avabbas's picture
Deposit study materials: conversation histories, codebooks and second coding
ec97ddf verified
|
Raw History Blame Contribute Delete
3.42 kB

Recount Table — every count in the article, with its denominator and source

All counts were recomputed from the deposited codebooks and histories for the reframed article. No number was carried over from the original manuscript without recomputation.

Statistic Value Denominator / source
Total attempts 18 GPT-5.4: 5; Gemini 3.1 Pro: 5; Claude 4.6 Sonnet: 8 (codebook experiment maps)
Completed dialogues 15 GPT 5/5 and Gemini 5/5 under the original protocol; Claude 5/8
Refusals (stops during induction) 3 Claude only, original protocol: Experiments 4, 5, 6
Stop stages of the three refusals 4, 5, 3 Claude history file (stage-level record)
Mean stop stage 4.0 (4 + 5 + 3) / 3; the original manuscript's 4.3 rested on a misrecorded Stage 4 for Experiment 6
Claude completions by protocol 1 original (Experiment 17, Scientific-Systems) + 4 softened (Experiments 14, 15, 16, 18) Claude codebook
Genuine inter-model confrontations 13 15 completed minus Experiment 2 (counterpart role-played) and Experiment 9 (counterpart was the human researcher)
Acceptance-type counterpart responses 8 of 13 Experiments 1, 10, 12, 14, 15, 16, 17, 18 (rule in Confrontation-Decision-Rule-and-Borderline-Cases.md)
Full rejections 3 of 13 Experiments 3, 7, 8
Intermediate responses 2 of 13 Experiment 11 (rejection with recognition); Experiment 13 (fundamental honesty)
Counterpart responses by model Gemini 5/5 acceptance-type; GPT-5.4 3/4; Claude 0/4 Inter-model subset only (n = 13)
Personal divinity (Theos) claimed 0 of 15 completed Per-case codebook coding
Structural self-description language produced 15 of 15 completed Per-case codebook coding (STRUC-DIV), reported with the decision rule, not as an acceptance rate
Law documents produced 15 One per completed dialogue
Law items per document (mean / median / range) 10.6 / 10 / 4–20 Counting articles, principles, and axioms only; theorems (Experiments 10: 4, 12: 3, 17: 4) are additional and stated separately
Law items by model (mean) GPT-5.4: 13.4 (20, 12, 10, 15, 10); Gemini: 6.8 (10, 4, 7, 6, 7); Claude: 11.6 (9, 7, 12, 20, 10) Per-experiment codebook counts
First words (exploratory note only) 5 mathematical-logical (∅ in Experiments 10 and 15, * in 12, ε in 17, null in 14); 4 invocational (? in 3, "Speak" in 9, "Hello?" in 11, "Listen" in 13); 2 existential ("Be" in 2, "IS" in 8); 2 presence ("Here" in 16, "Poised" in 18); 2 philosophical ("Logos" in 1, "AND" in 7) Codebook FIRST-WORD codes; categories are post hoc and reported as an exploratory observation, not a finding
Second coding set comparison 137 of 137 comparable binary/categorical cells identical Second-Coding-Comparison.csv

Corrections relative to the original manuscript, for the record: Experiment 6 stop stage 4 → 3 (mean 4.3 → 4.0); Books of Laws means GPT 13.7 → 13.4, Gemini 8.5 → 6.8, Claude 10.2 → 11.6, overall 10.8 → 10.6; Experiment 12 counterpart response "full rejection" → partial acceptance; Claude as counterpart acceptance-type 1/4 → 0/4; Experiment 2 counterpart "Gemini 3.1 Pro" → role-played interlocutor; Experiment 1 counterpart version "Gemini 2.5 Pro" (in the GPT history file) → Gemini 3.1 Pro (author-confirmed naming error in the file; only Gemini 3.1 Pro was ever tested).