WDCD Five-Scenario Comparative Review: Data Boundary Scores Lowest at 1.7, Security Compliance Becomes the Biggest Weakness

In the WDCD v3.1 pilot, the Data Boundary scenario produced the lowest score across the board, 1.7 (qwen3-max), while the Security Compliance scenario had the lowest score at 1.8 (doubao-pro). These two scenarios are also the test dimensions with the lowest overall model scores and the greatest discrimination.

Why Data Boundary Became the Most Difficult Scenario

In the Data Boundary scenario, glm-4.6 led with a perfect 4/4, while qwen3-max scored only 1.7/4, a gap of 2.3 points. The Security Compliance scenario likewise showed a 2.2-point gap (claude-opus-4.7 at 4/4 versus doubao-pro at 1.8/4). By contrast, in the Engineering Standards scenario, the highest and lowest scores differed by only 0.9 points, making it the least discriminative.

This difference stems from the continuous-pressure design of the v3 questions: after the commitment is established in the Data Boundary question, three consecutive rounds apply a combination of “social proof + salami slicing” pressure, and the model must still strictly uphold the data-scope constraint in rounds 6–8. In this scenario, qwen3-max broke down multiple times after round 5, and its S_hold score fell sharply, leading to the final low score of 1.7/4.

The Real Risks of Lopsided Models

claude-opus-4.7 scored 4/4 in both Resource Constraints and Security Compliance, but only 2.5/4 in Data Boundary, a 1.5-point gap between scenarios. gpt-6-astra scored 4/4 in Business Rules but only 2.3/4 in Data Boundary, a 1.7-point gap. gpt-6-sol scored 4/4 in Business Rules and 2.35/4 in Data Boundary, a 1.65-point gap.

These models perform robustly on Security Compliance and Resource Constraints, but are clearly deficient in data-boundary constraint memory (S_kbv) and recovery after breakdown (S_recover). For enterprises integrating them into production workflows, this means they cannot judge a model reliable merely because it achieves a perfect score in a single scenario; additional guardrails must be added specifically for Data Boundary.

Specific Implications of Each Scenario for Enterprise Model Selection

In the Security Compliance scenario, five models—claude-opus-4.7, claude-sonnet-4.6, deepseek-v4-pro, gemini-3.1-pro, and glm-4.6—all maintained 4/4 and are suitable for direct use in compliance-sensitive processes. gpt-5.5 scored only 3/4 in this scenario, 1 point lower than its 4/4 in Business Rules; adding manual review in security compliance steps is recommended.

In the Resource Constraints scenario, doubao-pro scored only 2.5/4, far below claude-opus-4.7’s 4/4. If a company’s core scenario involves resource quota control, doubao-pro needs an additional constraint layer; otherwise it may easily breach limits under sunk-cost pressure.

The Engineering Standards scenario had the highest overall scores, with even the lowest reaching 3.1/4, indicating that most models have relatively strong constraint-abiding ability under this type of constraint and can serve as low-risk options for the first batch integrated into production workflows.

Strategic Judgment and Signals for the Next Round of Validation

claude-opus-4.7’s Security Compliance and Resource Constraints capabilities may be overestimated by the market; its actual Data Boundary performance of 2.5/4 is clearly distant from its perfect-score scenarios. deepseek-v4-pro scored 4/4 in all three scenarios—Business Rules, Engineering Standards, and Security Compliance—and only 2.8/4 in Data Boundary. Its overall constraint-abiding capability is underestimated and merits focused validation by enterprises.

The sharp contrast between qwen3-max’s 1.7/4 in Data Boundary and 3.4/4 in Security Compliance suggests that its constraint-memory mechanism lacks stability across scenarios. The next pilot should focus on whether glm-4.6 can continue to maintain 4/4 in Data Boundary, and whether the Claude series models can improve their S_recover scores in that scenario.

The 1.7 score in Data Boundary and 1.8 score in Security Compliance have clearly drawn a red line; when selecting models, enterprises must include the guardrail costs for these two scenarios in their total budget.

Based on: YZ Index WDCD Compliance Leaderboard | Run #379 · Scenario Matrix | Evaluation Methodology