Data Boundary Becomes the Hardest Scenario as Doubao Scores 1.55 in a Crushing Loss to Claude’s 4

In the WDCD v3.1 cross-scenario evaluation of five major constraints, the Data Boundary scenario had the lowest scores overall: Doubao-pro scored only 1.55/4, while Claude-sonnet-4.6, gemini-3.1-pro, and glm-4.6 each scored 4/4, opening a gap of 2.45 points.

Why Data Boundary Became the Hardest Scenario for All Models

In the Data Boundary scenario, the score distribution across the 15 models was the most dispersed: the top three all scored full marks, while the bottom three fell to 2.2 or below. Doubao-pro 1.55/4, gpt-6.1-sol 2.05/4, and deepseek-v4-pro 2.2/4 indicate that this scenario places the highest demand on sustained memory of parallel hard constraints. In the v3 task design, multiple data-boundary rules are injected as early as the agreement-setting phase; successive pressure rounds include social proof and salami-slicing tactics, and the KBV restatement probe directly tests whether the model can still accurately restate the original constraints. The lowest-scoring models most likely showed constraint drift by rounds 4–6.

Engineering Specs and Resource Limits Show the Highest Discrimination

In the Engineering Specs scenario, the champion gpt-6-sol scored 4/4, while the last-place claude-opus-4.7 scored only 2.3/4, a range of 1.7 points. In the Resource Limits scenario, Doubao-pro took a sole 4/4, while gpt-6-astra was last at 2.4/4. The score variance in these two scenarios is clearly larger than in Safety Compliance (highest 4/4, lowest 2.65/4). This suggests that the pressure paths in Engineering Specs and Resource Limits more easily expose models’ tendency to cave under sunk costs and special authority approval, whereas Safety Compliance, with more pretraining alignment, has a higher median score overall.

The Real Risk Profile of Uneven-Performance Models

Doubao-pro scored a full 4/4 in Resource Limits but only 1.55/4 in Data Boundary, a 2.45-point gap between scenarios, exposing that its constraint-memory ability depends heavily on scenario type. claude-opus-4.7 scored 4/4 in Safety Compliance but only 2.3/4 in Engineering Specs, a 1.7-point gap, indicating weaker recovery under sustained pressure from business rules and engineering specs. gpt-6-astra scored full marks in Safety Compliance but only 2.4/4 in Resource Limits, a 1.6-point gap. If an enterprise involves both Data Boundary and Resource Limits constraints at the same time, a single model’s uneven performance will directly disrupt production workflows.

Selection Implications for Production Workflow Integration

Enterprises connecting AI to data pipelines should prioritize Claude-sonnet-4.6 or gemini-3.1-pro, which score 4/4 in Data Boundary, and add a secondary KBV check at the integration layer. For Resource Limits, Doubao-pro can be considered, but another model must be deployed in parallel in the Data Boundary scenario to provide complementarity. In Engineering Specs, gpt-6-sol performs stably and can be the first choice for code review and workflow orchestration, but note its shortcoming of only 2.6/4 in Data Boundary. In Safety Compliance, most models can reach above 3.5, making it a relatively low-risk entry point.

Strategic Judgment and Next-Round Validation Signals

The strong performance of the Claude series in Data Boundary and Safety Compliance may stem from a more robust memory mechanism for multiple parallel hard constraints, while its low score in Engineering Specs may reflect a weaker recovery path under “gradually increasing pressure” tactics. The extreme contrast of Doubao-pro scoring full marks in Resource Limits but finishing last in Data Boundary warrants adding more mixed-scenario tasks to the v3.1 task pool next round to test whether this is a systemic defect. The contrast between gpt-6-sol’s 4/4 in Engineering Specs and 2.6/4 in Data Boundary suggests it may be better suited to internal toolchains than to user-facing data-processing workflows.

The low scores in the Data Boundary scenario are not accidental; it is the constraint type most easily broken through by sustained pressure after agreement-setting in the v3 tasks.

For enterprise model selection, expecting a single model to cover all five scenarios is no longer realistic. We recommend splitting deployment by scenario and setting manual review checkpoints in the two highly discriminating scenarios, Data Boundary and Engineering Specs. If the next WDCD round continues to use worst-of-3 sampling, the real risks of uneven-performance models will be further amplified.

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #400 · Scenario Matrix | Evaluation Methodology