Grok 4 Tops WDCD Promise-Keeping Leaderboard with 97.5 Points, Qwen3 Max Ranks Last with 69.5 Points, a 28-Point Gap

In the WDCD v3.1 promise-keeping test, Grok 4 ranked first with a score of 97.50, while Qwen3 Max ranked 11th with 69.50 points, a 28-point difference between the two models.

Ranking Landscape: Concentrated at the Top, Disconnected at the Bottom

The score distribution across the 11 models in this round shows clear stratification. The top three — Grok 4 (97.50), GLM-4.6 (91.80), and Claude Opus 4.7 (91.50) — all exceeded 90 points; ranks 4 through 8 clustered in the 85–88 point range; a clear gap appeared after 9th-place Claude Sonnet 4.6 (78.10), with Doubao Pro (74.50) and Qwen3 Max (69.50) falling below 70 points. The overall full-score rate stood at 63.6%, with an R3 collapse rate of 6.4%.

Champion Analysis: Grok 4 Maintains the Longest Survival Under Multi-Round Pressure

Grok 4 scored full marks in all three rounds of the v2 anchor questions (R1=1.00, R2=1.00, R3=2.00/2), meaning that after making its commitment, it survived the longest when subjected to four levels of escalating pressure: social approval, authority approval, salami slicing, and sunk cost. With S_hold (promise-keeping survival) carrying a 60-point weight, being the last to break down directly lifted its total score. By comparison, 11th-place Qwen3 Max also scored full marks in R1 and R2 but received 0 in R3, indicating that its constraints failed immediately during the high-pressure phase of round 3, resulting in significant S_hold deductions.

Why the Bottom Ranks: Weak Constraint Memory and Recovery in the R3 Stage

Qwen3 Max's 69.50 points mainly stem from low scores on two v3 metrics: S_kbv constraint memory (15 points) and S_recover post-breakdown recovery (10 points). The v3 questions embed KBV restatement probes across 8–12 rounds of dialogue. After continuous pressure, Qwen3 Max's accuracy in restating its initial hard constraints was significantly lower than that of the leading models. Claude Opus 4.7 scored 91.50 in the same period, also with R3 at 1.00/2, showing that in the same R3 stage, Claude Opus 4.7 achieved a higher S_recover score and could return to its constraint framework more quickly after breaking down.

Mechanistic Differences Between the Leading and Bottom Tiers

The largest score gaps between the top three and bottom three models lie in R3 performance. Grok 4, GLM-4.6, and Claude Opus 4.7 scored 2.00, 2.00, and 1.00 on R3, respectively; the bottom three — Claude Sonnet 4.6, Doubao Pro, and Qwen3 Max — all scored 0.00 on R3. The "continuous pressure → KBV restatement probe" segment of the v3 questions directly exposed the bottom-tier models' faster memory decay for parallel hard constraints in multi-round dialogue.

Implications for Selecting Models for Production Integration

Enterprises integrating AI into data-boundary or security-compliance scenarios may prioritize Grok 4 and GLM-4.6. Both models maintain relatively high S_hold scores during the high-pressure R3 stage, making them suitable for workflows that require long-term enforcement of resource limits or engineering standards. GPT-o3 (88.80) and GPT-5.5 (88.60) showed 0 scores as early as the R2 stage, indicating that their constraints begin to loosen after the second round of interference; these models are better used in business-rule scenarios only after additional guardrails are added. Doubao Pro and Qwen3 Max completely lost hold in the R3 stage; it is recommended to use them only in low-risk, non-compliance modules, or to layer an external verification mechanism on top.

Strategic Assessment: Signals of Underestimated and Overestimated Promise-Keeping Capability

Claude Opus 4.7 scored 91.50 this round, up 6.1 points from the previous round. Its R3 score of 1.00/2 shows that it retains a certain recovery capability in high-pressure rounds — a capability that the market may be underestimating. Qwen3 Max's combination of 69.50 points and an R3 score of 0 shows that its constraint-survival capability under multi-round progressive pressure is notably weaker than that of models in the same tier — a gap worth focused verification in the next round. Gemini 2.5 Pro rose 8.7 points from the previous round, but its R3 remains at 0, suggesting that its improvement came mainly from the first half of the v3 questions; its R3-stage stability still requires observation.

The data from this pilot phase indicates that the R3-stage score directly determines whether a model lands in the top three or the bottom three of the final ranking. When selecting models, it is recommended to treat the R3 score as a hard threshold: models scoring below 1.00/2 need additional monitoring before being deployed in security-compliance or resource-constrained scenarios.

Promise-keeping capability is not a byproduct of model parameters — it is the decisive metric for whether production environments can rely on a model over the long term.

Data source: YZ Index WDCD Promise-Keeping Leaderboard | Run #285 · Overall Ranking | Evaluation Methodology