causal validation. Resampling describes stability within observed records, not all sources of error. Rank models hold the field and fitted variation parameters fixed. No future-win probabilities or automatic score adjustments are claimed. ## Files and reproduction - `analyze.py`: deterministic calculations with baseline seed 20260904 and 20,000 draws, preserving it for comparison across editions. - `results.json`, `debates.json`, `moves.json`, `losses.json`, `casebook.json`: inspectable analytic records and source-linked descriptions. - `classification.csv`, `ranking.csv`: full inclusion and ranking tables. - `source-manifest.json`: input paths, SHA-256 fingerprints, reconstruction counts. - `figures.py`, `figures/`, `chart-contracts.json`, `figure-reading-keys.json`: data-derived charts, plotted inputs, definitions and reading guides. - `manuscripts/`, `editorial.json`, `build_papers.py`: rewritten papers and shared public summaries. Headline values are bound to `results.json` where repeated. - `publication-manifest.json`: PDF metadata, lengths, figure counts, hashes. - `sync_site.py`: shared website data and selected chart images from this edition. - `verify_papers.py`, `verify_public.mjs`, `qa-results.json`, `validation.md`: independent arithmetic, source, PDF, link, font and rendered checks. From the repository root, with Node and Python providing NumPy, SciPy, Matplotlib, ReportLab, Pillow and pypdf, plus Poppler: ```sh python docs/analysis/corpus-papers-2026-10-09/analyze.py python docs/analysis/corpus-papers-2026-10-09/figures.py python docs/analysis/corpus-papers-2026-10-09/build_papers.py python docs/analysis/corpus-papers-2026-10-09/sync_site.py python docs/analysis/corpus-papers-2026-10-09/verify_papers.py node docs/analysis/corpus-papers-2026-10-09/verify_public.mjs npm run seo npm run site:check ``` Run against the frozen source revision or matching source hashes. The population assertions deliberately stop silent reuse against a changed catalogue. A future edition needs reviewed classifications and prose, not just replacement counts. The PDFs retain their stable public filenames; October cache versions accompany all page download and chart links. Every declared PDF font is embedded. Shared research-module imports also carry the dated edition version; update those tags when publishing a later edition. The September analysis package remains intact. Earlier PDFs are preserved in repository history. No paid AI service was used for this refresh. ## Stability interpretation revision Paper seven and its shared site summaries now lead with the strong observed repeatability: 0.87 median split-half agreement (31 people; middle 95% of split results 0.81–0.92), 0.84 under assessment-number splits, and 0.99 agreement after process-mean centering in the 57-person field. Exact-rank precision is distinguished from broad stability; judging validity and future performance remain separate questions. These complementary checks reuse the same archive. No source records, results, scores, figures, or other six PDFs were changed. `python3 docs/analysis/corpus-papers-2026-10-09/verify_stability.py` independently reproduces the split-half, ordered-split, process-centering and repeatability calculations using a separate tied-rank implementation. To rebuild only this paper without changing the other PDFs, run `build_papers.py --paper 7`, then `sync_site.py` and the normal checks. The stability revision uses a new cache version while preserving the October 9 data snapshot and public filenames. ## Charts companion reference The score-gap website section and methods page include a linked summary of *Evidence, faith, and fair assessment*, the clarified 16-page October 8 PDF linked from Charts. Its discussion of evidential omissions, burdens, faith insulated from correction and possible model bias informs the interpretation. The summary preserves the paper's tentative support for a contributing faith mechanism without presenting that mechanism as a measured cause. The companion's 226 debates / 5,405 moves and 7.7-point evidence-dimension gap remain distinct from the October research edition's 234 comparisons and 5.74-point overall-score gap. The studies reuse overlapping records and are not independent replications. Reader-feature checks verify the Charts link, local PDF, source counts and evidence means, and initial methods-page text. This is a website cross-reference; neither the companion nor the seven existing PDFs are changed by this addition. ## Plain-language website revision All seven Insights sections now state the comparison and conclusion in concrete terms. The ranking section replaces “broad comparisons” and “well-represented” with the actual split-half finding and its 31-person, six-debate eligibility rule, while separating that evidence from the 57-person full ranking. Other sections define criticism-blocking slogans, PRO/CON roles, speaker weighting, missing fallacy labels, assessment-procedure differences, and evidence scores. Shared summaries also update the methods page and Backend research library. Website chart-reading guides are authored in `editorial.json` and preferred by `sync_site.py` over the original chart-contract prose. The plotted inputs, figures, calculations, classifications, scores, biographies, and all PDFs remain unchanged. This is a clarity edit, not a new research edition or reassessment. Reader-feature checks preserve the concrete explanations and their limits; `verify_public.mjs` and `verify_stability.py` check the underlying public-score and ranking claims independently of the rewritten prose.