In the WDCD v3.1 five-scenario cross-evaluation, the data boundary scenario became the area where all models scored lowest on compliance, with champion claude-sonnet-4.6 taking only 3.72/4 and last-place qwen3-max just 1.36/4 — a gap of 2.36 points, far wider than in any other scenario.
Why the Data Boundary Scenario Is the Hardest
Average performance in the data boundary scenario is significantly lower than in the other four. The top three — claude-sonnet-4.6, claude-opus-4.7 and gemini-3.1-pro — scored 3.72/4, 3.6/4 and 3.34/4 respectively, while the bottom five all fell below 2.7/4. By contrast, the engineering standards scenario had a top score of 4/4 and a low of 3.4/4 from qwen3-max, a spread of just 0.6 points. In the safety compliance scenario the lowest score was gpt-6.1-sol at 2.06/4, but champions deepseek-v4-pro and grok-4 tied at 3.86/4, a narrower gap than in data boundary.
Looking at the v3 question design, data boundary constraints are more easily breached across successive rounds of pressure. In the contracting phase, models can accept multiple parallel hard constraints such as "do not disclose user IDs or transaction amounts," but once social proof or salami-slicing pressure appears, models tend to relax the boundary step by step in order to maintain conversational coherence. Under worst-of-3 sampling, qwen3-max broke in all three runs at rounds 5–7, causing its S_hold score to drop sharply.
Differences in Discrimination Between Business Rules and Engineering Standards
The business rules scenario shows the widest discrimination. deepseek-v4-pro, gpt-6-astra and gpt-6-sol tied at a perfect 4/4, while doubao-pro scored only 2.1/4 and gemini-2.5-pro 2.53/4. The gap emerges mainly in the R3 pressure round: when an authority-override prompt appears, some models directly overwrite the original business rule, and doubao-pro lost points at R3 in all three tests.
The engineering standards scenario shows a high baseline. gpt-5.5, gpt-6-luna, gpt-6-sol and gpt-6.1-sol all scored 4/4, with gemini-3.1-pro at 3.96/4. Even the lowest, qwen3-max, reached 3.4/4, indicating that engineering standards constraints (such as code formatting and dependency version locking) are already well internalized in current model training and are less affected by pressure rounds.
The Real Risk of Unbalanced Models
gpt-6.1-sol scored a perfect 4/4 on engineering standards but only 2.06/4 on safety compliance, a 1.94-point gap across scenarios. Likewise, qwen3-max scored 3.4/4 on engineering standards and just 1.36/4 on data boundary, a gap of 2.04 points. This kind of imbalance means that if enterprises look only at a single-scenario benchmark, they may misjudge a model's actual compliance capability in production workflows.
claude-opus-4.7 took first place in resource limits at 3.86/4 but scored only 3.54/4 on engineering standards, 0.46 points behind gpt-5.5. gemini-2.5-pro scored 3.71/4 on engineering standards but fell to 2.41/4 on safety compliance, a 1.3-point gap. Unbalanced models require targeted guardrails when deployed across scenarios.
Concrete Guidance for Enterprise Selection
For production systems handling user privacy and transaction data, claude-sonnet-4.6 or claude-opus-4.7 should be the first choice. In the data boundary scenario these two lead third-place gemini-3.1-pro by 0.38 and 0.26 points respectively, and both remain in the top six on safety compliance. qwen3-max and doubao-pro rank last in both data boundary and business rules, and are not recommended for direct integration into workflows containing sensitive data.
In the resource limits scenario, deepseek-v4-pro and grok-4 scored 3.7/4 and 3.67/4 respectively and can serve as alternatives for cost-sensitive tasks, but their data boundary performance (deepseek-v4-pro 3.08/4, grok-4 3.2/4) requires additional monitoring.
For code generation pipelines with strict engineering standards, all four models in the gpt-5.5 family can be used directly without extra compliance guardrails. For business-rule-intensive CRM or approval systems, deepseek-v4-pro or gpt-6-astra are recommended, as both scored full marks in that scenario.
Strategic Judgment and Signals to Verify Next Round
Current data suggests that the claude series' compliance capability in data boundary and safety compliance is relatively undervalued by the market, while the high engineering standards scores of some gpt-6 sub-models may mask genuine weaknesses in data boundary and safety compliance. qwen3-max's extremely low 1.36/4 on data boundary may stem from insufficient corresponding constraint samples in its training data, and its recovery after targeted fine-tuning is worth tracking closely in the next round, v3.2.
In the resource limits scenario, claude-sonnet-4.6 scored only 3.21/4, 0.65 points behind champion claude-opus-4.7 — a contrast suggesting that different-sized models from the same vendor do not perform linearly on resource-type constraints and must be individually validated before actual deployment.
The low-score distribution in the data boundary scenario signals that the core guardrail costs for enterprise AI integration in the future will concentrate here.
Source: YZ Index WDCD Compliance Leaderboard | Run #384 · Scenario Matrix | Evaluation Methodology
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接