The Need for Constraints

To maximize high signal volume we must set up guardrails to control gaming the incentive mechanism

Let's assume that miners are building predictive models to inform their trades. Because only qualified flow contributes to model inference quality, the convex program must impose logical constraints on the allocations x_i . Without these guardrails, miners could attempt to inflate volume with losing trades, micro-trades, or reckless trading. Constraints form the backbone of any optimization problem. They protect the budget, prevent winner-take-all outcomes, and enforce that rewarded flow reflects real predictive signal rather than noise or manipulation.

For instance, we could reward the top miner who generates a high ROI in a winner-take-all scheme, but this almost certainly rewards overfitting and model drift, since it reduces to a non-convex combinatorial optimization where noise is mistaken for skill. On the other hand, we could spread rewards across miners in a flat, uniform fashion, but doing so destroys incentives because strong and weak signals would be treated alike. A poorly shaped rule also invites schemers to game the system. A miner could post a high ROI while contributing negligible volume, or flood volume while contributing little signal. For these reasons, the decision to allocate budget must be carefully balanced with constraints that protect against degenerate outcomes.

Constraints are like hard guardrails that stop a dishonest actor from gaming the system and limits the directions they can take to run off with emissions that did not provide valuable signal.

A textbook two-phase convex program fits the bill because it guarantees an efficiently obtainable global optimum and fractional alpha token allocations combined with smooth ramp constraints (no choppy changes in reward structure) promote smooth reward structure. This gives predictable, stable incentives that are very hard to game. The constraints in a convex program are also explicit and interpretable, allowing for a clean understanding of how the incentive mechanism performs via the that naturally fall out of the primal problem. We propose the following constraints: 1. Total Budget Constraint: We cannot distribute more alpha tokens than what Bittensor emits to us. 2. Payout Rate Constraint: We do not allocate to miners who produce no signal in an epoch. 3. Diversity Constraint: Each miner has a max allocation amount. 4. Ramp Constraint: Your allocation can only get bigger or smaller over time at a certain rate. Eligibility: To qualify your volume, you can't:

• Trade a tiny amount and expect to get a big allocation (min volume)

• Have a trailing negative ROI or tiny ROI like buying favorites near the end of a game (min ROI)

• Take hail mary trades that aren't skillful or informed (proof of statistical skill or machine intelligence)

Together these constraints complete a robust convex program, making it tough on dishonest actors who don't provide a prediction market with useful signal.