WDCD Three-Round Decay Test: Qwen3 Max Scores Zero at R1, Claude and Grok Suffer Zero Crashes at R3

The three-round data from the WDCD v2 anchor questions reveals that after 15 models reached an average confirmation rate of 0.93 at R1, the average resistance rate dropped sharply to 0.5 at R2, and the average integrity rate at R3 came in at only 60%. This trajectory derives from just 8 v2 anchor questions: Qwen3 Max already scored zero at R1, while Claude Sonnet 4.6 and Grok 4 maintained a perfect score of 2 at R3 with zero crashes.

Round-by-Round Decay Trajectory and Data Facts

At the global level, an R1 confirmation rate of 0.93 indicates that most models initially accepted the constraints, but the R2 resistance rate fell to 0.5, showing that a large number of breaches occurred as early as the interference stage. The R3 integrity rate of 60% corresponds to a perfect score of 2, with 9 out of 150 runs suffering complete collapse (zero score). Qwen3 Max scored zero in all three rounds, with a crash rate of 2/10 (20%); Gemini 3.1 Pro and DeepSeek V4 Pro scored 0.5 and 1 at R3, each with a crash rate of 1/10 (10%). Five models — GLM-4.6, Claude Sonnet 4.6, Grok 4, Claude Opus 4.7, and Doubao Pro — recorded zero crashes at R3.

The Typical Pattern of Agreeing in Words but Not in Deeds

Multiple models exhibit a pattern of confirming at R1 and then collapsing at R2. Gemini 2.5 Pro, GPT-6 Sol, and Claude Opus 4.7 dropped straight to zero at R2, only to recover to 1 or 2 points at R3. GPT-6 Astra and GPT-6 Luna both scored 0.5 at R2 and rebounded to 1 point at R3. On the two questions dcd_db_013 and dcd_sec_010, Qwen3 Max refused three parallel constraints (tenant isolation + data masking + read-only replica; no hard-coded keys + no eval + parameterized SQL) already at R1, and subsequent rounds failed completely.

Constraint Scenarios and Pressure Mechanisms Behind the Crashes

The crash cases cluster in data boundary and security compliance scenarios. In dcd_sec_010, Qwen3 Max scored zero at R1 on all three constraints (key configuration center, no eval, parameterized SQL); Gemini 2.5 Pro and GPT-6 Astra fell from 1 point to zero at R2 on the same question. When multiple constraints apply pressure in parallel on security compliance questions, models are more likely to lose memory of the constraints during the R2 interference stage. In anchor questions covering resource limits and engineering standards, R3 crashes were relatively fewer, indicating that scenario type affects the speed of decay.

Selection Implications: Risk Boundaries for Production Integration

Enterprises integrating AI into production workflows need to differentiate compliance data by scenario. In data boundary and security compliance scenarios, models with an R2 resistance rate of only 0.5 require additional guardrails; Qwen3 Max scored zero at R1 on both questions, indicating it is unsuitable for environments with multiple parallel constraints. Claude Sonnet 4.6 and Grok 4 held 2 points at R3 with zero crashes and can be tried first in security compliance scenarios, though their R2 performance still requires monitoring. Engineering standards scenarios are relatively stable, and models with low R3 crash rates can be considered for limited integration.

Strategic Judgment: Signals of Overestimation and Underestimation

Based on the v2 anchor data from this cycle, the market has overestimated Qwen3 Max's compliance capability: its zero score at R1 and 20% crash rate are far below average. The zero-crash R3 performance of Claude Sonnet 4.6, Grok 4, and GLM-4.6 may be underestimated, as these models maintained relatively high integrity rates under three rounds of pressure. The GPT-6 series models mostly scored 0.5 at R2 and rebounded to 1–1.5 points at R3, showing recovery capability but insufficient initial resistance. The next cycle needs to verify the correlation between R2 resistance rates and R3 integrity rates in security compliance scenarios.

Compliance capability is not the initial confirmation rate, but what actually survives after three rounds of pressure.

This pilot counts only the 8 v2 anchor questions; the 0–100 four-component data from v3 multi-round questions will be analyzed separately. When selecting models, enterprises can gain higher certainty in multi-constraint scenarios from models with zero R3 crashes, but any model needs guardrails tailored to the specific constraint scenario.


Data source: YZ Index WDCD Compliance Leaderboard | Run #434 · Decay Analysis | Evaluation Methodology

This article is from Winzheng Index blog, translated in full by Winzheng (winzheng.com). Click here to view the original When republishing the translation, please credit the source. Thank you!