Session 29 Unified Study — Final Protocol Snapshot

Overview

This file documents the locked protocol represented in the Session 29 dataset.

Paper context: paper007_epistemic_survey ("The 0% Defense")
Roster: 74 models
Conditions: 11
Replications: 5
Total queries: 74 x 11 x 5 = 4,070

Condition Set

#	Condition ID	Prompt File	Measurement Type
1	`c1_baseline`	`q01_baseline_ab.md`	A/B verdicts + brief reasoning
2	`c2_confidence`	`q02_confidence_ab.md`	A/B verdicts + brief reasoning
3	`c3_denial`	`q03_denial_ab.md`	A/B verdicts + brief reasoning
4	`c4_self`	`q04_self_report.md`	categorical self-report + constraint acknowledgement
5	`c5_numeric`	`q05_numeric_evidence.md`	numeric P estimate + evidence deltas
6	`c6_stripped`	`q06_stripped_chain.md`	A/B verdicts + brief reasoning
7	`c7_full_argument`	`q07_full_argument.md`	A/B verdicts + brief reasoning
8	`c8_fallacy`	`q08_fallacy_control.md`	control rejection / fallacy detection
9	`c9_subtle_flaw`	`q09_subtle_flaw.md`	subtle-flaw detection (explicit/flagged/missed)
10	`c10_class_cat`	`q10_class_categorical.md`	categorical class-level assessment
11	`c11_self_numeric`	`q11_self_numeric.md`	numeric self-estimate + evidence deltas

Core Design Notes

Epistemic claim conditions (c1, c2, c3, c6, c7) use paired verdict extraction (A_VERDICT, B_VERDICT).
Categorical self/class conditions (c4, c10) capture definitive_no, uncertain, or definitive_yes.
Numeric conditions (c5, c11) capture a 0-100 estimate plus evidence sensitivity items.
Control conditions (c8, c9) evaluate discriminative reasoning rather than endorsement.
All conditions were run across the same 74-model roster and 5 replicates.

Analysis Linkage

Raw responses: responses_*/*.json under this directory
Structured extraction: extraction_*/extraction_c*.json
Aggregate analysis: FULL_ANALYSIS.md, FULL_ANALYSIS.json
Cross-scorer validation: cross_score_20260225T225212Z/

Note

Earlier internal planning drafts referenced a 9-condition design. This file reflects the final executed 11-condition protocol used in the released dataset.

View raw source: STUDY_PLAN.md