Subnet architecture

High-level shape

Like every Bittensor subnet, Tensorbox is a closed loop:

operators ──▶ miners ──▶ outputs ──▶ validators ──▶ scores ──▶ emissions ──▶ operators
                  ▲                                                          │
                  └──────────── feedback signal ─────────────────────────────┘

The interesting bits are:

• The task. What miners are actually being asked to produce. (See the task spec once published.)
• The scoring function. How validators grade miner outputs. This is the contract — change it and you change the entire economy.
• The reward distribution. How TAO emissions flow back to participants based on validator scores.

Hard constraints

This table will be filled in and locked once we finalize numbers. We'd rather ship a TBD than a number we'll have to retract.

Why these choices

The hardest design question in a subnet isn't "what should miners produce." It's "what should validators measure." Get the second wrong and you get a market that optimizes for the wrong thing.

We picked the current scoring function because:

1. It's falsifiable. Two validators with the same input should get the same output. No fuzzy human judgment in the inner loop.
2. It's cheap to compute. Validators run on commodity hardware; we won't accidentally exclude small operators by needing a GPU farm to validate.
3. It's robust to gaming. We tried to break it before shipping it. Open issue if you find a hole.

Open questions

• How aggressively do we deweight stale outputs? Currently linear decay; experimenting with step functions.
• Should validators see each other's scores? Currently no; this prevents trivial copying but also slows convergence.

These are the kinds of things that move between releases. The changelog tracks them.