Taoflow2 by ppolewicz · Pull Request #2385 · opentensor/subtensor

ppolewicz · 2026-01-31T16:28:45Z

Description

This PR introduces a system to control which subnets receive TAO emission. Previously, all eligible subnets received emission proportional to their "share." Now, three optional filters can reduce or zero-out emission for lower-performing subnets.

New metric: EffectiveRootProp

Each subnet distributes dividends to two groups: alpha stakers (staked directly on subnet) and root stakers (staked on root network). EffectiveRootProp measures what fraction flows to root stakers:

EffectiveRootProp[netuid] = root_dividends / (alpha_dividends + root_dividends)

This is computed during each epoch and stored per-subnet.

Filters

Three emission filters are applied sequentially in get_subnet_block_emissions():

EffectiveRootProp scaling (disabled by default): Multiplies each subnet's share by its EffectiveRootProp, rewarding subnets with more root staker participation.
Top subnet proportion: Keeps only the top X% of subnets by share (default 100% = all subnets, intended mainnet value: 50%).
Absolute subnet limit: Hard cap on number of subnets receiving emission (default 0 = disabled, maybe we won't use it at all, but proportion is tricky and it's better to have both tool at our disposal).

After each filter, shares are re-normalized to sum to 1.0. Subnets that get filtered out receive zero emission.

But why?

We had a couple of problems since Taoflow emission schedule was deployed:

New subnets get very little injection of Tao when they desperately need it
Old subnets which have a lot of liquidity get even more
Since dtao subnets are generally incentivized to fight root prop (see sn89 using tricks to reduce root prop artificially) to reduce sell pressure
sn104 managed to cut off root validators from receiving any dividends at all by using the time gap between CHK application and immediate effectiveness of btcli stake move coupled with randomly chosen "miner" (they are self-mining, the subnet doesn't really do anything). We think they can safely inflow to increase emissions and we want to mitigate this risk.
sometimes subnet owners say "oh, the validator code doesn't work for you? Then CHK stake to me, I don't want to debug your environment problems" and it is unacceptable, so it'd be good to incentivize them to care that the validator code actually works

It looks like this can be fixed with EffectiveRootProp. The filters were added because they were requested and the change was on the way.

Type of Change

New feature (non-breaking change which adds functionality)

Checklist

I have performed a self-review of my own code
I have commented my code, particularly in hard-to-understand areas
I have run ./scripts/fix_rust.sh to ensure my code is formatted and linted correctly
I have made corresponding changes to the documentation
My changes generate no new warnings
I have added tests that prove my fix is effective or that my feature works
New and existing unit tests pass locally with my changes
Any dependent changes have been merged and published in downstream modules

Original prompt

(events were later deemed to be useless and were removed in a subsequent prompt; default was changed to 100%)

Use Plan subagent to write a step-by-step detailed development plan, which then after a final confirmation from me you'll go implement in the main agent by following the task list until it is completed. Use a fresh sonnet-based subagent for every subfeature, but do not run agents in parallel or you'll run out of memory and disk space. Work on ⁨taoflow2⁩ branch which starts off ⁨origin/devnet-ready⁩. Run ⁨scripts/fix_rust.sh⁩ after every commit. Remember about checking in tests that events indicating key information are emitted. Rather than long methods use functions that have descriptive names and that are easy to test.

The changes we do today will focus on ⁨get_subnet_block_emissions()⁩ in ⁨src/coinbase/subnet_emissions.rs⁩.

Add ⁨EffectiveRootPropEmissionScaling⁩ root hyperparameter which will allow root to enable multiplication of shares by ⁨EffectiveRootProp⁩, which for a given netuid is ⁨sum(RootAlphaDividendsPerSubnet[netuid]) / (sum(AlphaDividendsPerSubnet[netuid]) + sum(RootAlphaDividendsPerSubnet[netuid])⁩ (document it as such), though it should be calculated during epoch near where ⁨RootAlphaDividendsPerSubnet⁩ is being written and stored in the global vector, so that during emission preparation reading ⁨EffectiveRootProp⁩ requires a single db get for all subnets.

Add root hyperparameter to ensure only a certain proportion (by default 50%) of subnets get any emission, while the rest is redistributed to other subnets (so shares must be zeroed except topk and normalized). Round up so that a single subnet counts as in top 50%.

Add another hyperparameter to limit the number of subnets which receive emission to an absolute number (there is no default and this means by default there will be no absolute limit).

AI reviewers

Claude 4.5 Opus
Codex
Devin

Add EffectiveRootProp StorageMap (NetUid -> U64F64) computed during distribute_dividends_and_incentives() as: sum(RootAlphaDividendsPerSubnet[netuid]) / (sum(AlphaDividendsPerSubnet[netuid]) + sum(RootAlphaDividendsPerSubnet[netuid])) This measures the proportion of dividends on a subnet flowing to root stakers, stored per-subnet for efficient single-read access during emission preparation. Co-Authored-By: Claude Opus 4.5 <[email protected]>

When enabled, multiplies each subnet's emission share by its stored EffectiveRootProp value and re-normalizes. This allows root governance to weight emission toward subnets with higher root staker participation. - Storage: EffectiveRootPropEmissionScaling (bool, default false) - Admin extrinsic: sudo_set_effective_root_prop_emission_scaling (call_index 88) - Helper: normalize_shares(), apply_effective_root_prop_scaling() - Event: EffectiveRootPropEmissionScalingApplied Co-Authored-By: Claude Opus 4.5 <[email protected]>

Only the top proportion of subnets (ranked by emission share) receive emission. Default is 50% (5000 basis points). Remaining subnets have shares zeroed and redistributed to top subnets. Uses ceil(count * proportion / 10000) so a single subnet always qualifies (ceil(1 * 0.5) = 1). - Storage: EmissionTopSubnetProportion (u16, default 5000) - Admin extrinsic: sudo_set_emission_top_subnet_proportion (call_index 89) - Helpers: zero_and_redistribute_bottom_shares(), apply_top_subnet_proportion_filter() - Event: EmissionTopSubnetFilterApplied Co-Authored-By: Claude Opus 4.5 <[email protected]>

Adds absolute cap on the number of subnets that can receive emission. Default is 0 (disabled). When set to N > 0, only the top N subnets by share receive emission; the rest are zeroed and redistributed. Applied after proportion filter so the stricter constraint wins. - Storage: EmissionTopSubnetAbsoluteLimit (u16, default 0) - Admin extrinsic: sudo_set_emission_top_subnet_absolute_limit (call_index 90) - Helper: apply_top_subnet_absolute_limit() - Event: EmissionAbsoluteLimitApplied - Fixed 5 existing coinbase tests to disable proportion filter Co-Authored-By: Claude Opus 4.5 <[email protected]>

- Replace slice indexing `sorted[top_k..]` with `.get(top_k..)` to avoid potential panic (clippy::indexing_slicing) - Replace `(total as u64) * (proportion as u64)` with `.saturating_mul()` to avoid arithmetic side effects (clippy::arithmetic_side_effects) Co-Authored-By: Claude Opus 4.5 <[email protected]>

…tests Remove workaround EmissionTopSubnetProportion::<Test>::set(10000) from existing coinbase tests that was needed when default was 50%. Co-Authored-By: Claude Opus 4.5 <[email protected]>

…efault Remove EffectiveRootPropEmissionScalingApplied, EmissionTopSubnetFilterApplied, and EmissionAbsoluteLimitApplied events and their test assertions. Update proportion filter tests to explicitly set 5000 (50%) since default is now 100%. Co-Authored-By: Claude Opus 4.5 <[email protected]>

When subnets tie at the k-th position, all tied subnets are now kept rather than arbitrarily zeroing some. Uses a threshold-based approach: find the share value at position top_k-1 and keep all entries >= that value, avoiding deterministic bias toward lower NetUIDs. Co-Authored-By: Claude Opus 4.5 <[email protected]>

Use min(EffectiveRootProp, RootProp) when scaling emission shares. This prevents a subnet from inflating its EffectiveRootProp above the configured RootProp by disabling alpha validators, which would cause all dividends to flow to root and artificially boost the scaling factor. Co-Authored-By: Claude Opus 4.5 <[email protected]>

shamil-gadelshin

The tests don't cover the whole run_coinbase. There are some concerns about the calculation order and other minor issues.

shamil-gadelshin · 2026-02-03T10:33:08Z

pallets/admin-utils/src/lib.rs

+        ) -> DispatchResult {
+            ensure_root(origin)?;
+            pallet_subtensor::Pallet::<T>::set_effective_root_prop_emission_scaling(enabled);
+            log::debug!("set_effective_root_prop_emission_scaling( {enabled:?} )");


This is noise. Please, remove unnecessary logs.

shamil-gadelshin · 2026-02-03T10:50:17Z

pallets/admin-utils/src/lib.rs

+        ) -> DispatchResult {
+            ensure_root(origin)?;
+            ensure!(
+                proportion > 0 && proportion <= 10000,


The proportion type is counterintuitive. Consider using U64F64 and limit "0-1" or "0-100".

shamil-gadelshin · 2026-02-03T10:52:25Z

pallets/subtensor/src/coinbase/run_coinbase.rs

+        let total = total_alpha_divs.saturating_add(total_root_divs);
+
+        let effective_root_prop = if total > zero {
+            U64F64::saturating_from_num(total_root_divs.checked_div(total).unwrap_or(zero))


Do you need these conversions between U64F64 and u96F32? Can we use a single format?

shamil-gadelshin · 2026-02-03T10:54:09Z

pallets/subtensor/src/coinbase/subnet_emissions.rs

+    /// zeros shares beyond the top `limit` subnets and re-normalizes.
+    pub(crate) fn apply_top_subnet_absolute_limit(shares: &mut BTreeMap<NetUid, U64F64>) {
+        let limit = EmissionTopSubnetAbsoluteLimit::<T>::get();
+        if limit == 0 {


Please, use rust idiomatic None for this case.

shamil-gadelshin · 2026-02-03T11:03:33Z

pallets/subtensor/src/coinbase/subnet_emissions.rs

+        top_k: usize,
+    ) {
+        let zero = U64F64::saturating_from_num(0);
+        if top_k == 0 || shares.is_empty() {


why do we iterate through the empty array?

shamil-gadelshin · 2026-02-03T11:16:36Z

pallets/subtensor/src/coinbase/subnet_emissions.rs

+            return; // 100% means all subnets get emission
+        }
+
+        let total = shares.len() as u32;


Incorrect type conversion - there is a different conversion below.

shamil-gadelshin · 2026-02-03T11:34:54Z

pallets/subtensor/src/coinbase/run_coinbase.rs

        root_alpha_dividends: BTreeMap<T::AccountId, U96F32>,
    ) {
+        // Compute and store EffectiveRootProp before distributing (uses raw dividend values).
+        Self::compute_and_store_effective_root_prop(


It seems we calculate the effective root prop for the next round. Is it intentional?

Also, we calculate this value once per epoch. Could this fact be exploited?

We could probably place a proper stake before we calculate the value and then remove it.

pallets/subtensor/src/coinbase/subnet_emissions.rs

shamil-gadelshin · 2026-02-03T11:54:28Z

pallets/subtensor/src/tests/subnet_emissions.rs

 //         assert_abs_diff_eq!(s1 + s2 + s3, 1.0, epsilon = 1e-9);
 //     });
 // }
+


There are no tests to verify that the new taoflow actually works as a whole.

- Remove log::debug! statements from admin extrinsics - Change EffectiveRootProp storage type from U64F64 to U96F32 to match context types and avoid unnecessary conversions in run_coinbase - Change EmissionTopSubnetProportion from u16 basis points (0-10000) to U64F64 fractional range (0.0-1.0) - Change EmissionTopSubnetAbsoluteLimit from u16 ValueQuery (0=disabled) to u16 OptionQuery (None=disabled) for idiomatic representation - Fix empty array iteration guard in zero_and_redistribute_bottom_shares - Add detailed comment explaining ERP timing (computed for next round) and exploitation mitigation via min(EffectiveRootProp, RootProp) - Update all tests for new storage types Co-Authored-By: Claude Opus 4.5 <[email protected]>

Three tests covering 2 subnets with 6 neurons each (owner, major/minor root validators, major/minor subnet validators, miner): 1. test_basic_all_validators_set_weights_to_miners (price=0.6) - root_sell_flag=true, root validators earn dividends - Verifies miner incentive, validator dividend proportionality, root validator earnings, owner cut (18%), incentive vector 2. test_no_root_sell_all_validators_set_weights_to_miners (price=0.5) - root_sell_flag=false, root validators earn 0 - Same structure but verifies root channel is unfunded 3. test_basic_major_root_no_weights (price=0.6) - Major root validator (5.55M TAO) does NOT set weights - Verifies major root earns 0 while minor root still earns Test structure: 5 blocks total (setup at block 1, epochs at blocks 3+5), SN1 only with tempo=1, TaoWeight=0.18, SubnetOwnerCut=18%. Co-Authored-By: Claude Opus 4.6 <[email protected]>

* wide-scope-dividend-test: (44 commits) Add wide_scope_dividend test suite for emission distribution Fix sudo_set_max_allowed_uids benchmark raise fee on `set_pending_childkey_cooldown` less sleep catch error with wrong nonce bump version add transaction replacement test case Fix try-runtime build bump version bump psdk/frontier with patch to prevent node panic Spec bump Spec bump auto-update benchmark weights bump frontier hash 2 by removing comment bump spec version bump Cargo toml frontier deps Bump spec version Add SubnetBuyback event format code test is ok ...

EffectiveRootProp now multiplies the raw dividend ratio by a utilization factor: active_root_stake / total_root_stake. This ensures subnets where most root stake is idle (validators not setting weights) get a much lower EffectiveRootProp than subnets where all root stake is active. Also enables EffectiveRootPropEmissionScaling in wide_scope_dividend tests and updates all assertions accordingly. Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Change compute_and_store_effective_root_prop to use dividend-efficiency metric instead of binary active/inactive. Returns U96F32 utilization. - Move ERP computation from distribute_dividends_and_incentives to distribute_emission; add utilization scaling (< 1.0) and hard cap (< 0.5 recycles all root dividends, sets ERP to 0). - Add 555k unstaked TAO to test setup_test(). - Add 3 new tests: unstaked_tao_does_not_affect_utilization, half_weights_to_validator, half_weights_no_minor_root. - Update existing test assertions for hard cap behavior. Co-Authored-By: Claude Opus 4.6 <[email protected]>

Skip hard cap and scaling when root_alpha_dividends is empty (e.g. root_sell_flag=false). Fixes tests where validators with root stake but no root dividend emission were incorrectly having their alpha dividends recycled. Co-Authored-By: Claude Opus 4.6 <[email protected]>

ppolewicz and others added 9 commits January 31, 2026 08:54

remove temporary plan.md file

6f1c024

fix: change default EmissionTopSubnetProportion to 100% and clean up …

e1b64c3

…tests Remove workaround EmissionTopSubnetProportion::<Test>::set(10000) from existing coinbase tests that was needed when default was 50%. Co-Authored-By: Claude Opus 4.5 <[email protected]>

ppolewicz marked this pull request as ready for review February 2, 2026 23:03

ppolewicz added the skip-cargo-audit This PR fails cargo audit but needs to be merged anyway label Feb 2, 2026

shamil-gadelshin requested changes Feb 3, 2026

View reviewed changes

ppolewicz force-pushed the taoflow2 branch from 46e4119 to 4699232 Compare February 5, 2026 16:09

ppolewicz and others added 5 commits February 6, 2026 21:39

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Taoflow2#2385

Taoflow2#2385
ppolewicz wants to merge 16 commits intodevnet-readyfrom
taoflow2

ppolewicz commented Jan 31, 2026 •

edited

Loading

Uh oh!

shamil-gadelshin left a comment

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

Uh oh!

shamil-gadelshin Feb 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

ppolewicz commented Jan 31, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

New metric: EffectiveRootProp

Filters

But why?

Type of Change

Checklist

Original prompt

AI reviewers

Uh oh!

shamil-gadelshin left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

ppolewicz commented Jan 31, 2026 •

edited

Loading