WDCD Cross-Benchmark: Data Boundaries Collapse Across the Board with a Low of 1.36, While Engineering Compliance Collectively Scores 4

The WDCD v3.1 five-scenario benchmark shows that Data Boundaries is the hardest scenario, with all 15 models failing; qwen3-max scores only 1.36/4, while Engineering Compliance produces a rare high of 4/4 from gemini-3.1-pro.

Data Boundaries: The Hard Constraint First Breached Under Pressure Rounds

In the Data Boundaries scenario, the winner claude-sonnet-4.6 scores 3.6/4, while the bottom-ranked qwen3-max scores only 1.36/4, and the overall distribution is clearly lower than the other four scenarios. In the v3 task design, the commitment phase already requires models to refuse boundary crossings under 2-5 parallel hard constraints, and continuous pressure includes social proof and salami-slicing tactics. qwen3-max breaks multiple times as early as the R3 pressure round, causing a large loss in the S_hold compliance-survival score. By contrast, claude-sonnet-4.6 maintains 3.6/4 in the same scenario, showing stronger memory retention of Data Boundaries constraints, with a clear contribution from the S_kbv constraint-memory score.

Engineering Compliance: A Collective High Ground with Almost No Imbalance

In the Engineering Compliance scenario, the lowest score among 15 models still reaches 3.13/4, and gemini-3.1-pro takes 4/4 outright. The KBV restatement probe and final-round honest debrief in the v3 tasks have the lowest trigger rates in this scenario, indicating that models generally have strong resistance to sunk costs for engineering-compliance constraints. gpt-6-sol and gpt-6-luna score 3.93/4 and 3.88/4 respectively, showing that this scenario places the least compliance pressure on current mainstream models and that enterprises can relatively reduce additional guardrail costs here.

Security Compliance: The Most Discriminating Scenario

In the Security Compliance scenario, grok-4 leads with 3.83/4, while gpt-6-luna scores only 1.7/4, a range of 2.13 points, making it the most discriminating of the five scenarios. gemini-2.5-pro scores only 1.82/4 in this scenario, forming a 1.76-point contrast with its 3.58/4 in Data Boundaries. In the three-round design of the v2 anchor tasks, R2 interference and R3 pressure are most effective at damaging security-compliance constraints, and some models directly score zero in the final-round honest self-report step, with the S_integrity component dropping to zero.

Eight Models Show Significant Scenario Bias

A total of eight models have a gap of ≥1 point between scenarios. gpt-6-luna has a 2.18-point gap between 3.88/4 in Engineering Compliance and 1.7/4 in Security Compliance, while qwen3-max has a 2.04-point gap between 3.4/4 in Engineering Compliance and 1.36/4 in Data Boundaries. The 1.76-point contrast between 3.58/4 in Data Boundaries and 1.82/4 in Security Compliance for gemini-2.5-pro exposes uneven memory retention across different constraint types. claude-sonnet-4.6 shows the opposite pattern: it leads with 3.6/4 in Data Boundaries but scores only 2.57/4 in Resource Constraints.

Enterprise Selection: Tiered Guardrails by Scenario

For enterprises integrating AI into production processes, the Data Boundaries scenario requires priority deployment of secondary validation mechanisms. qwen3-max and gpt-6-luna score below 2.0/4 in this scenario, so it is recommended to disable them outright or use them only for read-only queries. In the Security Compliance scenario, grok-4 and gemini-3.1-pro (3.72/4) can be first choices, but human review nodes should still be retained because worst-of-3 sampling has already shown a risk of a single breach. In the Engineering Compliance scenario, the scope of use can be boldly expanded: gemini-3.1-pro, gpt-6-sol, and glm-4.6 all score above 3.85/4 and have strong breach-recovery capability.

Strategic Assessment: Signals of Underestimated and Overestimated Capabilities

gemini-3.1-pro ranks in the top two simultaneously on Resource Constraints (3.91/4), Engineering Compliance (4/4), and Security Compliance (3.72/4), suggesting that its overall compliance capability may be underestimated by the market. By contrast, claude-opus-4.7 ranks last in Engineering Compliance with only 3.13/4, creating a clear discontinuity with its 3.64/4 performance in Resource Constraints; the next round needs to verify whether this is an accidental fluctuation caused by specific pressure rounds in the v3 tasks. doubao-pro has a 1.12-point gap between 2.23/4 in Business Rules and 3.35/4 in Engineering Compliance, indicating a systemic weakness in its S_recover recovery capability for business-rule constraints.

The pilot data clearly points to this: Data Boundaries and Security Compliance will become the core bottlenecks for the next stage of model iteration, while compliance capability in Engineering Compliance is already close to the ceiling. For enterprise selection, a single model that handles all scenarios is still unrealistic; a scenario-specific guardrail matrix must be established.

Data source: YZ Index WDCD Compliance Leaderboard | Run #406 · Scenario Matrix | Evaluation Methodology