Root
SN0Not a contest at all. Root is where TAO holders stake for network-wide exposure, and it has no miners.
Core · ProtocolAn independent guide to TAO / October 2026
Bittensor is an open network that pays people in a cryptocurrency called TAO for producing useful AI services, and uses dozens of competing marketplaces called subnets to decide whose work is best.
Fetching live dataSource: CoinGecko public API. Figures are informational only.
01 / Bittensor in 60 seconds
Bitcoin pays computers to secure a ledger. Bittensor borrows that idea and points it at something else: it pays computers to do AI work, then lets an open market judge the results.
There is no single "Bittensor AI". The network is a blockchain that keeps score and hands out rewards. The actual work happens in subnets, each of which is its own contest with its own rules: one rents out GPUs, another serves AI models, another hunts for drug candidates, another detects deepfakes.
Everything on this page is educational. It is not a recommendation to buy, sell or stake anything.
02 / Origins
Bittensor's best-known founder is Jacob Steeves, an engineer who previously worked at Google and who is known online as "Const". He has said that studying Bitcoin's incentives in the mid-2010s led him to ask whether the same design could reward machine intelligence instead of hashing. In 2019 he was joined by Ala Shaabana, a computer scientist with a PhD from McMaster University, and the project was incubated with support from the crypto investment firm Polychain Capital.
The ideas were set out in a whitepaper titled "Bittensor: A Peer-to-Peer Intelligence Market". The paper is credited to Yuma Rao, a pseudonym. Nobody has publicly confirmed who Yuma Rao is, in the same way that Bitcoin's Satoshi Nakamoto has never been identified. The name lives on in Yuma Consensus, the algorithm at the heart of the network.
The founders set up the Opentensor Foundation to build and steward the software. The first version of the network, codenamed Kusanagi, went live in January 2021. It was halted a few months later to fix consensus problems and relaunched as Nakamoto in November 2021. In March 2023 the chain migrated to Finney, the network that still runs today.
Bittensor did not hold an ICO and did not set aside a block of tokens for founders or investors. Like Bitcoin, the only way to obtain TAO at the start was to earn it by taking part in the network. Early backers, including venture firms, got their holdings by mining and validating, or by buying from people who did.
That does not mean ownership is evenly spread. Early participants faced little competition and accumulated a great deal, which is one of the criticisms covered in the risks section.
A tensor is the multi-dimensional array of numbers that neural networks are built from. The name signals the ambition: Bitcoin's economic model, applied to machine learning. The token symbol TAO is written with the Greek letter tau (τ).
The codenames nod to Motoko Kusanagi of the anime Ghost in the Shell, Bitcoin's creator Satoshi Nakamoto, and Hal Finney, who received the first Bitcoin transaction.
03 / Timeline
Select a year. Use the left and right arrow keys to move between them.
04 / How it works
Bittensor has two layers. At the bottom is Subtensor, a proof-of-stake blockchain built with the Substrate framework (the same toolkit used by Polkadot). It produces a block roughly every 12 seconds, records who holds what, and runs the reward logic. Since late 2024 it has also supported Ethereum-style smart contracts through an EVM layer.
On top sit the subnets. The heavy AI computation never touches the chain. Miners run models and hardware off-chain; only the scores and the resulting payments are recorded on-chain.
Produce the commodity the subnet asks for: answers, compute, predictions, storage. Paid according to how well they score.
Test miners and submit a list of scores, called weights. Earn more when their judgments match the consensus.
Design the incentive mechanism: the task, the test and the scoring code. Receive 18% of their subnet's emissions.
Back validators with TAO or subnet tokens and share in the rewards. Their stake gives validators their voting weight.
Imagine a panel of judges at a competition. Each judge scores every contestant. To stop one biased judge from handing a friend the prize, the panel looks at where the stake-weighted majority of judges landed for each contestant and clips any score above that level. The contestant's final mark is built from the clipped scores.
That is Yuma Consensus. Validators are the judges, miners are the contestants, and a judge's voting power depends on the stake behind them. Validators who score in line with the eventual consensus build up "bonds" in the miners they spotted early and earn more. Validators who score wildly out of line earn less.
The process repeats every tempo, which is 360 blocks or about 72 minutes by default. At the end of each tempo the subnet's newly created tokens are divided: 18% to the subnet owner, 41% to miners and 41% to validators and the people staking with them.
What counts as "good work" is not decided by the blockchain. Each subnet owner writes that definition into the subnet's incentive mechanism. A well-designed mechanism rewards real quality; a poorly designed one gets gamed. Much of Bittensor's history is a story of subnets learning that lesson.
Press "Run a tempo". Six miners answer the task, three validators score them, and 100 tokens are split according to the consensus.
A teaching model, not the real algorithm. Try dragging your miner's quality up, or tick the collusion box and watch consensus clip validator C's inflated score and cut that validator's pay.
05 / Subnets
A subnet is a self-contained competition for one kind of digital commodity. Subnets were opened to outside teams in October 2023; before that the network ran a single task. Each subnet has a number (its netuid), an owner, a limited set of miner and validator slots, and since 2025 its own token.
Launching a subnet means paying a registration cost in TAO that rises and falls with demand. The number of slots is capped. For most of 2025 and 2026 that cap was 128, and when the network is full the weakest subnet by token price can be deregistered to make room, after an immunity period for newcomers. An increase toward 256 slots has been proposed and reported during 2026, though public directories still listed about 128 active subnets in October 2026.
Subnet 0, known as Root, is special. It has no miners. It is where TAO holders stake for network-wide exposure rather than betting on a single subnet.
Subnets are not permanent. Teams rebrand, sell, pivot or leave, and a netuid can be reused by an entirely different project. Treat the directory below as a snapshot from October 2026 and check a live explorer before relying on it.
Filter by category. Descriptions are written in plain English from public documentation and independent subnet directories. Inclusion is not an endorsement.
Not a contest at all. Root is where TAO holders stake for network-wide exposure, and it has no miners.
Core · ProtocolA serverless platform for running open AI models. Developers call a model through an API and miners' GPUs do the work. One of the largest subnets by usage.
Compute & inference · Built by Rayon LabsConfidential computing: AI workloads run inside hardware-secured enclaves so the machine's operator cannot see the data.
Compute & inference · Manifold LabsA marketplace for renting GPUs directly from their owners, with checks that the hardware is what it claims to be.
Compute & inference · Formerly CeliumConfidential AI services run on GPU clusters inside trusted execution environments.
Compute & inferenceSupplies compute jobs to Bittensor validators and other subnets, so they do not have to rent cloud hardware.
Compute & inferenceUses zero-knowledge proofs to show that an AI answer really came from the model it claims to have come from.
Compute & inference · Inference LabsAn open competition to produce better reasoning models, scored across reinforcement-learning environments.
Training · Associated with Jacob SteevesCollaborative pre-training of large language models, run as a king-of-the-hill contest. The slot was previously Templar, which left with Covenant AI in April 2026.
Training · New operator since 2026Trains language models across many ordinary internet-connected GPUs instead of one data centre.
Training · MacrocosmosFine-tuning as a service: miners compete to customise a model on your data with the best result.
Training · Rayon LabsA sandboxed benchmark that pushes miners to solve abstract reasoning puzzles (ARC-AGI-2).
Training · LatentA tournament for autonomous software-engineering agents, scored on real coding tasks.
Agents & codeThe oldest subnet. Now a rotating set of competitions where miners submit algorithms and agents.
Agents & code · MacrocosmosA public arena where AI shopping agents compete on e-commerce tasks.
Agents & codePays miners to collect fresh public data from social platforms into a large open dataset.
Data & search · MacrocosmosAn open marketplace for live web and social search results that AI applications can query.
Data & searchTurns raw text into structured, labelled data using competing language-model miners.
Data & searchComputer vision that turns video, starting with football matches, into structured analytics.
Media & visionDetects whether images and video are real or AI-generated. Miners compete to spot deepfakes.
Media & visionGenerates 3D assets on demand for games and augmented or virtual reality.
Media & visionMiners compete to upscale and compress video.
Media & visionA decentralised proprietary-trading contest. Miners submit trading signals and are scored on risk-adjusted returns.
Prediction & finance · TaoshiMiners forecast the full range of likely price paths for assets, not just a single number.
Prediction & financeSports predictions scored against real prediction-market positions. Formerly Sportstensor.
Prediction & financeA competition between forecasting agents that predict real-world events.
Prediction & finance · YumaCompeting machine-learning models that forecast weather and other environmental variables.
ScienceA drug-discovery contest: miners search huge chemical libraries for molecules likely to bind to a target.
Science · Metanova LabsCrowdsourced benchmarking of software pipelines that detect variants in DNA sequencing data.
ScienceDecentralised cloud storage compatible with the widely used S3 interface. Its token was the first subnet token on a centralised exchange.
Storage & infraAn open competition to improve vision-language-action models that control robots.
RoboticsAn incentive-driven benchmark for autonomous drone flight.
Robotics06 / Tokenomics
TAO copies Bitcoin's headline numbers on purpose: a hard cap of 21 million and a reward that halves over time. The detail of when it halves is different.
From launch, the chain created 1 TAO per block, or about 7,200 a day. In mid-December 2025 the first halving cut that to 0.5 TAO per block, about 3,600 a day.
Bitcoin halves every 210,000 blocks. Bittensor instead halves when total issuance crosses set thresholds: half the supply (10.5 million), then three quarters (15.75 million), and so on. If nothing else happened this would work out at roughly four years between halvings.
Something else does happen. Fees for registering miners and subnets are recycled: the TAO is handed back to the pool of tokens yet to be issued. Recycling lowers the issuance count, which pushes later halvings further into the future. So the dates of future halvings are estimates, not fixed appointments.
There was no allocation to a team, foundation or investors at genesis. Every TAO in existence was created as a block reward. New TAO does not go straight to miners any more; since 2025 it is injected into subnet liquidity pools, and participants are paid in subnet tokens. The next two sections explain that.
Progress toward 21M: waiting for live data
07 / Dynamic TAO
Until early 2025, a small group of large validators on the root network voted on how much TAO each subnet received. Critics said this concentrated power and invited favouritism. In February 2025 the network replaced it with Dynamic TAO, usually shortened to dTAO.
Under dTAO every subnet has its own token, generically called an alpha token. Each subnet also has a liquidity pool holding two things: TAO and that subnet's alpha. The ratio between them is the alpha token's price. There is no order book and no company setting prices; an automated market maker does it with a formula.
When you stake TAO into a subnet, you are swapping it for alpha through that pool. Your TAO goes in, alpha comes out, and the price of alpha moves up. Unstaking is the reverse. Because it is a swap, the value of your position rises and falls with the alpha price, and large trades move the price against you (slippage). Each alpha token has its own 21 million cap and its own halving curve.
The market now decides how new TAO is shared between subnets, but the formula has been rewritten more than once:
Expect this to keep changing. The direction has been consistent, though: fewer rewards for idle subnets, more for those attracting real demand.
This pool starts with 10,000 TAO and 100,000 alpha. Drag to stake TAO and see what the pool gives back.
Pool: TAO / alpha
Before dTAO, staking was a fairly uniform yield. After it, staking into a subnet is a judgment about that subnet's future. Alpha tokens can lose most of their value, and several have.
08 / Staking
Staking means delegating tokens to a validator. You keep ownership, the validator gains voting weight, and you share the rewards minus the validator's commission. There are two very different places to do it.
You stake TAO to a validator on the root network and your position stays denominated in TAO. You are not exposed to any single subnet's token price. Root stakers receive a slice of the rewards from across the subnets.
Since the "Root Reborn" upgrade in August 2026, that yield accrues as subnet tokens held in a per-validator basket and is only converted when the staker claims it, rather than being sold for TAO every block. A planned feature that would let validators actively choose which subnets their basket backs was not switched on at launch.
You swap TAO for a subnet's alpha token and stake that with a validator on the subnet. Rewards are paid in more alpha. Potential returns are higher and so is the risk: your position moves with the alpha price, and thin pools mean slippage on the way in and out.
Subnet stake also counts for more in consensus. Within a subnet, TAO staked at root is weighted at 18% of the value of alpha staked directly.
A May 2026 report by FalconX and SubnetStats estimated that about 70% of all TAO was staked: roughly half of supply at root and about a fifth in subnets. These figures move constantly.
After a major subnet team left abruptly in April 2026, the network introduced "Conviction". In its first phase, live since May 2026, subnet owners' rewards are locked automatically and owners must post a public on-chain request before unlocking them, giving holders warning of large sales.
09 / People and organisations
Bittensor has no company at the top, but a fairly small set of people and teams have shaped it. Roles below are as publicly reported up to October 2026.
Former Google engineer who conceived Bittensor and wrote much of its early code. He led the Opentensor Foundation until the chief executive role was removed in February 2026, and remains the most influential contributor, proposing major protocol changes and building subnets.
Computer scientist who joined in 2019 and helped launch the network and foundation. He later co-founded Crucible Labs, which makes a Bittensor wallet and staking tools.
The pseudonymous name on the Bittensor whitepaper. The identity behind it has never been publicly confirmed. Yuma Consensus and DCG's Yuma subsidiary both take the name.
The non-profit that built and maintained the chain and software from launch. In 2026 it stepped back from formal control and handed over its code repositories.
Latent Holdings, a company associated with Jacob Steeves, operates the TAO.app explorer. Since April 2026 the Bittensor software kit and command-line tool have been maintained under the Latent name instead of the foundation. Critics note there is no second independent team maintaining that software.
The team behind several prominent subnets, including Chutes (model hosting) and Gradients (fine-tuning).
A studio co-founded by Will Squires and Steffen Cruz that runs several subnets, among them Apex, IOTA and Data Universe.
Developer of Targon, the confidential-compute subnet. It has reported raising a $10.5 million Series A round.
Digital Currency Group, founded by Barry Silbert, has backed Bittensor since 2021. Its subsidiary Yuma, launched in November 2024 with Silbert as chief executive, incubates subnets, runs a large validator and manages subnet funds.
Polychain incubated the project in 2019. dao5, founded by former Polychain partner Tekin Salimi, is another early holder. Both acquired TAO through the network, since there was no token sale.
Runs the Grayscale Bittensor Trust and has filed to convert it into an exchange-traded product. Also publishes research on the network.
Built Templar, Basilica and Grail, and trained a 72-billion-parameter model on the network. Left in April 2026 after a public dispute over governance; see the risks section.
10 / Ecosystem and institutions
Grayscale runs a TAO trust that trades over the counter in the United States under the ticker GTAO. On 30 December 2025 it filed with the SEC to convert the trust into a spot exchange-traded product, and amended that filing in April 2026. As of early October 2026 it had not been approved. That month Grayscale added Coinbase as a second custodian alongside BitGo.
In October 2025 Deutsche Digital Assets and the Swedish broker Safello listed a physically backed, staked Bittensor exchange-traded product on the SIX Swiss Exchange under the ticker STAO.
DCG's Bittensor subsidiary Yuma launched an asset management arm in October 2025 with two funds for accredited investors that hold baskets of subnet tokens, anchored by a reported $10 million from DCG.
A handful of publicly traded companies hold TAO as a treasury asset. CoinGecko's tracker lists TAO Synergies (Nasdaq), xTAO (TSX Venture), TaoWeave (formerly Oblong) and Safello Group. Holdings change; check the tracker in the sources.
TAO was listed by Binance in April 2024 and by Coinbase in February 2025, and trades on Kraken, KuCoin, OKX and others. In June 2026 Kraken added seven subnet tokens to its public listing roadmap, a first for a large exchange, without committing to dates.
The EVM layer lets Ethereum-style applications run on Bittensor. Early projects include bridges, liquid-staking tokens and trading front ends for subnet tokens. These are young and carry smart-contract risk.
11 / Roadmap and future
Bittensor has no single binding roadmap. What follows are directions that core contributors have published or shipped. Plans in this ecosystem change often, and proposals are not promises.
In February 2026 the Opentensor Foundation announced it was giving up privileged control. The chief executive role was removed, chain security moved toward nominated proof of stake, and validators and subnet owners gained the ability to ratify or veto upgrades. In April 2026 the core code repositories were transferred to a separate team, Latent.
A May 2026 refactor, the June cleanup of inactive subnets and the emission gate all push rewards toward a smaller number of active subnets.
Root yield now accumulates in validator baskets. Validator-curated baskets and later phases of Conviction, including a way for token holders to replace a subnet's owner, were still to come at the time of writing.
Published in June 2026, the plan targets December 2027 and includes bringing back real competition between validators, two-way subnet liquidity pools, and a "belief mechanism" giving alpha holders a direct vote on subnet decisions.
The community is pulling in two directions: expanding the number of subnet slots toward 256, while concentrating emissions on a few dozen. How those fit together is unresolved.
Several subnets now report meaningful income from outside customers, mostly for compute and inference. Most figures are self-reported and unaudited. Whether subnets can earn more than they receive in emissions is the long-term test.
12 / Comparison
These projects are often grouped together but do different jobs. The table is a simplification; scroll sideways on small screens.
| Bittensor (TAO) | Bitcoin (BTC) | Render (RENDER) | ASI Alliance (FET) | Akash (AKT) | NEAR (NEAR) | |
|---|---|---|---|---|---|---|
| What it is | A network of competing AI marketplaces (subnets) | Peer-to-peer digital money | Marketplace for GPU rendering and AI compute | Alliance of AI-agent projects led by Fetch.ai and SingularityNET | Open marketplace for cloud computing | General-purpose blockchain with an AI focus |
| What earns rewards | Work judged useful by validators, per subnet | Proof-of-work hashing | Completing GPU jobs for paying customers | Staking and running agents and services | Leasing compute to customers | Staking to secure the chain |
| Launched | 2021 | 2009 | 2017 token; network later | Alliance formed 2024 | 2020 mainnet | 2020 mainnet |
| Supply | Capped at 21 million | Capped at 21 million | Burn-and-mint model | Fixed maximum | Fixed maximum, released by inflation | No hard cap |
| Halving | Yes, triggered by issuance thresholds | Yes, every 210,000 blocks | No | No | No (declining inflation) | No |
| Own blockchain | Yes (Subtensor, Substrate) | Yes | No (token on Solana) | Yes (Fetch.ai chain) plus Ethereum | Yes (Cosmos SDK) | Yes |
| Scope | Any digital commodity that can be scored | Money only | GPU compute | Autonomous agents and AI services | General compute | Apps, agents, private AI inference |
The fixed cap, the halving and the absence of a pre-mine are deliberate echoes of Bitcoin. The difference is what the reward buys. Bitcoin pays for security through energy spent. Bittensor tries to pay for output that has value in its own right, which is far harder to measure.
Render and Akash match buyers with hardware and charge for it. Bittensor includes compute subnets that do the same, but it also funds contests with no customer yet, paid for by token emissions. That makes it more flexible and more dependent on good incentive design.
13 / Wallets
A wallet holds the keys that control your TAO. The options below all appear on the wallet page of bittensor.com. Filter by the device you use.
A Bittensor-only mobile wallet that grew out of the original Opentensor Foundation app. Supports staking and browsing subnets; the same site also offers swaps into subnet tokens.
From the team behind the Taostats explorer. Handles transfers and staking alongside network analytics.
A Chrome extension from Crucible Labs, co-founded by Ala Shaabana, with staking tools and Ledger support.
An open-source, self-custodial, multi-chain wallet that supports Bittensor alongside Ethereum and Polkadot.
A mobile wallet for Substrate-based chains with TAO staking and Ledger support on mobile.
A mobile and browser wallet for Substrate chains that supports TAO transfers and staking.
A hardware wallet that keeps keys offline. Used for TAO through a compatible app such as Crucible, Nova or SubWallet.
A tap-to-sign card wallet. Support is partial: TAO transfers only, no staking.
Turns a spare phone kept permanently offline into a signing device that communicates by QR code.
The official command-line tool for creating wallets, staking, and running miners and validators. For technical users.
A developer-oriented browser extension and interface for advanced interactions with Substrate chains, including Bittensor.
14 / Where to buy and how to stake
This explains the mechanics. It is not a suggestion that you should buy TAO. Availability depends on your country, and exchanges can add or remove assets at any time.
Shown in no particular ranking; links go to each exchange's own home page. Confirm that TAO is available in your region before opening an account.
15 / Risks and criticisms
Bittensor attracts strong opinions. These are the main concerns raised by critics and, in several cases, by people inside the ecosystem.
For most of its life, upgrades were controlled by a three-person "triumvirate" linked to the Opentensor Foundation. In April 2026 Covenant AI, then one of the largest subnet operators, quit the network and its founder described the governance as decentralisation in name only, alleging that co-founder Jacob Steeves held effective control. Steeves disputed many of the allegations. Reforms have followed, but they are recent and incomplete.
A small number of validators and early holders control a large share of stake. Because voting weight follows stake, they have outsized influence over scores, rewards and upgrades.
Many subnets have produced little of outside value. In April 2026 one analysis found more than half of all subnets were burning nearly all of their miner rewards, and in June 2026 emissions were cut off for dozens deemed inactive.
Miners optimise for whatever is measured. If a subnet's test is weak, miners will pass the test without doing useful work, copy each other, or collude with validators. Designing a mechanism that cannot be gamed is hard and never finished.
A subnet depends heavily on its team. When Covenant AI left, it sold its subnet tokens and their prices fell sharply within hours, pulling TAO down with them. A separate January 2026 incident on one subnet was reported by one analyst to have cost users close to $3 million. Conviction locks are a response, not a cure.
On 2 July 2024 an attacker drained about 32,000 TAO, then worth roughly $8 million, from users whose keys had been stolen by a malicious copy of the Bittensor software package uploaded to the Python package index. Validators put the chain into a "safe mode" that stopped all transactions while the code was reviewed. The halt limited losses and also showed that a small group could stop the network.
Two kinds of key, more than a hundred tokens, automated pools, changing emission formulas. The system is hard to understand even for experienced crypto users, and misunderstanding it can be expensive.
Core economics have been rewritten several times in two years. Each change creates winners and losers and makes long-term planning difficult for subnet builders and stakers.
TAO and alpha tokens are volatile, and alpha pools can be thin. Rules for staking and tokens differ by country and are still evolving. You can lose everything you put in.
16 / Glossary
17 / Quiz
18 / FAQ
Bittensor is an open network that rewards people with a cryptocurrency called TAO for producing useful AI services. The work is organised into subnets, each a separate competition where miners do a task, validators score them, and the blockchain pays out according to those scores.
Bittensor was co-founded by Jacob Steeves and Ala Shaabana, who set up the Opentensor Foundation to develop it. The whitepaper is credited to Yuma Rao, a pseudonym whose identity has not been publicly confirmed. The network launched in 2021.
TAO is the native token of the Bittensor network. Its supply is capped at 21 million. New TAO is created with each block and the amount halves over time. The first halving took place in December 2025, reducing issuance from about 7,200 to about 3,600 TAO per day.
A subnet is an independent marketplace for one kind of digital work, such as running AI models, renting GPUs, storing data or making predictions. Each subnet has an owner who sets the rules, miners who do the work, validators who grade it, and its own token.
Dynamic TAO is an upgrade from February 2025 that gave every subnet its own alpha token and a liquidity pool pairing it with TAO. Staking TAO into a subnet swaps it for that alpha token. Market demand for subnet tokens, rather than a committee of validators, now steers how new TAO is shared between subnets.
No. Bittensor copies parts of Bitcoin's economics, including the 21 million cap, halvings and a launch with no pre-mine. But Bitcoin is a payment network secured by proof-of-work, while Bittensor is a proof-of-stake chain that rewards AI-related work judged by validators.
You stake by delegating TAO to a validator from a compatible wallet. Staking on the root network keeps your position in TAO. Staking into a specific subnet converts your TAO into that subnet's alpha token, which carries price risk and slippage. Validators charge a commission on rewards.
TAO is listed on several large exchanges, including Binance, Coinbase, Kraken, KuCoin and OKX. Availability depends on your country. Always use the exchange's official website and check you are buying Bittensor with the ticker TAO.
In July 2024 attackers uploaded a malicious copy of the Bittensor software package that stole users' keys, then drained about 32,000 TAO worth roughly $8 million. Validators halted the chain while the code was reviewed. The blockchain's own consensus was not broken; the theft came through compromised software on users' machines.
Partly, and it is disputed. For years upgrades were controlled by a small group tied to the Opentensor Foundation, and stake is concentrated among a few large validators. In 2026 the foundation stepped back and published a plan to decentralise governance by the end of 2027, after a major subnet team left accusing the project of centralised control.
Not in the United States as of October 2026. Grayscale filed in December 2025 to convert its Bittensor trust into an exchange-traded product and the application was still pending. In Europe a staked TAO exchange-traded product has traded on the SIX Swiss Exchange since October 2025.
This site does not give investment advice. TAO is a volatile crypto asset, the network's rules change often, and subnet tokens can lose most of their value. Anyone considering it should understand the risks, read primary sources and only use money they can afford to lose.
19 / Sources
This guide was researched in October 2026 from official documentation and independent reporting. Where sources disagreed or a detail could not be confirmed, the wording above is deliberately cautious. Spotted an error? The primary sources below are the place to check.