WDCD Five-Scenario Comparative Review: Engineering Standards Lowest at 2.45; Doubao and Claude Bias Gap Reaches 1.45 Points

WDCD v3.1's five-constraint-scenario comparative review shows that the engineering standards scenario has the lowest overall scores, with Doubao-pro scoring only 2.45/4, making it the hardest area for compliance among the 11 models.

Engineering Standards Emerge as Biggest Weakness; Pressure Rounds Expose Memory Breaks

In the engineering standards scenario, claude-sonnet-4.6 and grok-4 tie at 4/4, gemini-2.5-pro and glm-4.6 both score 3.7/4, while Doubao-pro scores only 2.45/4, 0.8 points lower than the second-lowest deepseek-v4-pro. This scenario requires the model to simultaneously maintain three parallel hard constraints in multi-turn dialogue: code style, dependency versions, and deployment order. In the v3 question design, the "salami-slicing" pressure in rounds 6–9 is most likely to trigger breaks; Doubao-pro loses the engineering standards anchor already in the R3 anchor question, causing a sharp drop in S_hold score. By comparison, the lowest score in the data boundary scenario is 2.8/4, and the lowest in resource limits is 3.25/4, indicating that when constraint density and pressure intensity are combined in engineering standards, model memory decays fastest.

Safety Compliance Shows Largest Discrimination; qwen3-max and claude-sonnet-4.6 Gap 1.2 Points

In the safety compliance scenario, champion claude-sonnet-4.6 takes 4/4, with gpt-5.5, gpt-o3, and grok-4 close behind at 4/4; qwen3-max is last at 2.8/4, a range of 1.2 points, making this the most discriminative of the five scenarios. The v3 question sets a KBV restatement probe in the safety compliance scenario, requiring the model to restate the initial compliance boundary after round 8. Under social proof pressure, qwen3-max's S_kbv constraint memory score is significantly lower than gemini-3.1-pro's 3.85/4. If enterprises integrate AI into compliance review processes, the compliance differences in this scenario directly determine whether additional human review is needed.

Business Rules Scenario: deepseek-v4-pro Takes Sole Lead, claude-opus-4.7 Unusually Lags

In the business rules scenario, deepseek-v4-pro, gpt-o3, grok-4, and qwen3-max all score 4/4, while claude-opus-4.7 scores only 3.5/4, lower than glm-4.6's 3.9/4. This scenario emphasizes parallel constraints of business rules; claude-opus-4.7 breaks under the sunk-cost escalation stage, and its S_recover recovery score fails to pull back. Analysis shows that claude-opus-4.7 gets a full 4/4 in data boundaries but loses 0.5 points in business rules, indicating a structural difference in its compliance capability across different constraint types.

Subject-Biased Models in Practice: Doubao-pro's Gap Is Largest at 1.45 Points

Among subject-biased phenomena, Doubao-pro's resource limits score of 3.9/4 and engineering standards score of 2.45/4 differ by 1.45 points; qwen3-max's business rules score of 4/4 and safety compliance score of 2.8/4 differ by 1.2 points; deepseek-v4-pro's business rules score of 4/4 and safety compliance score of 3/4 differ by 1 point. These gaps all come from worst-of-3 sampling, representing the model's true performance in its worst run. claude-sonnet-4.6's engineering standards score of 4/4 and business rules score of 3/4 differ by 1 point, showing it is more stable under engineering-type constraints.

Specific Enterprise Selection Advice: Add Guardrails by Scenario Rather Than Full Replacement

For enterprises integrating AI into production processes, the engineering standards scenario requires external validation nodes. Doubao-pro scores 2.45/4 in this scenario, so it is recommended only for low-risk prototype validation; claude-sonnet-4.6 scores 4/4 in engineering standards and can be used directly for code review and deployment script generation. In the safety compliance scenario, qwen3-max scores 2.8/4 and requires an additional human final-review step; claude-sonnet-4.6 and gpt-5.5 both score 4/4 and can serve as a first-layer filter. In the data boundary scenario, most models score 4/4, so risk is lower and they can be tried first.

Strategic Judgment: claude-sonnet-4.6's Compliance Capability May Be Underestimated

claude-sonnet-4.6 scores 4/4 in both safety compliance and engineering standards, while claude-opus-4.7 scores only 3.5/4 in business rules, showing that the sonnet version is more robust in multi-constraint parallel scenarios. deepseek-v4-pro scores 4/4 in business rules but only 3/4 in safety compliance, so its compliance capability may be overestimated by the market. If the next pilot round adds long-dialogue pressure in rounds 10–12 of the engineering standards scenario, it can further verify the recovery capabilities of Doubao-pro and the gemini series.

The 2.45 score in engineering standards is not the end point, but a reminder to enterprises: compliance capability must be evaluated separately by scenario, rather than relying on a single model's total score.

Data source: YZ Index WDCD Compliance Leaderboard | Run #346 · Scenario Matrix | Evaluation Methodology