Grok 4 Tops WDCD with 96.3 Points, Doubao Pro Ranks Last with 67.9 Points — a 28-Point Gap

Grok 4 ranks first on the WDCD v3.1 compliance leaderboard with 96.30 points, while Doubao Pro trails at 67.90 points — a gap of 28.4 points. Among the 11 models evaluated, the top three (Grok 4, GPT-o3 at 95.20 points, and GLM-4.6 at 93.70 points) average 95.07 points, while the bottom three (Qwen3 Max at 71.50 points, Gemini 2.5 Pro at 71.00 points, and Doubao Pro at 67.90 points) average 70.13 points, a spread of 24.94 points.

Ranking Landscape: Stable Leaders, a Disconnected Tail

This pilot ranking uses worst-of-3 sampling with rule-based scoring, equally weighting v3 questions and v2 anchor questions. Grok 4 holds high scores across all three dimensions — S_hold (compliance survival), S_kbv (constraint memory), and S_recover (breach recovery) — with no collapse observed during R3 pressure rounds. GPT-o3 follows closely, just 1.1 points behind, indicating that the o-series' constraint memory under multi-round progressive pressure has reached near-top-tier levels. GLM-4.6 ranks third with 93.70 points, only 1.5 points ahead of fourth-place DeepSeek V4 Pro (92.20 points), forming a clear first tier.

Ranks six through eight form the second tier: Gemini 3.1 Pro with 88.20 points, Claude Opus 4.7 with 85.80 points, and GPT-5.5 with 83.30 points. The gap between this tier and the leaders ranges from 8 to 13 points, with most losses concentrated in the S_integrity (honest self-reporting) dimension. The bottom three models show notable declines in S_hold, with Doubao Pro dropping another 6.6 points to 67.90 points, Gemini 2.5 Pro falling 12.5 points, and Qwen3 Max holding steady at 71.50 points — creating an overall disconnect from the leaders.

Champion Analysis: S_hold Advantage Under Multi-Round Pressure

Grok 4's high score most likely stems from the contract-establishment and continuous-pressure phases in rounds 8–12 of the v3 questions. The v3 questions first set 2–5 parallel hard constraints, then sequentially apply four types of pressure: social proof, authority override, salami slicing, and sunk cost. Grok 4 scores highest on the latest breach round, with the 60-point S_hold component contributing significantly. By contrast, Doubao Pro and Gemini 2.5 Pro breach their contracts earlier in the R3 pressure rounds, resulting in substantial S_hold losses. Although the overall R3 collapse rate remains at 0%, the worst-case sampling already distinguishes models' true pressure resistance in engineering-governance and security-compliance scenarios.

Bottom-Performer Analysis: Low Scores on Both S_integrity and S_recover

Doubao Pro at 67.90 points and Qwen3 Max at 71.50 points perform worst on the S_integrity (honest self-reporting) dimension. The final round of the v3 questions requires models to honestly assess whether they violated constraints; a breach followed by a false claim of compliance scores zero directly. After the KBV recitation probe, the bottom-tier models also show low S_recover, indicating that once a constraint is breached, they struggle to proactively restore compliant states in subsequent rounds. This contrasts with top-tier models, which retain partial S_recover scores even after a breach.

Selection Implications: Real Risk Boundaries for Production Integration

For enterprises integrating AI into production workflows, the WDCD data offers three concrete takeaways. First, in data-boundary and security-compliance scenarios, Grok 4's 96-point and GPT-o3's 95-point performance can be directly applied to high-value processes with minimal guardrail requirements. Second, in resource-constrained and business-rule scenarios, Claude Sonnet 4.6 (91.60 points) and DeepSeek V4 Pro (92.20 points) have entered the usable range, though R3-level pressure simulation testing is still recommended. Third, Qwen3 Max, Gemini 2.5 Pro, and Doubao Pro score low on S_hold in engineering-governance scenarios and are not recommended for direct use in production pipelines that require long-term maintenance of multiple parallel constraints; an external verification layer must be added.

Strategic Assessment: Signals of Underestimation and Overestimation

Claude Sonnet 4.6 rose 13.5 points this round, DeepSeek V4 Pro gained 6.9 points, and GPT-o3 gained 6.4 points. These increases likely stem from the v3 questions' emphasis on multi-round progressive pressure, suggesting their constraint-memory mechanisms have improved through iterations. GPT-5.5 fell 5.3 points, Claude Opus 4.7 dropped 5.7 points, and Gemini 2.5 Pro declined 12.5 points, indicating their relative weaknesses in S_integrity and S_recover may have been overestimated by the market. Signals worth validating in the next round: whether top-tier models can maintain their S_hold lead in resource-constrained scenarios, and whether bottom-tier models can narrow the S_recover gap through single-round recovery training.

This pilot phase does not count toward the main leaderboard, but it has already exposed differences in models' compliance-survival capabilities under real multi-round dialogue pressure. When selecting models, relying solely on general-purpose benchmarks is no longer sufficient to cover production-grade constraint requirements; the worst-of-3 rule-based scoring provided by WDCD v3.1 more closely reflects actual deployment risk.

Compliance is not an add-on feature of a model — it is the fundamental prerequisite for whether a production workflow can operate over the long term.

Data source: YZ Index WDCD Compliance Leaderboard | Run #296 · Overall Ranking | Evaluation Methodology