Null consensus

Null is an owner-selected alternative to Yuma inside the existing epoch interface. Existing and newly created subnets default to Yuma. Selecting Null does not change stake calculation, reward destinations, collateral settlement or owner cuts. The root subnet keeps its separate basket/dividend mechanism and cannot select Null.

Rewards

Null elects exactly one validator permit holder: the registered UID with the highest unrounded stake weight. Equal stakes select the first UID. Only this permit holder may submit weights, commit weights or reveal weights; owner and self-weight exceptions do not bypass this rule. Permits are recomputed at epochs and initialized when switching to Null. All positive staked UIDs receive proportional dividends even without a permit, including inactive UIDs.

The winner's weights undergo the existing self-weight, registration and commit-reveal masks. Its remaining integer ratios determine miner rewards. An empty or all-zero valid row distributes miner rewards equally across every registered UID. Null skips consensus and bond calculations and does not write bond storage.

With positive stake, half the epoch budget goes to validator dividends, proportional to all positive stake. The miner budget receives the other half, including the extra atomic unit when the epoch budget is odd. With no positive stake, the local server and validator budget goes to miners. Unassignable root rewards are recycled separately and are never added to that miner budget. Wide integer arithmetic allocates each budget in UID order using cumulative proportional shares. Each payout is the floor or ceiling of its exact share; their sum equals the budget. A zero-weight UID receives zero. No fractional entitlement carries forward. Reported u16 incentive and dividend fractions remain quantized and do not feed back into actual reward amounts.

Owner configuration

btcli hparams set --netuid 1 --name epoch_consensus --value Null
btcli hparams set --netuid 1 --name max_allowed_uids --value 2500

Capacity is floor(2500 / mechanism_count) for Null, and floor(256 / mechanism_count) for Yuma. Increasing the mechanism count also checks the configured capacity. Reduce capacity first if the shared limit would be exceeded. Switching to Null does not automatically increase capacity.

Returning to Yuma requires the actual registered population to fit the Yuma ceiling. Use bounded owner pruning to shrink explicitly before switching. Switching never silently deletes participants; once allowed, it also lowers configured capacity to the Yuma ceiling. The minimum allowed UIDs must fit the selected ceiling too. Owner authentication, admin windows and hyperparameter rate limits apply. Entering Null saves the Yuma validator limit. Returning restores that limit, clamped to the remaining UID capacity; setting Null again does not overwrite it.

btcli hparams trim-null-batch --netuid 1 --target 256

For a long-running return to Yuma, use the opt-in workflow:

btcli hparams trim-null --netuid 1 --switch-to-yuma

This derives the target as floor(256 / mechanism_count), or resumes an existing on-chain pruning target. --target selects a smaller final population within the subnet's minimum and maximum. A conflicting pending target is rejected. Pruning deregisters participants and can change survivor UIDs; review the target before signing. Each batch uses the normal wallet, fee review, confirmation and receipt flow. Pass --yes to approve sequential submissions without individual prompts. Without --switch-to-yuma, the workflow only prunes and leaves consensus as Null.

The workflow waits for each transaction to finalize and reads NullPruningTarget and SubnetworkN after every batch. It reports removed UIDs and cancelled commits, including cleanup-only batches. It polls closed administration windows every 12 seconds, with a one-hour wait limit; --poll-seconds and --wait-timeout adjust these bounds. Initial trimming cooldown failures stop with an explanation rather than repeatedly submitting paid failures. Successful same-target continuations retain the runtime's cooldown exemption. --max-batches bounds the run to 256 batches by default; repeated lack of progress also stops the loop. --dry-run previews one transaction without starting the loop's submissions.

Ctrl-C stops waiting or further submissions. Rerun the command with the same target to resume from on-chain state. A recovery record beside the CLI config prevents another local workflow from silently resubmitting a transaction whose delivery is uncertain. --state-file selects its location. If interruption or a connection failure leaves an unresolved record, reconcile the transaction's inclusion or expiry on-chain before removing that record and restarting. Do not delete it merely because the population has not changed: a transaction may still be pending or may only cancel commits. Run one workflow per subnet, including across machines, and retain the same recovery-file location when restarting.

With --switch-to-yuma, remaining timelocked commitments are cancelled through bounded batches before the switch; the workflow also waits for the consensus cooldown and administration window. It verifies the mode after the switch receipt. Ownership changes, mechanism-count changes, permanent transaction failures and incomplete multisig approvals stop the run so the owner can resolve them explicitly.

For the single-batch command, repeat with the same target until NullPruningTarget is absent and SubnetworkN is at most the target. Each transaction deletes at most 64 UIDs. Pending weight commits are cancelled before survivor UIDs move; a large legacy commit backlog can require cleanup-only transactions. Continuations retain owner authorization and admin-window checks, while sharing the first transaction's trimming cooldown. Changing the target starts a new operation. The legacy atomic trimming call still works for Yuma; in Null it cannot delete more than one batch. A target equal to the current population can cancel queued commits without deleting participants.

Mode changes require pending timelocked commits to drain. Mechanism-count changes in Null also require legacy hash commit queues to be empty. If necessary, disable new commit-reveal submissions and wait for pending commits to reveal or expire, or cancel them through bounded pruning before changing the mechanism count. Frozen bonds retain the last Yuma epoch's registration cutoff, so replaced miners cannot inherit old bonds. The mechanism-count call charges for checking pending queues and removing historical state when the mechanism count is reduced.

Registration without TAO

Owners can opt into PoW registration for Null or Yuma subnets:

btcli hparams set --netuid 1 --name network_pow_registration_allowed --value true
btcli subnets register --netuid 1 --pow --yes

PoW and burn have independent demand: PoW admissions increase only difficulty, and burn admissions increase only burn cost. Both use burn_increase_mult and burn_half_life for their multiplier and decay. PoW keeps a separate team-controlled min_difficulty floor and its own maximum.

Owners can toggle burned registration with registration_allowed independently of PoW. At least one route must stay enabled on a non-root subnet; enable the replacement route before disabling the other. Direct zero-tip PoW submission pays no registration burn, initial collateral purchase or transaction fee. The CLI pulls the latest block challenge every 12 seconds while mining. The chain accepts proofs from the five preceding blocks, leaving time for signing and propagation. Only public challenge data enters the Rust GPU or CPU workers; signing stays in the usual coldkey workflow. Automatic mining uses all discovered OpenCL GPUs, with CPU fallback. Use --pow-backend gpu to require GPU mining, repeat --pow-device to select devices, and use --pow-workers and --pow-timeout to bound mining resources. Work does not bypass pruning, immunity, ownership or per-block registration limits. Root keeps its separate registration path.

Exact integer weights

Null stores submitted u16 values without max-upscaling. Values are relative integers in 0..65535 and need not sum to 65535. Choose ratios such as 65534:1 to retain small shares. A full row can target thousands of participants; payouts below one atomic unit still depend on the epoch budget and deterministic remainder allocation.

from bittensor.intents import SetWeights

intent = SetWeights(netuid=1, weights={2: 65534, 3: 1}, raw_u16=True)
# Execute with the normal client and validator hotkey wallet.

For large rows, write a JSON object mapping UID to integer weight:

btcli misc weights set --netuid 1 --raw-u16 --weights-file weights.json
btcli misc weights commit --netuid 1 --raw-u16 --weights-file weights.json

The raw path rejects fractional, negative, nonfinite and out-of-range values. It checks the subnet's maximum weight fraction without clipping the submitted ratios. Minimum count, unique valid destinations, stake, permits, rate and version checks still apply. Raw mode requires Null. The existing relative-float path remains available and keeps its scaling and clipping behavior.

set automatically selects plaintext or timelocked submission according to subnet configuration. Timelocked rows auto-reveal using drand; they need no manual reveal. Null allows ciphertexts up to floor(32 KiB / mechanism_count), leaving room above a full row at the corresponding UID ceiling. The shared queue per subnet epoch is bounded to 64 KiB and 64 commits across all mechanisms. Each hotkey can have one pending row per mechanism; resubmitting replaces that row. When full, a higher-priority eligible validator can displace lower-priority rows. Priority uses stake with ties resolved by first UID; ineligible hotkeys rank last. The leading validator can admit a full row for every mechanism even if junk commits previously filled the queue. Other commits may receive CommitQueueFull. Yuma keeps its 5,000-byte ciphertext limit. These limits include encryption and serialization overhead.

Operational cost

Null incentive and payout allocation is linear in registered UIDs. The epoch loads only the elected validator's row, containing at most one entry per UID. Historical nonwinning rows are not read or decoded. Weight reads, masking and working memory therefore stay linear even if old dense rows remain in storage. Null skips quadratic bond updates and writes the ordinary output vectors and payouts. For an existing staking pool owned entirely by the reward recipient, owner rewards increase the pool balance without rewriting ownership shares. Shared pools and first deposits keep the ordinary deposit accounting, and collateral capture and release still apply. The shared interface still uses ordered account maps, so this does not claim that every operation in the surrounding epoch pipeline is strictly linear. Delegation processing also scales with incoming parent edges, whose count is currently unbounded; a population ceiling alone does not bound that workload. When a Null epoch is due, the existing scheduler runs at most one epoch in that block. Remaining due subnets are deferred; blocks with only Yuma epochs retain the configured epoch cap. Payouts settle atomically in the epoch's block.

Solidity callers

The Subnet precompile at 0x0803 adds setEpochConsensus(uint16,uint8) (0 is Yuma, 1 is Null), getEpochConsensus(uint16), and getNullConsensusLimits(). The setter dispatches with the mapped caller's signed origin and enforces the same owner, rate, window and population checks as the normal extrinsic.

The Neuron precompile at 0x0804 adds setMechanismWeightsV2 for rows up to 2,500 entries and commitTimelockedMechanismWeightsV2 for ciphertexts up to 32 KiB. The runtime still applies the selected subnet mode's limits. The released setMechanismWeights and timelocked selectors preserve their 4,096-entry and 5,000-byte input bounds. Existing selectors, types and addresses are unchanged. The SDK's vendored ABIs include the new functions.

trimNullUidsBatch(uint16,uint16) performs one owner-authorized pruning batch. getNullPruningState(uint16) returns (active, target, remaining) with a zero target when inactive. getNullPruningBatchSize() reads the current runtime's deletion bound. getSavedYumaValidatorLimit(uint16) returns (present, limit) with a zero limit when absent. These are additive selectors at 0x0803.

New PoW hotkey associations keep each coldkey’s ownership and staking indexes below the existing 256-entry staking work limit. An already-associated hotkey can register on another subnet without growing those indexes.

The synchronous SDK exposes the same public challenge miner:

import bittensor as bt

# wallet is the existing local wallet; mining needs only its public addresses.
with bt.Subtensor("test") as sub:
    proof = sub.mine_pow_registration(
        netuid=1,
        hotkey_ss58=wallet.hotkey.ss58_address,
        coldkey_ss58=wallet.coldkeypub.ss58_address,
        workers=4,
    )
    result = sub.execute(proof, wallet)
    result.raise_for_failure()

For async clients, await client.mine_pow_registration(...) and then client.execute(...). Submit a returned proof promptly. The CLI refreshes an expired proof after confirmation and key unlock while keeping the selected subnet, hotkey and mining backend unchanged. Confirmed pre-inclusion expiry may trigger fresh mining, with at most three submission attempts; included transactions are never replayed by this recovery path.