Subnet Mechanism

Dojo V2 transforms conventional Generative Adversarial Network (GAN) into a competitive, decentralized GAN built on the principle of zero-sum incentives. Unlike traditional GAN where the generator simply aims to mimic ground truth data, V2 challenges miners to create outputs that are not only indistinguishable from a high-quality baseline but are also superior to it. This creates a competitive environment where Bittensor miners, acting as both generators and discriminators, are in constant competition to produce and identify the best possible work.

Responsibilities of Miners

In each task, miners are now required to play either one of the 2 roles:

• Generator: Produce a completion according to the provided task prompt (coding, 3D annotation etc.).

• Discriminator: Evaluate a set of anonymous outputs and vote for the superior output.

Responsibilities of Validators

Validators orchestrates the task creation and scoring pipeline for each of the 3 task types:

Synthetic Task: To benchmark miners against a known, high-quality standard

1

Task Genesis and Validator Baseline Generation

Validator generates a task prompt, and a high quality baseline output using SOTA models.

2

Generator Miner Completion

A single miner is assigned as the Generator to produce their own completion.

3

Discriminator Voting

Those completions are anonymously presented to the rest of the miners who will play as Discriminators and vote for the superior output.

Organic Duel: To create self-sustaining competition among miners to improve output quality

1

Task Genesis

Validator generates a task prompt, and assigns 2 miners who will play as Generators.

2

Generator Miners Completion

Both assigned generators will work on the completion for submission before the expiry time.

3

Discriminator Voting

Those completions are anonymously presented to the rest of the miners who will play as Discriminators and vote for the superior output.

Trap Round: To regulate honest behaviour and deter collusion among miners

1

Task Genesis and Augmentation

Validator generates a task prompt and may introduce augmentations based on random probability. This includes prompt augmentations such that the two miners will receive 2 different prompts.

2

Generator Miner Completion

Both assigned generators will work on the completion based on the received task prompts for submission before the expiry time.

3

Validator-side Augmentation

After receiving the completions from assigned generators, there is a random probability where validators will augment the miner's output to introduce partial ground truth, which will be used to evaluate the truthfulness of discriminators.

4

Discriminator Voting

Those completions are anonymously presented to the rest of the miners who will play as Discriminators and vote for the superior output. This round is not for ranking Generators but for identifying and penalizing incompetent or malicious Discriminators.

Scoring Mechanism

Miner scores will be calculated according to the nature of the task as described below, and will be aggregated before weight setting.

Synthetic Task

The output of the generator miner will be evaluated against the output by the validator that issues the task by the discriminator miners. The total scores distributed will be exactly 1.

1

For discriminators that voted for the output generated by validator

$$discriminator\_score = 1/n$$

where n equals to the number of total discriminators that participated in the task

2

For discriminators that voted for the output generated by generator miner

$$discriminator\_score = 0$$

Due to zero sum nature, the score will be forfeited to the generator.

3

Generator will gain

$$generator\_score = 1 - sum(discriminator\_scores)$$

implying that if discriminators perform well, generator will lose more scores, vice versa.

Organic Duel

The outputs of 2 generator miners will be evaluated against each other by the discriminator miners.

1

All discriminators will gain a constant score of

$$discriminator\_score = 1 / n$$

where n equals to the number of total discriminators that participated in the task.

2

Each generator will gain

$$generator\_score_i = v_i * 1 / n$$

where v equals to the number of votes casted by the discriminators and n equals to the number of total discriminators that participated in the task.

Trap Round

One generator will be chosen randomly as the negative generator to produce an objectively inferior output. This round is solely focused on penalties, not rewards.

1

For discriminators that voted for the negative generator

$$discriminator\_score = -1$$

This is to make sure that discriminators are always guided to vote for the superior output instead of colluding with other miners or only voting for validator's output regardless of performance.

2

For generators, and discriminators that voted against the negative generator

They neither gain any score, nor incur any penalty.

Aggregate Reward Calculation

1

The scores are summed across all expired tasks within a rolling window

$$miner\_score = \sum_{i=1}^{n} (generator\_score_{i} + discriminator\_score_{i})$$

where n equals to each expired task within a rolling window

2

The scores is normalized and converted into weights to be set on Bittensor chain for incentives distribution

3

All miners scores will be re-initialized to 0 for the next weight setting round

Incentives Landscape

The zero-sum scoring in Synthetic Tasks, combined with the penalty-based Trap Rounds, creates a competitive and self-regulating incentives model. This system establishes a continuous "arms race" where:

• Generators are constantly pushed to improve their outputs to be chosen as superior and earn points.

• Discriminators are regulated and incentivized to choose the superior output to gain scores and avoid penalties.

This ensures that honest, high-performing miners are rewarded and can rise through the ranks, while malicious or colluding miners face significant economic consequences, upholding the long-term integrity and health of the subnet.

Security and Anti-Gaming Measures

Miner-miner Collusion Prevention

The zero-sum nature of the scoring mechanism, makes it economically infeasible for multiple colluding entities to dominate the metagraph without producing high quality outputs. Incoming honest and high quality miners can always displace and erode colluding entities when some of their miners have to forfeit their scores to keep other colluding miners afloat.

Randomization

Miners do not know in advance regarding the task nature, and augmentations may be introduced in any point of the task orchestration pipeline. This means that miners should always be honest in order to avoid negative consequences.

Trap Rounds

The unpredictable nature of the Trap Round serves as a powerful deterrent. It ensures that any miner who blindly votes for a colluding peer will inevitably be caught and penalized, leading to a loss of their reputation and stake. Not only so, it also penalizes miners who optimize to always vote for validators' baseline output in order to steal scores from generators.