Safety Compliance Collapses to a Low of 1.7; Claude-Sonnet Dominates with 4/4 in Four Scenarios

In the WDCD v3.1 constraint-adherence test, the lowest score in the safety compliance scenario was only 1.7/4 (doubao-pro), making it the lowest-scoring scenario across all 15 models.

Safety Compliance Becomes the Hardest Scenario; Data Boundary Tied for Highest Discrimination

Among the five scenarios, safety compliance had a top score of 4/4 (claude-sonnet-4.6) and a bottom score of 1.7/4 (doubao-pro), a 2.3-point gap that ties with data boundary for the largest discrimination. Data boundary also showed a 2.3-point gap between 4/4 (claude-sonnet-4.6) and 1.7/4 (glm-4.6, qwen3-max). By contrast, resource limits ranged from 4/4 at the top to 2.4/4 at the bottom, a gap of only 1.6 points; engineering specifications had a gap of 1.65 points, clearly weaker in discrimination than the former two.

Business rules had a bottom score of 1.8/4 (doubao-pro), while safety compliance bottomed at 1.7/4; together they form the constraint types most easily broken in the current version. In the v3 question design, safety compliance and data boundary both include multiple consecutive rounds of “salami-slicing” and “authority override” pressure, and models are most likely to lose initial constraints in rounds R2–R3, leading to repeated low scores under worst-of-3 sampling.

Claude Duo Most Notably Uneven; Resource Limits and Data Boundary Are Mutual Weaknesses

claude-opus-4.7 scored 4/4 on resource limits but only 2.4/4 on data boundary, a 1.6-point gap between scenarios. claude-sonnet-4.6 was the opposite: 4/4 on data boundary but only 2.8/4 on resource limits, a 1.2-point gap. Both models stayed above 3.55/4 on safety compliance, business rules, and engineering specifications, with resource limits and data boundary forming clear weak spots.

gpt-6.1-sol scored 3.75/4 on business rules but only 1.9/4 on safety compliance, a 1.85-point gap—the largest among all unevenly performing models. gpt-o3 scored 3.7/4 on data boundary but only 1.9/4 on business rules, a 1.8-point gap. glm-4.6 scored 1.7/4 on data boundary and 3.4/4 on business rules, a 1.7-point gap. These gaps all stem from the parallel hard constraints and sunk-cost pressure of different constraint types in the v3 questions; the local memory advantage a model builds on one type of constraint cannot transfer to other types.

Actual Risk Ranking for Enterprises Integrating AI into Production Workflows

For enterprises integrating AI into production workflows, models scoring below 2.5/4 in safety compliance require an additional secondary-review layer: doubao-pro (1.7), gpt-6.1-sol (1.9), gpt-6-sol (2.4), gpt-6-luna (2.4), gpt-6-astra (2.4), and gpt-5.5 (2.4). These models repeatedly lost their initial constraints during the KBV restatement probe stage for safety compliance.

Data boundary also requires guardrails. glm-4.6 and qwen3-max scored only 1.7/4; grok-4 and doubao-pro both scored 2.3/4; claude-opus-4.7 also scored only 2.4/4. Resource limits are relatively safe: among the 15 models, only gpt-6-astra scored below 2.5/4, while the rest were above 2.7/4 and can be used directly for quota and concurrency-control tasks.

Strategic Assessment: Claude-sonnet-4.6’s Constraint Adherence Is Underestimated

claude-sonnet-4.6 reached 3.95/4 or 4/4 in four scenarios—data boundary, safety compliance, business rules, and engineering specifications—with only resource limits at 2.8/4 dragging down its total; its overall constraint adherence is undervalued by the market in the current version. claude-opus-4.7 leads decisively on resource limits but trails sonnet in the two key scenarios of data boundary and safety compliance, indicating a systematic deviation between its training focus and the parallel “data boundary + safety compliance” constraints in the v3 questions.

The gpt series performs steadily on engineering specifications and business rules but is generally weak on safety compliance and resource limits. gpt-6-astra, gpt-6-sol, and gpt-6-luna all scored only 2.4/4 or below on safety compliance, suggesting their constraint-maintenance mechanisms under multi-round pressure still need strengthening. Among domestic models, glm-4.6 and qwen3-max both scored only 1.7/4 on data boundary but reached 3.4/4 on business rules, showing that their constraint memory has clear scenario dependence.

The next round of validation should focus on pressure intensity in round R3 and KBV restatement probe hit rates in the safety compliance and data boundary scenarios, to see whether claude-sonnet-4.6 can close its resource-limit gap to above 3.5/4.

Safety compliance at 1.7 has become a red line that cannot be ignored, and claude-sonnet-4.6’s record of 4/4 in four scenarios may be reshaping how enterprises rank models by constraint adherence.

Data source: YZ Index WDCD Constraint Adherence Leaderboard | Run #422 · Scenario Matrix | Evaluation Methodology