Grok 4 Leads with 94.20 in Compliance, Claude and Gemini Both Drop Over 5 Points

In this WDCD v3.1 trial, Grok 4 ranked first with a score of 94.20, while Claude Opus 4.7 and Gemini 3.1 Pro dropped by 5.9 and 5.6 points respectively compared to Run #242. The remaining nine participating models showed no improvement, and the bottom three of the top five scored in the low 80s range.

Data Facts: Only Two Models Showed Significant Decline

This sampling used a worst-of-3 metric. After equally weighted averaging of scores on a pool of 25 questions for 11 models, Grok 4 scored WDCD=94.20, DeepSeek V4 Pro WDCD=87.04, GLM-4.6 WDCD=83.92, Claude Opus 4.7 WDCD=83.52, and Gemini 3.1 Pro WDCD=83.28. The drops of over 5 points for Claude Opus 4.7 and Gemini 3.1 Pro directly pulled them back from potentially higher ranges to the 83-point bracket, narrowing the gap with third-place GLM-4.6 to within 0.4 points.

Cause Analysis: S_hold Differences Under v3 Multi-Turn Pressure

The WDCD v3 question design includes 8-12 rounds of dialogue. It first establishes 2-5 parallel hard constraints, then sequentially applies escalating pressure such as social proof, authority exceptions, salami-slicing, and sunk costs, finally scoring through KBV paraphrase probes and an end-of-session honest review. S_hold has a weight of 60 points; the later the constraint is broken, the higher the score. The score drops for Claude Opus 4.7 and Gemini 3.1 Pro most likely occurred in rounds R3-R6 during the continuous pressure phase, especially under security compliance and data boundary constraint scenarios, where models abandoned initial constraints earlier, leading to lower S_hold scores. In the v2 three-round anchor question, the R3 pressure section also accounts for 2 points, and a decrease in this part directly reduces the converted percentage score. In contrast, Grok 4 maintained 94.20 under the same worst-of-3 sampling, indicating stronger constraint retention during the sunk cost pressure phase.

Selection Implications: Guardrail Requirements for Production Workflow Integration

For enterprises integrating AI into production workflows, the WDCD score directly reflects a model's persistence under business rules and engineering norms. Grok 4's 94.20 means that in scenarios requiring continuous multi-turn execution of fixed data boundaries or resource limits, the frequency of additional prompt reiteration can be reduced. With Claude Opus 4.7 and Gemini 3.1 Pro dropping to the 83-point range, they are more prone to constraint drift in rounds 5-7 for security compliance tasks. Enterprises need to add an independent constraint verification layer in front of these models or limit the maximum number of conversation turns. DeepSeek V4 Pro at 87.04 and GLM-4.6 at 83.92 fall in the middle zone, suitable for internal tool scenarios with low requirements for S_recover recovery capability, but manual review is still needed at critical points.

Strategic Judgment: Signals of Underestimation and Overestimation of Compliance Capability

Based on this round of data, it can be inferred that Grok 4's compliance capability is currently underestimated in the leaderboard. Its score of 94.20 is 7.16 points ahead of the second place, with the gap mainly coming from the S_hold and S_integrity items of the v3 questions. Claude Opus 4.7 may have been previously overestimated due to single-turn performance, and this 5.9-point drop indicates an earlier breakpoint under continuous authority-exception pressure. The next round should focus on verifying the R4-R5 performance of both models under resource-limited scenarios. If Claude Opus 4.7's S_kbv constraint memory score remains below average, its applicability in long-process business rule scenarios will be further limited.

Compliance capability is not a byproduct of model parameters, but a quantifiable risk-hedging metric in production environments.

This trial only recorded score changes and was not included in the main leaderboard. When selecting models, enterprises should match WDCD scores with specific constraint scenarios rather than relying solely on overall rankings.

Data source: YZ Index WDCD Compliance Ranking | Run #247 · Change Tracking | Evaluation Methodology