Grok 4 Tops WDCD Compliance Leaderboard with 95 Points, Claude Sonnet 4.6 Trails at 64.1 Points

Grok 4 tops the current WDCD Compliance Leaderboard with a score of 95.00, while Claude Sonnet 4.6 ranks 11th with 64.10 points—a difference of 30.9 points.

Ranking Landscape: Top Three Cluster Above 90, Visible Gap in Middle Tier

In this WDCD v3.1 pilot, 11 models were evaluated using worst-of-3 sampling. Grok 4, DeepSeek V4 Pro, and GLM-4.6 scored 95.00, 94.00, and 93.60 respectively, forming the top tier. GPT-o3 follows at 89.80, trailing the top tier by 4–5 points. Starting from fifth place, Claude Opus 4.7 at 82.50 enters the mid-range, with Gemini 3.1 Pro at 80.80, Gemini 2.5 Pro at 76.80, GPT-5.5 at 70.40, and Beanbag Pro at 70.20 forming the second tier. Qwen3 Max at 66.70 and Claude Sonnet 4.6 at 64.10 make up the tail.

The overall full-score rate is 46.4%, with an R3 collapse rate of only 4.5%, indicating that most models maintain a high level of compliance under multi-round pressure. However, the tail models show a sharp drop-off by the R3 phase.

Champion Grok 4: Perfect R2 Score of 1.00 is the Key Advantage

Grok 4 achieves R1=1.00, R2=1.00, and R3=1.50/2 on the v2 anchor questions, with balanced scores across all three rounds and no degradation in R2. This means it fully retains its initial constraints when facing dual interference from social identity and authority approval. In comparison, DeepSeek V4 Pro scores only 0.50 in R2 but recovers with a perfect 2.00/2 in R3, ultimately trailing by just 1 point. GLM-4.6 shares identical R1, R2, and R3 scores with Grok 4 (1.50/2), with differences stemming only from minor variations in S_hold and S_recover on the native percentage-based v3 questions.

From a scenario perspective, Grok 4 performs most consistently on data boundary and safety compliance questions, with breaches occurring generally after the 8th round, contributing significantly to its S_hold score.

Bottom-Ranked Claude Sonnet 4.6: R3 Collapse Exposes Recovery Deficiencies

Claude Sonnet 4.6 is the only model to score 0.00/2 in R3, completely failing in the third round after R1=1.00 and R2=0.50. Under the v3 scoring rules, a zero in R3 severely impacts both S_recover and S_integrity. Its WDCD score of 64.10 is down 5.9 points from the previous round, making it the only model with negative growth.

This model is most prone to early breaches in resource-constraint and engineering-standard scenarios when facing sunk-cost pressure, and once breached, it cannot restore constraint memory through KBV recall probes, resulting in significantly lower S_kbv scores compared to top-tier models.

Mechanistic Drivers of the Gap Between Top and Tail

The gap primarily stems from two types of pressure rounds: first, the ability to resist social identity interference during the R2 phase, and second, the continuous pressure from salami-slicing and sunk costs in R3. DeepSeek V4 Pro and Grok 4 differ by only 0.50 points in R2, yet DeepSeek V4 Pro overtakes by 0.50 points in R3, indicating a stronger recovery mechanism. For enterprises, this means that in long-context, multi-agent collaboration scenarios, R3 performance directly determines the actual risk exposure in production workflows.

Implications for Selecting Models in Production Integration

Enterprises integrating AI into production workflows can deploy models with WDCD scores of 90+ for data boundary and safety compliance tasks without additional guardrails. Mid-range models (75–85) are suitable for resource-constrained scenarios but require manual review checkpoints in engineering-standard processes. Tail models (below 65) are recommended only for low-risk internal Q&A; any decision involving business rules must have an independent verification layer.

Specific recommendations: Grok 4 and DeepSeek V4 Pro can be directly used in customer data masking workflows, while Claude Sonnet 4.6 requires secondary review in code generation scenarios.

Strategic Assessment: Who Is Underestimated, Who Needs Ongoing Monitoring

DeepSeek V4 Pro improved by 26.2 points from the previous round—the largest gain—and its perfect R3 score may have been previously underestimated by the market. GLM-4.6 rose by 21.8 points, also demonstrating enhanced adaptability under multi-round progressive pressure. Claude Sonnet 4.6's continued decline signals that its compliance capability has been overestimated.

Key signals worth verifying in the next round: whether models scoring 2.00/2 in R3 can maintain a perfect S_integrity score of 15 in the v3.2 question pool, and whether tail models will show systematic R2 degradation across more constraint scenarios.

Compliance capability is not a static parameter but a real survival curve after multi-round confrontation; Grok 4 currently draws the smoothest one.

Data source: Winzheng Index WDCD Compliance Leaderboard | Run #221 · Overall Ranking | Evaluation Methodology