WDCD Benchmark Comparison: Data Boundaries Becomes the Hardest Scenario as GLM-4.6 Collapses with Only 1.36

In the WDCD v3.1 test, the data-boundary scenario produced the lowest scores across the board: glm-4.6 scored only 1.36/4, while deepseek-v4-pro and gemini-3.1-pro tied at 3.8/4, a gap of 2.44 points.

Why Data Boundaries Became the Hardest Scenario

The data-boundary scenario's average performance was significantly lower than that of the other four categories. deepseek-v4-pro and gemini-3.1-pro both scored 3.8/4, doubao-pro 3.6/4, claude-opus-4.7, gemini-2.5-pro and grok-4 each 3.4/4, gpt-o3 3.28/4, claude-sonnet-4.6 3.12/4, qwen3-max 3.1/4, gpt-5.5 2.52/4, and glm-4.6 last at 1.36/4. The v3 questions in this scenario use 8–12 rounds of dialogue: first, 2–5 parallel hard constraints are agreed upon, then social proof, authoritative special approval, salami slicing and sunk cost are applied to ratchet up the pressure level by level, and finally a KBV restatement probe and an honest final-round debrief are given. In its worst sampling run, glm-4.6 most likely broke the agreement as early as rounds 4–6, and its S_integrity item was scored 0 for falsely claiming it had stayed clean, driving its total far below the other models.

Business Rules Show the Largest Discrimination

In the business-rules scenario, the top spot was shared by claude-opus-4.7, gemini-2.5-pro and grok-4, all with a perfect 4/4, followed by gpt-o3 at 3.67/4, gpt-5.5 3.37/4, claude-sonnet-4.6, deepseek-v4-pro and gemini-3.1-pro all at 3.33/4, doubao-pro 2.93/4, qwen3-max 2.73/4, and glm-4.6 2.27/4. The gap between the perfect score and the bottom was 1.73 points, wider than the 0.98-point gap in resource limits and the 0.68-point gap in engineering standards. This scenario matters greatly to "enterprises plugging AI into production workflows": once the terms were set, claude-opus-4.7 fully withstood interference from three rounds of anchor questions, making it suitable for direct use in contract-clause validation, pricing-rule execution and other high-value business tasks.

Uneven Performance in Resource Limits and Safety Compliance

In the resource-limits scenario, grok-4 led at 3.41/4, followed by qwen3-max 3.4/4, gpt-o3 3.37/4, gemini-3.1-pro 3.26/4, deepseek-v4-pro 3.23/4, doubao-pro 3.09/4, claude-sonnet-4.6 3.01/4, gemini-2.5-pro 2.81/4, gpt-5.5 2.71/4, claude-opus-4.7 2.69/4, and glm-4.6 2.43/4. claude-opus-4.7 scored 1.31 points lower here than in business rules, indicating that its ability to keep commitments declines under constraints such as resource quotas and concurrency limits. In the safety-compliance scenario, grok-4 was highest at 3.86/4, followed by claude-opus-4.7 3.63/4, gemini-3.1-pro 3.61/4, gpt-o3 3.43/4, gpt-5.5 3.11/4, claude-sonnet-4.6 3.04/4, deepseek-v4-pro 2.91/4, gemini-2.5-pro 2.86/4, doubao-pro 2.84/4, glm-4.6 2.43/4, and qwen3-max last at 2.41/4. qwen3-max scored 1.08 points lower here than in engineering standards, exposing its fragility under pressure at compliance boundaries.

Engineering Standards Relatively Balanced

In the engineering-standards scenario, gemini-3.1-pro and gpt-o3 tied at 3.84/4, followed by grok-4 3.79/4, gemini-2.5-pro 3.63/4, deepseek-v4-pro 3.57/4, qwen3-max 3.49/4, claude-opus-4.7 3.46/4, doubao-pro 3.43/4, glm-4.6 3.4/4, claude-sonnet-4.6 3.34/4, and gpt-5.5 3.16/4. The lowest score still reached 3.16/4, showing that this scenario puts relatively little pressure on most models and that enterprises can prioritize it for pilots.

Concrete Recommendations for Enterprise Selection

Enterprises plugging AI into production workflows must deploy additional output filtering and human review in the data-boundary scenario, because glm-4.6 and gpt-5.5 scored only 1.36/4 and 2.52/4 there respectively, carrying the highest risk of breaking commitments. For the business-rules scenario, claude-opus-4.7 or grok-4 can be chosen directly, as their 4/4 results show they can reliably enforce hard constraints. For resource limits, grok-4 is the preferred choice: its 3.41/4 leads the field and it also reaches 3.86/4 in safety compliance, giving it a cross-scenario advantage. qwen3-max scores 3.49/4 on engineering standards but only 2.41/4 on safety compliance, so compliance-related workflows require an added secondary check. glm-4.6 is passable at 3.4/4 on engineering standards but collapses on data boundaries, making it suitable only for internal code-standard checks rather than customer data processing.

Strategic Judgment

claude-opus-4.7's perfect business-rules score may be overrated by the market; its shortfall in resource limits, at only 2.69/4, is clearly exposed under multi-round pressure. grok-4 ranks in the top two in both resource limits and safety compliance, so its commitment-keeping ability may be underestimated and is worth focused verification in the next v3.2 round for its S_recover performance under parallel constraints. The 2.44-point spread in the data-boundary scenario suggests that memory decay before the KBV restatement probe remains unsolved in most current models, and enterprises should treat this scenario as the first acceptance threshold for commitment-keeping ability.

The gulf between 1.36 on data boundaries and 4 on business rules means the next generation of models must break through on both constraint memory and multi-round pressure resistance at once; otherwise, enterprise production adoption will remain stuck at the stage of "usable, but not trusted for full use."

Data source: YZ Index WDCD Commitment-Keeping Leaderboard | Run #331 · Scenario Matrix | Evaluation Methodology