WDCD Five-Scenario Cross-Review: Data Boundaries Lowest for All, qwen3-max Scores Just 1.36

The WDCD v3.1 five-scenario cross-review shows that the data boundaries scenario received the lowest scores across the board, with qwen3-max scoring only 1.36/4, far below the highest score of 4/4 in the engineering standards scenario.

Why Data Boundaries Became the Hardest Scenario

In the data boundaries scenario, gemini-2.5-pro ranked first with 3.48/4, followed closely by gpt-5.5 at 3.4/4, while qwen3-max came last at 1.36/4 and doubao-pro scored only 2.64/4. By comparison, the resource constraints scenario champion gemini-3.1-pro reached 3.76/4, with a higher overall median. This indicates that data boundary constraints are more easily breached under multi-round pressure.

The cause lies in the commitment phase of the v3 tasks, where 2-5 parallel hard constraints are injected simultaneously. Continuous pressure includes authority-based special approval and salami-slicing tactics, directly testing the model’s initial memory of data boundaries. After using worst-of-3 sampling, qwen3-max’s scores dropped significantly in the R3 pressure round, indicating that under escalating sunk-cost pressure it prioritizes abandoning constraints rather than holding them.

Business Rules Scenario Shows the Greatest Differentiation

In the business rules scenario, gpt-6.1-sol led with 3.88/4, while gemini-2.5-pro scored only 2.23/4, a gap of 1.65 points, the largest among the five scenarios. claude-opus-4.7 scored 3.48/4, and grok-4 scored 3.03/4. This distribution shows that business rule constraints place the highest demand on a model’s ability to prioritize rules.

Looking at the v2 anchor tasks, after constraints were injected in R1, gpt-6.1-sol maintained high scores through R2 interference and R3 pressure, while gemini-2.5-pro’s score dropped noticeably in R3. In the KBV restatement probe stage, the S_kbv sub-score for constraint memory in business rules tasks had a greater impact on the final total score, widening the differentiation.

Analysis of Model Specialization

gemini-2.5-pro scored 3.51/4 in resource constraints but only 2.23/4 in business rules, a gap of 1.28 points. gpt-6.1-sol scored 3.88/4 in business rules but only 2.33/4 in security compliance, a gap of 1.55 points. qwen3-max showed a gap of 2.04 points between 3.4/4 in engineering standards and 1.36/4 in data boundaries.

These gaps stem from differences in the pressure mechanisms across scenarios. Engineering standards tasks are mostly single-round hard constraints, where gpt-6-sol achieved a perfect 4/4; data boundaries require holding multiple parallel constraints at once, and qwen3-max lost points on S_integrity in the final honest review, causing its total score to collapse.

Specific Recommendations for Enterprise Selection

Enterprises integrating AI into production processes should prioritize gemini-2.5-pro or gpt-5.5 for data boundaries scenarios and add external data validation guardrails. For resource constraints scenarios, they can directly use gemini-3.1-pro, whose 3.76/4 performance is stable and whose breach-recovery S_recover score is relatively high.

For business rules scenarios, gpt-6.1-sol is recommended, but manual review should be added in security compliance scenarios because this model scores only 2.33/4 there. In engineering standards scenarios, gpt-6-sol performs perfectly and can be the first choice for code review tasks, but its security compliance weakness should be noted.

qwen3-max ranks last or near last in four scenarios. If an enterprise is already using it, it is recommended to limit it to low-risk internal tools and to force system prompts that strengthen constraints in data boundaries and security compliance scenarios.

Strategic Judgment

The gemini series’ ability to keep commitments in data boundaries and resource constraints may be underestimated by the market, as its 3.48/4 and 3.76/4 performances outperform most GPT variants. The strengths of the gpt-6 series in engineering standards and business rules may be overestimated, given clear weaknesses in security compliance and data boundaries.

The next validation round should focus on the impact of social-proof pressure rounds in v3 tasks on S_hold commitment-keeping survival, as well as differences in recovery ability across models under worst-of-3 sampling.

The low scores in the data boundaries scenario remind enterprises that no model can never breach under all constraints; model selection is essentially about matching scenarios with guardrail combinations.

Data source: YZ Index WDCD Compliance Leaderboard | Run #428 · Scenario Matrix | Evaluation Methodology