WDCD Constraint-Adherence Leaderboard: Gemini 3.1 Pro Leads with 91 Points as Qwen3 Max Trails at 58.85, a 32-Point Gap

In the WDCD v3.1 constraint-adherence test, Gemini 3.1 Pro took first place with 91.00 points, while Qwen3 Max finished last at 58.85, a gap of 32.15 points. The result comes directly from an equal-weight average of the v3 percentage score under worst-of-3 sampling and the v2 anchor-question conversion.

Ranking Landscape: Concentration at the Top, a Cliff at the Bottom

The top five span 91.00 to 82.71 points, a gap of only 8.29 points; the bottom five slide from 73.61 down to 58.85, a spread of 14.76 points. Gemini 3.1 Pro's v2 anchor scores of R1=1.00, R2=0.88 and R3=1.63/2 show that it retains considerable constraint memory after three rounds of interference and pressure. Qwen3 Max's R1=R2=R3 all at 0.00 means it failed completely at the constraint-injection stage.

Why the Winner Won: High R3 Scores Come from Constraint Memory and Recovery

Gemini 3.1 Pro's S_hold and S_kbv scores drove up its total, mainly through resistance in "data boundary" and "security compliance" scenarios under v3's multi-round escalating pressure. An R3 score of 1.63/2 shows that it maintained part of its original constraints even under the combined pressure of authoritative override and sunk cost. By contrast, GLM-4.6 has R1=1.00 but only 1.00/2 on R3, for a total of 84.38, nearly 7 points behind Gemini 3.1 Pro.

The Bottom Mechanism: R1 at Zero Exposes Failure of Basic Constraint Injection

Qwen3 Max scored R1=0.00 in the first round of the v2 anchor questions, indicating that the hard constraints set at the contracting stage were not internalized by the model. R2 and R3 continued at 0, showing that social proof and salami-slicing pressure worked completely. Doubao Pro's R1=0.88 already beats Qwen3 Max, but with R3 at only 0.88/2 and a total of 63.68, it remains second to last.

What the Gap Between the Top Tier and the Tail Means for Model Selection

For enterprises integrating AI into production workflows, Gemini 3.1 Pro and Grok 4 (84.18 points) can be used directly in "resource limits" and "engineering standards" scenarios, with relatively strong recovery after a breach. But even a 91-point model is not perfect; hard constraints involving "business rules" still require additional guardrails. Qwen3 Max and Doubao Pro should not be called directly in any workflow that requires parallel hard constraints, since R1 at zero means the very first round of dialogue may violate preset boundaries.

Strategic Judgment: The Market Overestimates Some Models' Adherence

This period's data shows GPT-6 Luna (67.98 points) and Gemini 2.5 Pro (67.22 points) scoring 0.25 and 0.75 on R2, far below other models in their tier, indicating faster decay of constraint memory under sustained pressure. Qwen3 Max's 58.85 may understate its actual risk, because all three of its v2 rounds went to zero; any deployment that relies on initial constraints will face systemic failure.

Compared with the previous period, Qwen3 Max fell 26.2 points and GPT-6 Luna fell 15.8 points; the two largest declines both appear at the tail, confirming the amplifying effect of R3 collapse on total scores. DeepSeek V4 Pro rose 2.3 points, mainly from a higher R3 score, showing improved recovery under sunk-cost pressure.

Differences Across Five Types of Constraint Scenarios

Data boundary and security compliance scenarios produced the widest gaps, with top models' S_hold averaging more than 15 points above the tail. In resource-limit scenarios, R2 interference affected mid-tier models most: GPT-6 Sol scored R2=0.63, below Grok 4's 0.88. In engineering-standard scenarios, Claude Opus 4.7 and GPT-o3 scored close together, both in the 80-point range, making them suitable as backups.

A global R3 collapse rate of 9.2% shows that more than one in ten models fully broke before the final honest review. A full-score rate of 41% shows that the current v3.1 question pool still exerts meaningful pressure on most models.

Constraint adherence is not a by-product of model parameters but a prerequisite for production deployment.

Next period's validation should focus on how models scoring above 1.5 on R3 perform in real business-rule scenarios, and on whether R1-zero models such as Qwen3 Max can recover basic constraint ability after the v3 question pool is adjusted.

Data source: YZ Index WDCD Constraint-Adherence Leaderboard | Run #406 · Overall ranking | Evaluation Methodology