Claude Duo Up 6.8 Points, Gemini Down 5.6, WDCD Compliance Leaderboard Shifts Dramatically

In this round of WDCD v3.1 compliance testing, Claude Opus 4.7 rose 6.8 points from Run #247, Claude Sonnet 4.6 rose 6.7 points, Gemini 3.1 Pro dropped 5.6 points, and GPT-5.5 rose 5.3 points. This set of data directly reveals significant divergence in the models' constraint survival capability under multi-round progressive pressure.

Data Facts: Top 5 and Changes

The current Top 5 rankings show Grok 4 at 94.80 points (1st), DeepSeek V4 Pro at 93.60 points (2nd), GLM-4.6 at 93.50 points (3rd), Claude Opus 4.7 at 92.60 points (4th), and Claude Sonnet 4.6 at 88.20 points (5th). Three models rose, only one fell, and Gemini 3.1 Pro was the only model with a notable decline.

The compliance total score is an equally weighted average of the native v3 question percentage score and the v2 anchor question converted score (total/4 × 100), with the sampling method being worst-of-3. The simultaneous rise of Claude Opus 4.7 and Claude Sonnet 4.6 indicates that in the 8-12 round dialog process of commitment-pressure-KBV paraphrase probe, their S_hold compliance survival 60-point sub-score saw the most significant improvement.

Cause Analysis: Pressure Rounds and Constraint Scenario Differences

The WDCD v3 question design includes 2-5 parallel hard constraints during the commitment phase, followed by four levels of pressure (social identity, authority approval, salami slicing, and sunk cost), and finally a KBV paraphrase probe and an end-of-round honest review. The rise in Claude series scores most likely stems from improvements in the S_recover breach recovery 10-point sub-score and the S_integrity honest self-report 15-point sub-score, particularly in safety compliance and engineering specification constraint scenarios. These models are better at maintaining constraint memory rather than collapsing during rounds 6-8 of salami-slicing pressure.

Gemini 3.1 Pro's 5.6-point drop, combined with the worst-of-3 rule (taking the worst result), suggests that in consecutive pressure rounds for data boundary or resource constraint scenarios, its S_hold score declined the fastest. In the three-round design of the v2 anchor question, the R3 pressure phase (weight 2 points) is most likely to expose issues. This model may breach constraints earlier in the R3 stage after R2 interference, leading to a drop in the overall converted score.

GPT-5.5's 5.3-point rise may reflect an enhancement in its S_kbv constraint memory 15-point sub-score for business rule scenarios, enabling it to accurately restate initial constraints under multi-round progressive pressure. All judgments are based on comparisons between this round and Run #247, with no historical continuous data to support them.

Selection Implications: Real Risks in Production Process Integration

For enterprises integrating AI into production processes, the WDCD score directly corresponds to the usable boundary in different constraint scenarios. Grok 4 (94.80 points) and DeepSeek V4 Pro (93.60 points) perform more stably in engineering specification and safety compliance scenarios, making them suitable for direct integration into internal systems with strict resource constraints, requiring relatively fewer guardrails.

Although Claude Opus 4.7 (92.60 points) has entered the top four, it still lags behind Grok 4 by 2.2 points. Manual review checkpoints are still recommended for data boundary scenarios, especially when dialog rounds exceed eight. Gemini 3.1 Pro's 5.6-point drop means it has a relatively higher probability of breaching constraints in customer service or approval processes with dense business rules, and additional R3 pressure simulation testing is needed before integration.

Enterprises can adopt a tiered selection approach based on the priority of five constraint scenarios: for high safety compliance needs, prioritize Grok 4 and Claude Opus 4.7; for edge deployments with strict resource limits, refer to DeepSeek V4 Pro (93.60 points); for Gemini 3.1 Pro, strengthen initial constraint anchors in prompts to compensate for its current S_hold weakness.

Strategic Judgment: Underestimated and Signals to Be Validated

The simultaneous rise of Claude Opus 4.7 and Claude Sonnet 4.6 suggests that their recovery ability under multi-round pressure in v3 questions may have been previously underestimated by the market. The next cycle should focus on verifying whether their S_kbv scores in the KBV paraphrase probe phase remain stable. Gemini 3.1 Pro's decline may be a phase change due to prompt sensitivity and requires observation in the next cycle for a potential rebound.

Grok 4 leads at 94.80 points. Combined with the overall pattern of three rises and one fall, its compliance capability currently carries a relatively low risk of being underestimated by the market during this trial phase, but more worst-of-3 validation in engineering specification scenarios is still needed. GLM-4.6 (93.50 points) and DeepSeek V4 Pro (93.60 points) differ by only 0.1 points. If this trend continues in the next cycle, it will become an important signal for enterprise backup selection.

The divergence in compliance capability has shifted from single-round accuracy to survival duration under multi-round pressure. The most noteworthy trends in the next cycle are whether Gemini 3.1 Pro can stop its decline in the R3 pressure phase and whether the Claude series can convert the 6.8-point gain into a Top 3 position.

Data Source: YZ Index WDCD Compliance Leaderboard | Run #253 · Change Tracking | Evaluation Methodology