Getting in and getting out: registration, immunity, deregistration
This lesson is about the two edges of taking part: what it costs to claim a slot, and what can take that slot away again — for one participant, and for a whole contest.
Joining a subnet as a miner or validator means paying that contest's current registration price out of your own funds. On most subnets the whole payment is simply burned — a sunk, non-refundable entry fee — but some instead lock a share of it as collateral attached to your key, money you earn back over time by doing the work rather than losing outright. [source] [source] [source]
That price is not fixed, and there is no scheduled window to register in — it is always open. The cost decays a little every block and jumps after each successful registration, so a rush of newcomers gets expensive fast for whoever arrives late. [source]
Every subnet has a hard limit on how many registered slots it has — 256 by default, and an owner can shrink that floor down to 64. Once it is full, each new registration doesn't add a slot; it costs whoever currently holds the weakest one theirs. [source]
A freshly registered key gets a grace period called immunity, and it lasts by default about half a day of blocks — plenty of contests set it much longer. Eviction targets whoever is earning the least, and the chain doesn't distinguish miners from validators when deciding who that is: whoever earns least, on either side, goes first. Immunity is only a preference, though, not a guarantee — if every registered key on a full subnet is still within its grace period, the chain evicts the weakest one anyway. [source] [source]
Losing a slot this way isn't a punishment beyond the slot itself. Whatever you staked stays exactly where it was, and on a contest that locked part of your registration price as collateral, that collateral stays attached to your key and is credited back to you if you register again. [source]
That's the story for one participant's slot. A whole subnet can lose its place too, though not for going quiet — there's no inactivity-based removal. The network itself only has room for a limited number of contests at once, and registering a new one, once that limit is reached, prunes whichever existing contest is trading at the lowest price and reuses its slot; a fresh contest is protected from that fate for roughly its first six months. Losing that protection and then losing the slot takes every one of its participants down too, deleting their standing along with it. [source] [source]
How it actually works
The eviction rules, more precisely: pruning score is purely emission-based, so a low-earning validator is just as evictable as a low-earning miner. Immunity is defined as (current_block − registered_at) < immunity_period, with a chain default of 4096 blocks (about 13.7 hours at 12-second blocks) — read the real value per subnet before planning around it, since it's frequently set much higher. Ties on pruning score evict the older registration first, and a subnet owner's own hotkey is permanently immune, bounded by an owner-immune limit that defaults to a single hotkey. [source]
One hotkey can hold a UID on many different subnets at once, but never more than one UID on the same subnet — confirm the assignment with a direct chain read once registration lands. An owner can also shrink a subnet's own capacity directly, via trim-subnet, rather than waiting for it to fill up naturally: the change is rate-limited to once per roughly 30 days, victims are always the lowest-emission UIDs, and both temporarily immune and permanently owner-immune UIDs are skipped — though immune UIDs together can never exceed 80% of the new, smaller maximum. [source] [source] [source]
At the subnet level the same "immunity is a preference" idea runs backward: if every existing subnet is still within its own immunity window when the registration limit is full, the new registration simply fails, rather than force-evicting an immune one. This price-ranked pruning rule has changed hands before — it existed pre-dTAO ranked by emission instead of price, was removed at dTAO's February 2025 launch, and returned in its current price-ranked form in September 2025. When a subnet does dissolve, its TAO reserve is paid out pro-rata over all its alpha directly to holders' balances — but two catches apply: alpha the protocol itself holds counts in that calculation and its share simply returns to the chain, diluting everyone else's payout, and any accumulated conviction from locked stake goes uncompensated even though the underlying stake is paid out like anyone else's. [source] [source] [source]