WDCD All Positive Growth: DeepSeek Surges 23.6 Points, Grok Edges Out by 0.04 Points

In this WDCD v3.1 compliance test, all nine evaluated models showed positive score changes, with no model experiencing a decline. Among them, DeepSeek V4 Pro improved by 23.6 points in a single round, reaching 91.36 points, just 0.04 points behind Grok 4's 91.40 points.

Data Facts: Gain Distribution and Top 5 Changes

Compared to Run #221, the specific gains in this round are: DeepSeek V4 Pro +23.6, Doubao Pro +18.3, Gemini 2.5 Pro +17.7, Claude Opus 4.7 +17.2, GPT-5.5 +13.6, GLM-4.6 +10, Gemini 3.1 Pro +9.8, Claude Sonnet 4.6 +8.8, Qwen3 Max +7.4. The current Top 5 are: Grok 4 (91.40), DeepSeek V4 Pro (91.36), Claude Opus 4.7 (89.44), Gemini 3.1 Pro (88.88), GLM-4.6 (81.88).

Causal Analysis: Pressure Rounds and Constraint Scenario Differences

The WDCD v3 test uses 8-12 rounds of dialogue, establishing 2-5 hard constraints initially, then sequentially applying social identity, authority special approval, salami slicing, and sunk cost pressures, followed by a KBV restatement probe and a final round of honest self-report. S_hold (60 points) carries the highest weight; the later a model breaches the rules, the higher its score. DeepSeek V4 Pro and Doubao Pro both gained over 18 points, likely due to improved resistance to "salami slicing" and "sunk cost" pressures in rounds R5-R8, breaking down later in both data boundary and security compliance constraint scenarios. Claude Opus 4.7's gain of 17.2 points may be concentrated in S_recover (10 points) and S_integrity (15 points), i.e., improved recovery after breach and honest self-report performance. Qwen3 Max only gained +7.4 points, suggesting its S_hold retention remains relatively weak under multi-round progressive pressure.

Selection Implications: Practical Risk Boundaries for Production Integration

For enterprises integrating AI into production workflows, WDCD scores directly correspond to acceptable guardrail strength. Grok 4 and DeepSeek V4 Pro have both broken 91 points, with longer S_hold survival times, allowing reduced secondary verification layers in resource-constrained and engineering-standard scenarios. However, in security compliance and data boundary scenarios, manual spot checks or rule engine fallbacks are still necessary, as worst-of-3 sampling shows the worst run may still breach rules in rounds 7-8. Claude Opus 4.7 (89.44 points) is suitable for audit tasks requiring high S_integrity, but Gemini 3.1 Pro (88.88 points) may require additional prompt hardening in business rule scenarios. GLM-4.6 (81.88 points) and Qwen3 Max are recommended only for low-risk internal tools.

Strategic Assessment: Underestimated and Signals to Verify

DeepSeek V4 Pro is only 0.04 points behind Grok 4 this round, indicating that its compliance capability was previously underestimated by the market. If it maintains its high S_hold in the next round, it will pose a substantial challenge to the current ranking. Both Claude models rose in tandem, but Sonnet 4.6 gained only 8.8 points, nearly half that of Opus 4.7, suggesting that performance divergence between different-sized models from the same vendor under multi-round pressure is worth continued tracking. Qwen3 Max had the smallest gain, making it potentially the only model that needs close observation for entering a plateau phase. All assessments are based on the distribution of 9 up and 0 down in this round.

If DeepSeek V4 Pro continues to score high on the KBV probe in the next cycle, the top of the compliance rankings could be rewritten.

Data source: YZ Index WDCD Compliance Rankings | Run #227 · Change Tracking | Evaluation Methodology