447Release

Conviction Normalization

What shipped

This release changes one rule: who counts as holding enough conviction to take ownership of a subnet. The gate now measures one hotkey alone. Previously the threshold was compared against the sum of all conviction on the subnet, while the winner was simply the hotkey with the most — so anyone's locks, including the owner's own defensive locks, counted toward a challenger's quorum. Now the winning hotkey must clear the bar with its own conviction only. [source] The threshold itself rises from 10% to 18%. [source]

The source is the project's own release note for this upgrade, dated August 2026 and billed as "The 18% single-hotkey ownership gate". [source] It is written as a correction to the release before it. The note says the v446 upgrade changed how the conviction ownership gate is calculated, basing the threshold on eligible alpha — outstanding alpha minus protocol-owned and burned alpha — instead of outstanding alpha; the accounting is more faithful, but the denominator shrank, so the gate fell, and the TAO a challenger needed to take over a subnet dropped on most networks, and on some dropped a lot. [source] That same change is described from the other side on the v446 page.

The release states its own purpose plainly: to bring the expected TAO cost of a takeover back to — and above — its pre-446 level, by turning exactly two parameters. [source] The rule that results is that ownership transfers when, on a subnet at least one year old, the highest-conviction hotkey's own conviction exceeds 18% of eligible alpha [source] — where eligible alpha is SubnetAlphaOut - SubnetProtocolAlpha - AlphaBurned. [source]

What it changes for the ecosystem

18% is not a new number on this network. The subnet owner cut — the share of a subnet's emissions paid to its owner — is 18%, and the release's argument for the choice is that the bar to claim ownership should match the reward of holding it: to earn the owner's 18% of emissions, you must first hold 18% of the subnet's eligible alpha as matured conviction on a single hotkey. One number then describes both sides of the trade, the note argues, instead of a 10% figure with no economic anchor. [source] That owner's cut is the same 18% lesson 04 traced through a payout period.

For subnet owners the symmetry produces a defence rule the release states outright. To win, a challenger must clear 18% and out-hold the owner's own conviction; an owner's locks mature instantly while a challenger's take about 43 days, so an owner who watches their subnet can always lock enough to stay ahead before a challenge matures. [source] Its instruction to owners who want a standing defence is exactly that: keep more conviction locked on the owner hotkey than any challenger holds, and top up at any point before a challenge matures. [source]

For a challenger, coalitions still work the natural way: backers who want to support a challenger lock toward the challenger's hotkey, and their conviction lands in that hotkey's aggregate. What no longer works is winning with a small position because unrelated lockers happened to push the subnet-wide total over the line. [source] Lesson 10 made the same point about judging power: what decides an outcome here is what one holder holds, not how many holders agree.

The release prices the change rather than only describing it. It quotes a takeover of every mainnet subnet that has passed the one-year age gate — 88 subnets — under three rule sets: pre-446, v446 as it runs today, and this upgrade. Each cost is presented as a real AMM quote: the TAO needed to buy enough alpha from the pool, with fees and slippage, so that one fresh hotkey both beats the current conviction leader and clears the gate. [source] The note also plots the same costs per subnet on a log scale, sorted cheapest to dearest, and pins that chart to Finney block 8,844,344. [source] The columns are the rule set, the median cost, the count of subnets takeable for under 1,000 τ, and the count where the required position is not fillable in one AMM buy: [source] [source] [source] [source]

The release's own reading of those rows is that the v446 curve sits below the pre-446 curve across the board — the accidental lowering this upgrade corrects — and that the new curve sits above both: the median cost roughly doubles against pre-446, the number of subnets takeable for under 1,000 TAO falls from 23 to 5, and on 21 subnets the required 18% position cannot be assembled from the pool at all, because buying that much alpha in one market exhausts what the AMM will sell at any price. [source]

Nothing about existing locks or conviction resets, and no takeover becomes easier — the note's claim is that every change in the upgrade raises the bar. For a subnet to change hands, all of the following must now hold at once: [source]

Anyone running a takeover dashboard has to recompute it. The gate is per-hotkey conviction against 18% of eligible alpha; subnet-wide totals no longer gate anything. [source] The subnet-convictions read now projects each hotkey against the 18% bar directly. [source]

What to watch

The release's own account of the immediate effect is that conviction becomes a slower instrument. On every mainnet subnet past the age gate at the time it was written, it says, the conviction leader is the owner's own hotkey — no challenger leads anywhere, let alone above 18%. A challenge that starts now must buy against the pool and then wait out maturity in full view. Conviction remains the path by which a committed, majority-scale backer can eventually earn a neglected subnet, which the note says is by design, but it is no longer a lever a small, well-timed position can pull. [source]

Unverified The note states no deployment status for this release. Its comparison table labels one rule set "v446 today" and this one "This upgrade", which reads as though the v446 rule was the live one when the note was written — but that is an inference from a column header, not a statement, and it says nothing about whether spec 447 has since been deployed. Which rule the chain is actually applying has to be read from chain state, not from this page.
Unverified Every figure in the comparison is the project's own AMM quote at a single block, not an independent measurement, and pool depth is exactly the quantity those quotes are most sensitive to. The 21 subnets where an 18% position cannot be assembled at any price are the release's strongest number and also its most fragile: that is a statement about how thin those pools were at one block, and a deeper pool would move a subnet out of that column without any rule changing. It is also worth naming what the column means — on those subnets the ownership rule is not a high bar but a closed door for as long as the pool stays that thin. The release presents the count as evidence of safety; whether an unreachable gate is the same thing as a safe one is not a question this document asks.
How it actually works

The gate as it now computes

The gate measures one hotkey alone. Previously the threshold was compared against the sum of all conviction on the subnet, while the winner was simply the hotkey with the most — so anyone's locks, including the owner's own defensive locks, counted toward a challenger's quorum. Now the winning hotkey must clear the bar with its own conviction only. [source] The threshold rises from 10% to 18%. [source] The denominator is the one v446 introduced: [source] [source]

eligible alpha = SubnetAlphaOut - SubnetProtocolAlpha - AlphaBurned

Which gives the transfer condition: [source]

ownership transfers when, on a subnet at least one year old,
the highest-conviction hotkey's OWN conviction > 18% × eligible alpha

Coalitions still resolve through a single hotkey: backers who want to support a challenger lock toward the challenger's hotkey, and their conviction lands in that hotkey's aggregate. What stops working is winning with a small position because unrelated lockers pushed the subnet-wide total over the line. [source]

The defence asymmetry

A challenger's conviction approaches its locked mass on the MaturityRate timescale — about 43 days to most of full weight — and the position is visible on chain through subnet-convictions for that entire period. [source] An owner's locks mature instantly, so out-locking a challenger at any point before its position matures ends the attempt. [source] [source]

What was measured

The release priced a takeover of every mainnet subnet past the one-year age gate — 88 subnets — under pre-446, v446 as it runs today, and this upgrade, quoting for each the TAO needed to buy enough alpha from the pool, with fees and slippage, so that one fresh hotkey both beats the current conviction leader and clears the gate. [source] The accompanying chart is a log-scale plot of cost per subnet under each rule set, sorted cheapest to dearest, at Finney block 8,844,344. [source] Its stated result: the median cost roughly doubles against pre-446, subnets takeable under 1,000 TAO fall from 23 to 5, and on 21 subnets the 18% position cannot be assembled from the pool at any price. [source]

For indexers and dashboards

The gate to compute is per-hotkey conviction against 18% of eligible alpha; subnet-wide totals no longer gate anything. [source] The subnet-convictions read projects each hotkey against the 18% bar directly, so a dashboard can take the projection rather than recomputing it. [source] [source]