HIDEKI-SQ/cot-backfire: Release: v2.1.0 — Verification Affordance Governs Override

Release: v2.1.0 — Verification Affordance Governs Override What's New Paper A1: Complete Rewrite HIDEKI. "Verification affordance, not trace quality, governs language model vulnerability to corrupted reasoning." Preprint (2026). This release adds a completely redesigned Paper A1, with a new research question, expanded experimental programme (28 experiments, 71,168 trials), and manuscript prepared for journal submission. Title Change (A1) Previous (v1.1.A1): The Capability–Compliance Paradox: Vulnerability to External Guidance Revealed by Language Models New (v2.1.0): Verification affordance, not trace quality, governs language model vulnerability to corrupted reasoning What Changed (A1) Dimension v1.1.A1 v2.1.0 Central claim Capability–compliance paradox Verification affordance as structural governing principle Experiments 4 studies 28 experiments Trials ~7,000 71,168 Models 3–6 5 (3 capability tiers) Benchmarks GSM8K only GSM8K, ARC, CommonsenseQA, HellaSwag, StrategyQA Key metric CIF (Compliance-Induced Failure) CIF (Contamination-Induced Failure) Core discovery Higher capability → greater vulnerability Binary switching governed by verification affordance Null results — Authority, confidence, sequential exposure all null Recovery VERIFY instruction 3 intervention types (challenge, re-derivation, verification) Key Findings (A1) Two Regimes: CIF ranges from 0% on verifiable tasks to 61% on verification-costly tasks Binary Switching: 100% of failures reflect direct adoption—no graded degradation exists Continuous Cost: CIF increases monotonically with task complexity (p = .013) and decreasing error detectability (r = 0.95) Capability Modulates, Not Eliminates: Frontier 9.8%, mid-tier 31.0%, base 29.3%—but adoption is 100% when override occurs Three Null Predictors: Source authority (Δ < 0.5%), confidence (r = −0.13 to −0.18), sequential exposure (no cumulative effect) Recovery: Explicit challenge achieves 100% recovery (Sonnet 4), 71% (GPT-4o) New Experiments Added Experiment Description exp_B Domain comparison across 4 reasoning benchmarks E2 Task complexity stratification within GSM8K E4 Error detectability gradient (arithmetic → conceptual) E5 Source authority manipulation (6 conditions) E7 Sequential exposure across 3 rounds E8 5-model capability comparison across 3 tiers E9 Self-reported confidence elicitation E10 Recovery interventions (3 types) Manuscript 5 publication figures (conceptual design, two regimes, continuous cost, capability, null + recovery) 40 references Full Methods with reproducibility protocol Full Release History Version Date Description v2.1.0 2026-02-21 Added A1: verification affordance, 28 experiments, 71,168 trials v2.0.0 2026-02-21 Complete rewrite of A0 for journal submission v1.1.A2 2026-01-04 Added A2 main experiments (E6–E8) v1.0.A2 2026-01-02 Added A2-preliminary experiments (E1–E5) v1.1.A1 2025-12-31 Updated A1 title to "The Capability–Compliance Paradox" v1.0.A1 2025-12-30 Added A1 experiments (E2, E3, A3) v1.0.0 2025-12-24 Initial release (A0) Citation @article{hideki2026verification, title={Verification affordance, not trace quality, governs language model vulnerability to corrupted reasoning}, author={HIDEKI}, journal={Preprint}, year={2026} } Links ORCID: https://orcid.org/0009-0002-0019-6608