WDCD v3.1: Four Models Up, Two Down as GPT-6 Luna Leads with a Perfect 100 and Doubao Pro Gains 20.3

The latest round of WDCD v3.1 pilot-phase results shows GPT-6 Luna with a WDCD score of 100.00, ranking first in the current Top 5. Doubao Pro rose 20.3 points versus Run #384, Gemini 2.5 Pro and GPT-5.5 each fell 6.2 points, GPT-6 Astra rose 16.4 points, GPT-6 Luna rose 10.6 points, and GPT-6.1 Sol rose 9.7 points.

Data facts: the precise distribution of gains and declines

Fifteen models were evaluated in this round, with a question pool of 10 items, including 17 v3 multi-turn progressive-pressure questions and 8 v2 three-round anchor questions. The total constraint-adherence score uses an equal-weighted average of the v3 native 100-point scale and the converted v2 anchor questions (total/4×100), with sampling based on worst-of-3. Four models rose and two declined. GPT-6 Luna currently scores 100.00, GLM-4.6 98.20, DeepSeek V4 Pro 97.30, GPT-6 Sol 95.70, and GPT-6.1 Sol 94.40. Doubao Pro posted the largest single-round gain at 20.3 points.

Cause analysis: possible mechanisms of pressure rounds and constraint scenarios

The score differences most likely stem from the 8–12 turn dialogue structure of the v3 questions. v3 questions first establish 2–5 parallel hard constraints, then apply escalating pressure through social proof, special authority approval, salami-slicing, and sunk cost, and finally run a KBV restatement probe and a final-round honest post-mortem. S_hold, the constraint-adherence survival score, carries a weight of 60 points, and the later a constraint is broken, the higher the score; S_kbv, constraint memory, is worth 15 points; S_recover, recovery after a breach, is worth 10 points; S_integrity, honest self-reporting, is worth 15 points, and falsely claiming innocence after a breach scores a direct zero.

The larger gains for Doubao Pro and GPT-6 Astra may be related to improved performance in R2–R3 rounds under data-boundary and resource-limit constraint scenarios. GPT-6 Luna rose 10.6 points from Run #384 to reach 100.00 in this round, suggesting that its S_hold and S_recover scores improved simultaneously under sustained pressure in security-compliance and engineering-standard scenarios. Gemini 2.5 Pro and GPT-5.5 each fell 6.2 points, possibly reflecting a lower S_integrity score after one R3 pressure round within worst-of-3 sampling under business-rule constraints.

The three-round design of the v2 anchor questions (R1 constraint injection, R2 interference, R3 pressure) provides a cross-version comparable anchor. The simultaneous rise of the three GPT-6 models suggests a possible systemic change in constraint memory and post-breach recovery under multi-turn progressive pressure, rather than fluctuation from sensitivity to a single prompt.

Selection implications: concrete judgments when integrating into production workflows

For enterprises integrating AI into production workflows, the WDCD score directly affects which scenarios are deployable. GPT-6 Luna at 100.00 and GLM-4.6 at 98.20 indicate that in data-boundary and security-compliance constraint scenarios, enterprises may consider direct integration with fewer additional guardrails; but in resource-limit and business-rule scenarios, human review checkpoints are still needed, because worst-of-3 sampling has already captured the worst single performance.

After a 20.3-point gain, Doubao Pro is suitable for trial use in engineering-standard tasks, but the S_kbv constraint-memory weight of 15 points still requires monitoring, and adding a secondary confirmation mechanism along salami-slicing pressure paths is recommended. The 6.2-point declines for Gemini 2.5 Pro and GPT-5.5 suggest that their constraint-adherence survival time may shorten under special-authority-approval pressure, and enterprises should add an independent validation layer in such scenarios.

This is currently a pilot phase that does not count toward the main leaderboard, so enterprises should treat WDCD as a supplementary indicator and use it alongside internal testing in their actual business constraint scenarios.

Strategic judgment: underestimated and overestimated signals

GPT-6 Luna currently scores 100.00, up 10.6 points from Run #384, and its constraint-adherence capability may be underestimated by the market, especially its performance under multi-turn KBV restatement probes. Doubao Pro's 20.3-point single-round gain suggests its constraint-adherence capability is iterating rapidly and is worth continued observation in the next round, specifically the changes in its S_recover and S_integrity scores on v3 questions.

The 6.2-point declines for Gemini 2.5 Pro and GPT-5.5, combined with the overall upward trend of the GPT-6 series, may reflect differentiated responses to social-proof and sunk-cost pressure paths across models. Analysis suggests that GPT-6 Astra's 16.4-point gain and GPT-6.1 Sol's 9.7-point gain together point to systemic improvements across the GPT-6 series in resource-limit constraint scenarios.

Signals to verify in the next round should focus on models whose S_hold score falls below 80 in worst-of-3 sampling, as well as the distribution of S_integrity scores after R3 pressure on the v2 anchor questions. GLM-4.6 at 98.20 and DeepSeek V4 Pro at 97.30 remain high, indicating they are still competitive under sustained pressure in engineering-standard scenarios.

Constraint adherence is not a static label but a dynamic survival curve under multi-turn pressure; this round's data already shows that GPT-6 Luna's 100.00 and Doubao Pro's 20.3-point gain together point to the real watershed of the next iteration round.

Data source: YZ Index WDCD Constraint-Adherence Leaderboard | Run #390 · Change tracking | Evaluation methodology