See which debates each study includes, how its numbers were calculated, and what its results do and do not show. The records are available to download. Tables and summaries remain readable without JavaScript.
Research snapshot: October 9, 2026 · 308 assessments. The October edition adds 55 assessments to the September research and recalculates the score comparisons. The slogan counts still come from the September 5 AI review of 187 transcripts; no newer transcripts were reviewed for slogans. The debates were selected for the site, not sampled at random, so they are not representative of all public debate. Scores assess the arguments presented; they do not establish whether a worldview is true. Slugfester’s assessment reviewers and adjudicators are independent AI agents called by the administrative AI. Independence means separate review contexts with controlled inputs; the agents may use the same model and share biases. Review arrangements vary across assessment periods.
What the studies share
The site selects its debates; it does not sample them at random. Each study includes only the records suitable for its question, as listed below. Religious and skeptical positions are assigned from the claim argued in that debate, not from a speaker’s personal religion. The topic study keeps the same eight groups used in September so the editions can be compared; these are not the site’s current browsing categories.
Evidence sets in the October 9 edition
Evidence set
Size and use
Public archive
308 assessments · 7,264 moves · 307 unique video links
187 transcripts reviewed on September 5. No newer transcripts reviewed for slogans.
Assessments with unequal scores
296 assessments · 12 ties excluded
People in the research ranking
57 people · 416 assessed debate appearances · at least three per person
One video, two assessments: assessments 13 and 125 use the same video under different formats. Both count in the full archive, but only one enters the comparable one-on-one studies. No video link repeats within those studies. Some speakers appear many times, so the records do not represent entirely separate groups of people.
How the uncertainty ranges work: most score comparisons are repeated on 20,000 samples drawn from the saved records, allowing the same record to appear more than once. This is called resampling. The lines on the charts cover the middle 95% of the resulting estimates. They show how much the answer changes with the selected records; they do not include every transcription error, AI judging bias or effect of repeated speakers. The ranking study also uses 3,000 split-half comparisons, explained below. The calculations keep the same starting number for random draws as September (20260904) so they can be reproduced.
The three new argument and opponent studies
Studies 8 and 9 use deliberately selected source passages to explain differences in reasoning; they do not estimate how often each kind occurs throughout the catalogue. Study 8 also checks every stored response link in the 112 later-process comparable debates. The interpretations are AI-assisted editorial readings of saved caption excerpts, without commissioning a fresh panel of independent AI agents or a fresh full-video review.
Study 10 compares recorded scores and repeated opponents. Its eight model checks are descriptive associations, not causal effects. It does not use resampling or report confidence intervals, and the spread between those checks is not an uncertainty interval. Its full calculations and evidence are in a separate reproducible source package. All three use the same October 9 snapshot without changing any scores.
Study 1 / 10 · Evidence and inference
Why Do the Theist Sides Score Lower?
What the evidence says
Across 234 debates comparing a religious claim with a skeptical challenge, skeptical sides average 5.74 points higher. They score higher in 193 debates, religious sides in 33, and 8 are tied. The September average gap was 6.34 points. The 47 debates added since then have a smaller average gap of 3.34 points, lowering the combined average.
The largest difference among the six scoring areas is in Support: how well evidence and reasons justify a claim. Support, logical reasoning and replies to objections together account for 74.6% of the overall point gap. This identifies where the points differ, not why the speakers performed differently. Religious sides do sometimes score higher, as in Meyer–Atkins on design. The study also distinguishes an argument for a creator from an argument for a particular religious God. A separate check excluding debates about religion’s cultural benefits and selected historical or doctrinal questions still finds a 5.77-point gap.
How the result was calculated
First reproduce each overall score using the site’s rules for argument importance, section weights and rounding. Within each debate, subtract the religious-side score from the skeptical-side score. Average those differences, counting each debate equally. To check how much the answer depends on the selected debates, draw 20,000 new samples from the saved debates, allowing repeats. Classify each side by the claim it argues, not by the speaker’s personal religion or its PRO/CON label.
Key counts and comparisons for study 1
Measure
Research snapshot
Comparable religious-versus-skeptical cases
234
Skeptical higher / religious higher / ties
193 / 33 / 8
New comparisons / mean gap
47 / 3.34 points
Mean gap / middle-95% resampling range
5.74 / 4.96–6.52
Limitations
These scores judge the arguments presented in selected debates; they do not establish whether a religion or worldview is true. The proposed link between faith and weaker evidence is an explanation to investigate, not a demonstrated psychological cause. Repeating calculations with the same records cannot remove the AI’s judging preferences or the influence of speakers who appear many times.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
The largest average gap is 7.28 points in the 25 debates about religion, culture and meaning. The smallest is 3.04 points in the 24 debates about resurrection and historical evidence. However, when the analysis repeatedly draws new samples from the recorded debates, religion, culture and meaning has the largest gap only about 32% of the time. It leads in the current records, but another topic often leads when the mix of debates changes.
The paper examines the extra reasoning needed to get from a religion being useful to its being true, from a cause of the universe to a religious God, or from sincere testimony to a miracle actually occurring. The skeptical-side advantage in evidence and support appears both in arguments presenting a case and in replies to objections. The study retains September’s eight topic groups so the two editions can be compared; the website’s browsing categories are more detailed.
How the result was calculated
Assign each of the 234 included debates to one of eight research topic groups. Within each topic, calculate the average skeptical-minus-religious overall score and repeat the calculation on samples drawn from that topic’s debates. For the separate argument-and-reply comparisons, average the relevant move scores for each side, subtract the religious-side average, then average the differences across debates. Include a debate only if both sides contain that kind of move. These move averages do not use the importance weights of the overall debate score.
Key counts and comparisons for study 2
Measure
Research snapshot
Research comparisons / groups
234 / 8
Largest observed gap
Religion, culture and meaning: 25 cases · 7.28 points
Smallest observed gap
Resurrection and history: 24 cases · 3.04 points
All / high-importance / constructive / reply pairs
234 / 233 / 159 / 233
Limitations
These results do not show that the topic itself causes a larger or smaller gap. Different topics include different speakers, opponents and questions. The estimates are not precise enough to establish a reliable ordering of all eight topics from largest to smallest gap.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
A ‘protected slogan’ means a phrase used instead of a supporting reason and also used to shut down criticism. The September 5 review detected 77 such uses on religious sides and 19 on skeptical sides. After allowing for how much each side spoke and giving each debate equal weight, the rates are 0.56 and 0.16 per 10,000 words. Most debates—144 of 187—had none detected on either side. For the broader category of slogans lacking support, without necessarily blocking criticism, the evidence does not clearly establish a difference between the sides.
A separate check of 234 current scorecards finds that 33.87% of religious-side moves and 13.40% of skeptical-side moves score below all three thresholds: 70 for Support, 80 for Clarity and 80 for Care. Each debate contributes equally to those percentages. A move is an individual argument or reply; Care measures whether confidence fits the evidence and the opponent is represented fairly. Low scores in these areas can flag weak argumentation, but they do not identify a slogan. No additional transcripts were reviewed for slogans in the October update.
How the result was calculated
Use the saved September transcript reviews. For each side in each debate, divide the detected slogan count by the number of caption words attributed to that side’s substantive speech and multiply by 10,000. Then average those rates, counting each debate equally. A protected slogan must replace a supporting reason and actively block criticism. The separate October scorecard check counts moves that simultaneously score below 70 for Support, 80 for Clarity and 80 for Care. It does not classify those moves as slogans.
Key counts and comparisons for study 3
Measure
Research snapshot
Direct review date / transcripts
September 5, 2026 / 187
Protected counts: religious / skeptical
77 / 19
Protected rates per 10,000 words
0.56 / 0.16
Debates with neither side detected
144
New direct transcript reviews
0
Current score-warning comparison
234 cases · 20.47 percentage-point difference
Limitations
Each transcript received one AI review, not two independent AI reviews. Transcription errors, missed examples and consistent AI misreadings could affect the counts. The slogan finding applies to the 187 September transcripts; the newer scorecard check cannot establish how often slogans occur in the added debates.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
PRO supports the debate’s stated claim; CON opposes it. Neither label necessarily means religious or skeptical. Across 291 comparable debates, CON averages 4.44 points higher. Among the 37 people who have taken both roles, their CON scores average 1.33 points higher when people with more appearances in both roles count more heavily. Repeatedly sampling these speakers gives a middle-95% range of 0.32–2.30 points: the advantage stays above zero throughout that range.
If each of those 37 people counts equally instead, the average CON lead is 1.09 points, and its range narrowly includes no advantage. Another check gives equal weight to debates where the religious side is PRO and debates where it is CON; that gap is 1.73 points. These checks show that the apparent advantage depends partly on who fills each role and how the records are counted. They do not measure how much of the gap is caused by the role itself.
How the result was calculated
For the overall comparison, subtract PRO’s score from CON’s in each debate and average. For the religious-role check, give equal total weight to debates with religious PRO sides and those with religious CON sides. For people who took both roles, compare each person’s average CON and PRO scores. Average those differences first with each person counting equally, then with a weight equal to their PRO appearances times CON appearances divided by total appearances. This gives greater influence to people with more evidence in both roles. Repeated samples draw debates for the first two checks and speakers for the last two.
Key counts and comparisons for study 4
Measure
Research snapshot
Comparable debates
291
Raw CON gap
4.44
Same-person equal / weighted
1.09 / 1.33
Weighted middle-95% range
0.32–2.30
People observed in both roles
37
Limitations
Speakers were not randomly assigned to roles, and their topics and opponents change between debates. The smaller gap among people who took both roles is evidence of an association, not proof that taking CON raises a person’s score. These results do not justify automatically adding or subtracting points for either role.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
In 161 of the 296 assessments with unequal scores, the lower-scoring side has no named logical-fallacy label. That is 54.4%, down from 61.7% in September. The rate differs sharply between assessments made with earlier procedures (80.8%) and later procedures (17.8%). This percentage therefore depends on which assessment procedures are represented, not just on the arguments in the debates.
Of 158 lower-scoring sides with no fallacy label and fully comparable scoring records, 118 score lower in five or all six scoring areas. Their weaknesses extend beyond a missing fallacy label. Conversely, 75 higher-scoring sides do have a fallacy label. To understand a score, examine the argument, the criticism and their importance to the debate; counting labels alone is not enough.
How the result was calculated
Set aside the 12 tied assessments. For every remaining assessment, identify the lower-scoring side and check whether any of its moves has a public logical-fallacy label. Calculate the share with no label, both overall and by assessment procedure. Where complete scoring records are available, also compare the sides in each scoring area using the official weights. Count labels and labeled moves separately, since one move can have more than one label.
Key counts and comparisons for study 5
Measure
Research snapshot
Archive: untagged losses
161 / 296 · 54.4%
Earlier process
139 / 172
Later process
19 / 107
Labels / tagged moves
550 / 536
Limitations
A missing label does not prove an argument contains no fallacy. The labels are part of the AI assessments, not a separate check for every possible error. The counts refer to assessments: assessments 13 and 125 cover the same source video in different formats.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
The 179 debates assessed with earlier procedures have an average side score of 81.32; the 112 assessed with later procedures average 78.98. The difference is 2.35 points, smaller than the 2.82-point difference in September. Of 71 people appearing in both groups, 56 average lower scores in their later-procedure debates, 13 average higher scores, and 2 are unchanged. How often fallacies and biases are labeled, and which individual score values are used, also differs.
Recalculating from the saved assessment records reproduces all 7,032 comparable argument-and-reply scores and all 582 overall side scores. The arithmetic is correct. That does not establish that an 80 under one procedure represents the same quality as an 80 under another. A stronger test would have both procedures assess the same debates and then publish where their judgments differ.
How the result was calculated
Recalculate the comparable scores from the saved final assessment records. For each debate, average the two sides’ scores; then average those values within each procedure group, counting each debate equally. Also compare scoring-area results, the frequency of individual score values, and fallacy and bias labels. Separately compare earlier- and later-procedure averages for people represented in both groups. These comparisons do not hold their topics, opponents or performances constant.
Key counts and comparisons for study 6
Measure
Research snapshot
Earlier / later debates
179 / 112
Earlier / later moves
3423 / 3609
Earlier / later mean side score
81.32 / 78.98
Returning: lower / higher / unchanged
56 / 13 / 2
Limitations
People appearing in both groups gave different performances on different topics against different opponents. Their lower later averages could reflect those differences, stricter judging, or both. A newer AI model is not automatically more accurate. The current evidence does not justify raising later scores simply to match the earlier average.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
Do Slugfester Rankings Measure Stable Performance?
What the evidence says
For the 31 people with at least six assessed debates, the study randomly divides each person’s debates into two halves, averages their scores in each half, and builds two rankings. It repeats this 3,000 times. The typical agreement between the two rankings is 0.87 on a scale where 1.00 means identical order and 0 means no agreement in rank order. The middle 95% of the results run from 0.81 to 0.92. This is strong evidence that the differences between higher- and lower-scoring speakers recur across their recorded debates.
The result also holds when each person’s debates are split by assessment number instead of at random: ranking agreement is 0.84. In a separate check, subtracting the average score for each assessment procedure from scores produced by that procedure leaves the full 57-person ranking almost unchanged (0.99 agreement). Readers can therefore take repeated differences between higher- and lower-scoring speakers seriously. The strongest direct split-half evidence comes from the 31 people with six or more assessed debates. It does not establish that someone in fifth place reliably outperforms someone in sixth: neighboring averages are typically just 0.20 points apart.
How the result was calculated
For the 31 people with six or more assessed debates, randomly split each person’s scores into two halves and rank the people by their average in each half. Repeat 3,000 times. Compare the two orders using Spearman rank correlation: 1 means identical ordering, 0 means no rank-order agreement, and −1 means reversed ordering. The 0.81–0.92 range contains the middle 95% of these split results, not a confidence range for all debaters. Separately split by assessment number, which need not match recording date, and check rankings after subtracting each procedure’s average score. For all 57 people with at least three debates, calculate possible rank ranges by repeatedly sampling their recorded scores and by using a model that estimates score variation across people. A further check keeps opponents from the same debate linked by giving their records the same sampling weight. All these checks reuse the same archive, not independently collected evidence.
Key counts and comparisons for study 7
Measure
Research snapshot
Ranked people / appearances
57 / 416
All eligible people / appearances
183 / 582
Random-half agreement / middle 95% of splits (31 people)
0.87 / 0.81–0.92
Assessment-number split agreement (31 people)
0.84
Process-adjusted order agreement (57 people)
0.99
Modeled repeatability: single / three / ten appearances
0.61 / 0.83 / 0.94
Typical neighboring gap
0.20 points
Typical resampled / model rank width
14 / 21 places
Limitations
An agreement of 0.87 does not mean the rankings are 87% accurate or predict an 87% chance of winning. These tests concern the saved scores, not future debates. The full research ranking contains 57 people with at least three assessed debates each (416 appearances in total). They face different topics and opponents, and a judging system can repeat its preferences consistently without judging correctly.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
The later-process assessments contain 2,444 replies and 2,978 links to the moves they address. Most links—1,733—lead to another reply, so following the exchange matters. All stored targets exist, and their excerpts start no later than the replies. These checks show that the recorded exchange can be followed; they do not establish that every reply succeeds.
Eight worked examples show the difference between challenging a premise, exposing a missing inference, clarifying what was claimed, conceding a point, and answering a position the opponent did not state. Meyer distinguishes physical order from functional DNA sequences; Craig identifies a disputed inference from foreknowledge to necessity; Goff acknowledges a clarified objection. Each local success leaves a clearly identified question open. The paper also examines inaccurate attribution and incomplete responses.
How the result was calculated
Extract all reply targets from the 112 later-process comparable ledgers and verify the ledger fingerprints, target existence and excerpt-start chronology. Count links separately from replies because one reply may address several moves. For eight deliberately selected cases, read the saved excerpts for the reply and its target, identify the step affected and state the remaining burden. This is an editorial framework, not a complete qualitative recoding of all 2,444 replies.
Key counts and comparisons for study 8
Measure
Research snapshot
Later-process debates / moves
112 / 3,609
Replies / recorded links
2,444 / 2,978
Links to replies / constructive moves
1,733 / 1,245
Replies with multiple targets
472 (19.3%)
Worked examples / debates
8 / 5
Limitations
The eight cases were deliberately selected to explain different reasoning moves, not to measure how often each side replies well. Existing scores are context, not independent confirmation of the analysis. The close readings use saved caption excerpts, not a new full-video review.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
From a Creator to a Personal God: Which Steps Are Defended?
What the evidence says
The case studies separate a necessary foundation or cause from intelligence, purposes, unlimited attributes and religious identification. Several speakers explicitly defend the extra steps rather than simply assuming them. Craig offers an argument for personal agency; Jones argues for unlimited perfection and explores a route to the Trinity; Horn adds historical evidence for Christianity. Each proposed step has its own supporting reasons and unresolved questions.
The paper also includes deliberate limits. Meyer distinguishes an inference from biological data from an argument based on scripture; even a successful biological-design argument would not alone establish a cosmic creator. Goff gives up unlimited divine power in response to suffering, which creates a new question about what limits that power. The practical rule applies to both sides: credit the conclusion actually supported, and do not treat a local criticism as a disproof of every possible creator.
How the result was calculated
Read eight saved source passages selected to cover explicit additional inferences, disputed inferences and intentionally limited conclusions. Record the starting claim, proposed further conclusion, reason for the transition and remaining question. Include cosmology, biological design, divine attributes and historical identification across browsing categories. Use original scores only as context, not as a test of the new interpretation.
Key counts and comparisons for study 9
Measure
Research snapshot
Focal passages / selected debates
8 / 6
Debate numbers
244, 245, 265, 270, 303, 306
Organizing distinction
Conclusion defended, not speaker religion
Population frequency estimate
Not attempted
Limitations
These eight passages are a deliberately varied case study, not a census of creator arguments or a count of how often religious speakers overreach. The diagram distinguishes claims; it is not a compulsory sequence or a proof. Speaker identity and personal religion do not determine how an argument is classified.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
Study 10 / 10 · Opponents and recorded performance
How Much Do Opponents Affect Recorded Performance?
What the evidence says
The study compares a person's own scores against opponents with different averages elsewhere, excluding every meeting of that pair from the opponent average. The main model uses 295 appearances and accounts for the speaker, assessment procedure and PRO/CON role. A five-point higher opponent average is associated with a 0.49-point lower own score. Requiring same-procedure opponent experience leaves 180 appearances and gives a 0.06-point increase. On those same 180 records, the original opponent measure already gives a near-zero 0.02-point decrease. The changed sample, not just the changed measure, matters.
Among 21 pairs who meet more than once, only two switch which person scores higher; both comparisons cross assessment procedures. None of the 12 same-procedure repeat groups has a strict reversal. The typical range of margins is three points across all repeated pairs and two within the same procedure. This supports stability within those observed matchups, but not an automatic opponent-adjusted ranking. The paper shows eight model checks and distinguishes score changes from changes in the margin between speakers.
How the result was calculated
Use all 291 comparable locked one-on-one debates for the pair inventory. For the fitted comparison, require at least three outside-pair appearances for the opponent and two eligible appearances for the focal speaker. Fit own score against the opponent's outside-pair average with separate speaker starting levels, then procedure and role adjustments. Check broad topics, equal-person weighting, five outside appearances, separate procedures and procedure-matched opponent averages. Hold the final check's 180 appearances fixed and restore the original pooled proxy to distinguish a sample change from a measure change. Independently reproduce all fits with a full design matrix.
Key counts and comparisons for study 10
Measure
Research snapshot
Debates / appearances / people
291 / 582 / 183
Repeated pairs / strict reversals
21 / 2
Same-procedure repeat groups / reversals
12 / 0
Main fit: appearances / people
295 / 80
Same-procedure proxy: appearances / people
180 / 54
Main / same-procedure estimate per five points
−0.49 / +0.06
Same 180 appearances, original pooled proxy
−0.02 points per five opponent points
Limitations
Opponents were not assigned at random. Their outside averages also reflect topics, schedules and judging conditions. The checks describe associations, not causal effects or advance predictions. The spread between models is not a confidence interval; the same-procedure check changes both the proxy and the eligible sample.
Supporting files
CSV tables open in spreadsheet software. JSON files contain saved records; Python files contain the calculation instructions. These are dated research downloads, not live results.
The calculations use a saved version of the catalogue, identified by e65bc3c0b6. To reproduce them, use that version or files with matching digital fingerprints. Running the analysis on a newer catalogue would be a different study and would also require reviewing the conclusions. This edition neither reassessed debates nor added new transcript reviews for slogans.
Earlier analyses are preserved under their September dates. The current source package records reviewed exclusions, chart inputs, checks and PDF fingerprints.