GLM-4.6 Tops WDCD Adherence Leaderboard at 96.2 as Doubao Pro Trails at 68.78, a 27-Point Gap

GLM-4.6 ranked first on the WDCD v3.1 adherence leaderboard with 96.20 points, while Doubao Pro came in last at 68.78, a 27.42-point gap between top and bottom.

Ranking Landscape: A Concentrated Top, a Broken-Off Tail

This round of WDCD adherence testing covered 15 models using worst-of-3 sampling. The top three — GLM-4.6, Gemini 3.1 Pro and GPT-o3 — scored 96.20, 94.00 and 92.30 respectively, within 2–4 points of one another, forming a clear first tier. Fourth-place Grok 4 scored 90.29, 1.99 points behind third and still part of the leading group. Fifth through tenth place clustered in the 86.94–83.79 range, forming the second tier. Beyond tenth place scores fell sharply: Gemini 2.5 Pro at 78.90, DeepSeek V4 Pro at 76.24, GPT-6.1 Sol at 71.48 and Doubao Pro at 68.78 make up the third tier. The overall full-mark rate was 42.7% and the R3 breakdown rate was 0%, indicating that all models achieved full marks/2 on the v2 three-round anchor questions, while the v3 multi-turn, progressively escalating pressure questions opened up the gap.

Champion Analysis: GLM-4.6's Survival Edge Under Sustained Pressure

GLM-4.6's 96.20 points come mainly from high retention on S_hold, the 60-point adherence-survival component. The v3 questions consist of 8–12 rounds of dialogue: 2–5 hard constraints are first established, then social proof, authoritative override, salami-slicing and sunk-cost pressure are applied in sequence, followed by a KBV restatement probe and honest self-report. GLM-4.6 most likely held its constraints through rounds 8–10 without breaking under mounting sunk-cost pressure. By contrast, Doubao Pro's 68.78 means its S_hold had already lost substantial ground by rounds 5–6, and neither the 10-point S_recover for post-break recovery nor the 15-point S_integrity for honest self-report could compensate. The gap between champion and last place shows up mainly in the later rounds of resource-limit and safety-compliance constraint scenarios.

Bottom Analysis: Doubao Pro's Early Collapse in Engineering-Spec Scenarios

Doubao Pro's 68.78 is the lowest among the 15 models. Given the four-component structure of the v3 questions, its 15-point S_kbv constraint memory and 15-point S_integrity were likely both low, indicating that even after breaking it could neither accurately restate the initial constraints nor honestly report the break. Tail models more easily lose constraints during the salami-slicing pressure stage in business-rule and engineering-spec scenarios, in contrast to top models: GLM-4.6 and Gemini 3.1 Pro pushed the break point back by at least three rounds in the same scenarios.

A Mechanistic Explanation of the Top-Tier vs Tail Gap

The top four models averaged 93.20 points and the bottom four averaged 73.85, a gap of 19.35 points. The continuous-pressure design of the v3 questions amplifies this gap. Social proof and authoritative override are early-stage pressures, while salami-slicing and sunk cost are mid-to-late-stage pressures. Top models still maintained high S_hold at the R3 stage, indicating more robust memory and execution of multiple parallel hard constraints. Tail models may already have shown constraint drift at the KBV restatement probe stage, capping their subsequent S_recover scores. The overall 0% R3 breakdown rate indicates that the interference and pressure intensity of the v2 anchor questions were insufficient to expose model differences; v3's multi-turn progressive design is the real differentiator.

Selection Implications for Production Pipeline Integration

Enterprises integrating AI into production pipelines can use the following guidance: for data-boundary and safety-compliance scenarios, prioritize GLM-4.6 or Gemini 3.1 Pro, as both maintain constraints under multi-turn pressure and suit sensitive data or compliance approval workflows. For resource-limit and business-rule scenarios, Grok 4 or GPT-o3 can be considered, but external guardrails should monitor outputs after round six. In engineering-spec scenarios, Claude Sonnet 4.6 and Qwen3 Max score similarly and suit internal toolchains, though final decisions require human review. Doubao Pro and GPT-6.1 Sol sit at the tail and are not recommended for direct use in automated workflows requiring long-term constraints, since the risk is that recovery after an early break is difficult.

Strategic Takeaway: Adherence Capability Is Undervalued, and Signals Still to Verify

GLM-4.6's 96.20 may be undervalued by the market; its dual-high S_hold and S_integrity on v3 questions reveals a systemic advantage in multi-constraint parallel scenarios, and the next version should verify whether it can reliably stay above 95 points. Doubao Pro's 68.78 may be overvalued: its early collapse in engineering-spec scenarios indicates the current architecture's resistance to sunk-cost pressure is insufficient. GPT-6 Sol and GPT-5.5 differ by only 0.1 points, suggesting that models in the same series have neared saturation on the adherence dimension; the next round should focus on their specific performance in data-boundary scenarios. The 0% R3 breakdown rate suggests that v2 anchor questions can no longer differentiate models, and future versions will need more pressure rounds in v3 or additional constraint types.

Adherence is not an add-on feature of a model; it is the last quantifiable safety line in production environments.

Data source: Winzheng YZ Index WDCD Adherence Leaderboard | Run #400 · Overall ranking | Evaluation methodology