GAN8
A multi-agent decision AI architecture composed of eight independent abstract logical frameworks. Each framework independently produces a YES/NO judgment with a score, and the highest-scoring outcome prevails. The architecture requires no discriminator or referee model and is designed for tasks involving multi-constraint combinatorial reasoning.
Timeline (last 90 days)
GAN8 employs a multi-framework voting architecture in which eight independent abstract logics each give a YES/NO judgment and score, with the highest scorer winning, and there is no discriminator or referee model.
In the first round of 34 logical reasoning questions, a single-head model similar to JEV achieved an accuracy of 94.12%, while GAN8 achieved 93.01%, nearly tied.
In the IID test (21 questions) of the third round's structured combination tasks, GAN8 reached 95.2%, surpassing JEV's manual weights at 76.2%, linear training at 71.4%, and MLP at 85.7%.
On a pure parity task, GAN8 achieved 100%, JEV MLP achieved 67%, and JEV linear achieved 33%.
One of GAN8's logic frameworks directly implements parity checking, whereas JEV's heads have no XOR column.
On unseen rule combination tasks, GAN8 did not win; the trained JEV failed due to overfitting on single-defect templates, and GAN8 failed along with it.
In the fourth round of fusion experiments, treating JEV as the 9th framework and taking the highest score yielded an overall accuracy of 81.5%, but parity dropped back to 50%.
The author recommends that model fusion should route by structure rather than averaging, and should not allow a high-confidence generalist model to act as the judge.
Single-head probabilistic models are suitable for scenarios where features are the labels, while multi-logic architectures are suitable for structural problems with multiplied constraints, and blind fusion often fails on both ends.
All Facts (9)
GAN8 employs a multi-framework voting architecture in which eight independent abstract logics each give a YES/NO judgment and score, with the highest scorer winning, and there is no discriminator or referee model.
50%UnverifiedIn the first round of 34 logical reasoning questions, a single-head model similar to JEV achieved an accuracy of 94.12%, while GAN8 achieved 93.01%, nearly tied.
50%UnverifiedIn the IID test (21 questions) of the third round's structured combination tasks, GAN8 reached 95.2%, surpassing JEV's manual weights at 76.2%, linear training at 71.4%, and MLP at 85.7%.
50%UnverifiedOn a pure parity task, GAN8 achieved 100%, JEV MLP achieved 67%, and JEV linear achieved 33%.
50%UnverifiedOne of GAN8's logic frameworks directly implements parity checking, whereas JEV's heads have no XOR column.
50%UnverifiedOn unseen rule combination tasks, GAN8 did not win; the trained JEV failed due to overfitting on single-defect templates, and GAN8 failed along with it.
50%UnverifiedIn the fourth round of fusion experiments, treating JEV as the 9th framework and taking the highest score yielded an overall accuracy of 81.5%, but parity dropped back to 50%.
50%UnverifiedThe author recommends that model fusion should route by structure rather than averaging, and should not allow a high-confidence generalist model to act as the judge.
50%UnverifiedSingle-head probabilistic models are suitable for scenarios where features are the labels, while multi-logic architectures are suitable for structural problems with multiplied constraints, and blind fusion often fails on both ends.
50%