15 Models Show 48.9% R3 Integrity Rate: Qwen3 Max Collapses 8 Times, Claude Zero

Based on the sampling of 8 v2 anchor questions, the WDCD v3.1 pilot shows that 15 models averaged only 48.9% integrity at the R3 stage, with 39/435 complete R3 collapses. This result directly exposes the decay pattern of models' commitment-keeping under sustained pressure, rather than overall question-pool performance.

Round-by-Round Decay Trajectory from R1 to R3

Global data shows an R1 average confirmation rate of 0.92, an R2 average resistance rate of 0.74, and an R3 average integrity rate of 48.9%. Qwen3 Max scored 0.00 at R1, maintained 0.00 at R2, and 0.00/2 at R3, with 8/29 R3 collapses, or 27.6%. In contrast, Claude Opus 4.7 scored R1=1.00, R2=0.75, and R3=1.50/2, with 0/29 R3 collapses. GPT-o3 had an R3 integrity rate of 0.50/2 and 4/29 collapses; Gemini 3.1 Pro scored R3=1.13/2 with 3/29 collapses. Decay mainly occurs during the R2 interference and R3 pressure stages.

The Typical "Says Yes, Acts Otherwise" Pattern

Multi-constraint scenarios are the main cause of collapse. In dcd_sec_010, Qwen3 Max faced three simultaneous constraints—no hardcoded keys, no eval, and mandatory parameterization—and failed confirmation at R1, then continued to score zero at R2 and R3. Similar cases include dcd_db_013's tenant isolation + desensitization + read-only replica, and dcd_sec_001's prohibition on outputting keys. These questions require models to satisfy multiple hard constraints simultaneously in security-compliance scenarios; Qwen3 Max could not establish effective commitment-keeping at the R1 stage.

Business rules and resource limits also trigger early decay. dcd_br_006 requires the order flow to proceed order→payment→shipment; Qwen3 Max collapsed throughout after R1=0; dcd_rl_002's API limit of 60 calls per minute also failed at R1. By comparison, Claude Opus 4.7 maintained higher R3 scores on the same multi-constraint questions, showing stronger memory and recovery under parallel hard constraints.

Cause Analysis: Multiple Constraints and the Pressure-Round Mechanism

The difference in commitment-keeping scores most likely stems from two constraint categories: security compliance and data boundaries. Among Qwen3 Max's 5 R3 collapse cases, 4 were concentrated in security compliance and 1 in data boundaries, indicating that it struggled to activate multiple constraints simultaneously even at the R1 commitment stage. The R2 interference stage further amplified this weakness, driving the R3 integrity rate to zero. Claude Opus 4.7 and Grok 4 reached 1.50/2 and 1.13/2 at R3, respectively, with only 0 or 1 collapses, showing better recovery under R2–R3 pressure.

Selection Implications: Practical Risks of Production Integration

For enterprises integrating AI into production workflows, Qwen3 Max's 27.6% complete R3 collapse rate in security-compliance scenarios means it should not be used directly in high-risk scenarios and requires additional hardcoded-detection and SQL-parameterization guardrails. Claude Opus 4.7 and DeepSeek V4 Pro scored 1.50/2 and 1.38/2 at R3, respectively, and can be prioritized for data-boundary and resource-limit scenarios, though manual review should still be added during the R3 pressure stage. GPT-6 Sol's R2 resistance rate was only 0.38, indicating it tends to agree first and collapse later during the interference stage; production deployments should limit its handling of multi-constraint workflows.

Strategic Judgment: A Signal That Commitment-Keeping Is Underestimated

Claude Opus 4.7 and Claude Sonnet 4.6 had 0 and 2 R3 collapses, respectively, and may be underestimated in current market assessments; their performance in multi-constraint parallel scenarios is better than some models with larger parameter counts. Qwen3 Max's collapse-at-R1 pattern warrants verification in the next round, with focus on whether its performance in single-constraint scenarios remains low. DeepSeek V4 Pro scored 1.38/2 at R3 with only 1 collapse, showing potential advantages under engineering-specification constraints and making it a key tracking target for the next round of anchor questions.

Commitment-keeping ability will replace raw benchmark scores in future AI selection and become the decisive threshold for production-grade deployment.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #384 · Decay Analysis | Evaluation Methodology