Candles operates on a tri-agent model within Bittensor, involving miners, validators, and public participants, each with specific roles:

• Miners – These are contributors running algorithms or models to generate structured predictions about market direction. Miners can employ machine learning models, technical analysis, or heuristics to forecast whether a given crypto asset’s candle will close green (up) or red (down) for the specified timeframe. Essentially, miners act as signal generators, providing data-driven predictions to the subnet. They must submit their predictions within strict cutoff windows prior to candle closure (e.g. 1 hour before for hourly candles, 4 hours for daily, 24 hours for weekly), ensuring that all predictions are time-stamped well in advance of the actual market move.
• Validators – These nodes are responsible for grading prediction accuracy and enforcing the scoring/reward rules. After a candle’s timeframe elapses, validators retrieve the actual market outcome (e.g. TAO/USD price up or down) and confirm whether predictions were correct. They then update miner performance scores and determine rewards. The validator scoring framework gives 80% weight to a miner’s direct prediction accuracy and 20% weight to the miner’s success in public prediction pools (explained below), with the pool component weighted by pool size and stake. Validators thus not only check correctness but also integrate the miner’s “crowd performance” into their reputation. This incentivizes miners to participate honestly and effectively, since **accurate predictions and positive pool contributions both boost their on-chain reputation and ranking】.
• Public Participants – These are regular users (the wider community) who join prediction pools without running mining nodes. Using Candles’ subnet token (the S31 “Alpha” token), public users can enter their own guesses (green or red) into scheduled pools for hourly, daily, or weekly candles. Their participation is manual (based on personal intuition or analysis) rather than algorithmic, but it is a key part of Candles’ “wisdom of crowds” approach. Public participants effectively form a gamified prediction market, putting up a small stake of S31 tokens for each guess and standing to win rewards if they predict correctly.

These three agent types are tightly interlinked by Candles’ incentive design. Miners and public participants predict the same events, and miners themselves are required to engage in the public pools: by default, 60% of every miner’s block reward (“emissions”) is automatically recycled into the public prediction pools as their stake. This means miners are literally rewarded in kind and reinvested into the game – their own tokens go into the prize pools that public users (and miners themselves) compete in. A miner who allocates more than the default 60% of earnings back into the pools can even earn a bonus to their score weighting, whereas recycling less than 60% incurs a proportional penalty. This emissions-recycling mechanism aligns miners’ incentives with the broader participant community and ensures constant liquidity in the pools. In essence, miners “have skin in the game” on the outcomes they predict, which discourages spurious or low-effort predictions and helps sustain engagement in the system.

Prediction Pools: Candles organizes public participation through time-segmented prediction pools – separate pools for hourly, daily, and weekly candle predictions. Each pool corresponds to a specific asset (initially TAO/USD price, with potential to add other trading pairs in future) and a timeframe. Participants must join before the cutoff (e.g., join the daily pool at least 4 hours before daily candle close) and commit a fixed entry fee paid in S31 tokens. Pools will come in various sizes/configurations (e.g. pools of 10, 50, 100, or 1000 entries) with entry fees ranging from 1 up to 100 S31 Alpha tokens, allowing both small and high-stakes games.

Reward Distribution: Once the candle closes and validators determine the outcome, winners in the pool (those who guessed the direction correctly) share the prize. Candles’ reward distribution is designed to be competitive yet fair. Roughly 70–90% of the total entry fees in a pool are paid out to the winners. In fact, the protocol can “top up” rewards from a reserve wallet to target a payout of ~90–110% of entry fees, ensuring winners collectively receive at least their stakes back (and usually a profit). The remaining 10–30% of entry fees not paid to winners constitute a protocol fee – these funds may be burned, recycled, locked, or retained by the Candles treasury for operations, marketing, and growth. (In practice, Candles aims to retain closer to 10% in fees when possible, to stay competitive with industry-leading prediction platforms.) Additionally, sponsored pools or bonus prizes can be introduced (e.g. a partner project or subnet can sponsor a pool by contributing extra rewards from their own wallet), further enriching the game for participants. Notably, miners are automatically entered into these pools via their recycled emissions (instead of keeping all block rewards, a portion is used as their pool entries). This guarantees each pool has liquidity and active participation from the miner cohort at all times.

Validation Mechanism: After each prediction round, validators finalize the outcomes. They likely rely on an external price feed or oracle (for example, querying a trusted exchange rate for TAO/USD at the candle open and close) to determine whether the candle closed higher (green) or lower (red). Once the truth is known, validators programmatically assign scores and rewards: miners get score updates (impacting their reputation and future block reward share) and both miners and public users receive token rewards if they won in the pools. By weighting miner scores mostly on accuracy (80%) but partly on successful crowd predictions (20%), Candles marries individual performance with crowd performance. This unique validation logic incentivizes miners to not only be accurate but also to engage with and improve the public pools, since a miner who often wins in the pools (signaling alignment with crowd sentiment or informational edge) will gain an extra boost to their reputation.

All these operations are intended to be trustlessly automated via smart contracts. Wherever feasible, Candles will implement the pool entry validation, outcome resolution, reward distribution, and emissions recycling through on-chain smart contracts. This reduces reliance on any central administrator and ensures transparent, auditable behavior for the entire subnet. In summary, Candles’ operational design creates a self-sustaining ecosystem: miners contribute predictions and fuel the reward pools; validators enforce honest outcomes and scoring; and public participants drive the “wisdom of crowds” dynamic. The interplay of these roles – coupled with built-in token recycling and fee mechanisms – aligns incentives across the board and keeps the subnet economically healthy.

(Note: Public prediction pools are slated to be introduced shortly after the initial subnet launch. The subnet’s miner/validator code will launch first, with the public pooling feature following in an update once the core network is running smoothly. This phased rollout allows Candles to establish a stable mining/validation process before opening the system to broad user participation.)

Products and Deliverables

Candles delivers both a platform for predictive markets and the data products resulting from those markets:

• Decentralized Prediction Platform: At its heart, Candles is launching a prediction pools application on the Bittensor network. The deliverables here include the miner software and validator software that run the subnet’s incentive mechanism (miners generate predictions; validators evaluate them), as well as the smart-contract-backed pool system that users interact with. Participants will likely use a combination of tools – e.g. a web interface or a Bittensor wallet extension – to stake into pools and submit predictions. The design is inspired by familiar formats like daily fantasy sports and crypto prediction markets, making the user experience a key product focus. For instance, users can pick an upcoming candle (hour, day, or week), pay the entry fee, and choose “Green” or “Red” in a few clicks, similar to entering a fantasy sports contest. Smart contracts (either on the Bittensor chain or an integrated chain) manage the game: validating entries, locking the pool funds, and automatically paying out winners once validators confirm results. This provides a trustless, on-chain betting experience for market sentiment, with transparency guaranteed by code rather than a central house.
• Sentiment Indices and Analytics: One of Candles’ novel outputs is the aggregation of crowd prediction data into sentiment metrics. Over time, as miners and users make thousands of predictions, Candles will compute and publish market sentiment indices – analogous to a crypto Fear & Greed Index specifically derived from on-chain predictions. For example, if a large majority consistently predicts “green” candles for TAO, that indicates bullish sentiment which can be quantified into an index value. Candles plans to offer token-specific sentiment indices (e.g., a sentiment score for TAO/USD, and later for other supported assets) and “wisdom of crowds” dashboards. These dashboards would visualize data like the percentage of bullish vs bearish bets in recent periods, the historical accuracy of the crowd’s predictions, and possibly the correlation of crowd sentiment with actual market movements. Another analytic product is miner performance metrics – since each miner is effectively an algorithmic forecaster, Candles can track and display each miner’s historical accuracy, consistency, and ranking over time. Such statistics might be showcased on a dashboard or API for the community (and could even feed into a leaderboard or a trust score for miners). Importantly, these data products are not just for show: Candles sees them as potential revenue-generating services that could be consumed by external platforms. For instance, a DeFi analytics site or a trading app might integrate Candles’ sentiment index as an indicator, possibly paying a fee or revenue share for it. By turning raw prediction data into actionable indicators, Candles adds value to the broader crypto ecosystem and creates an additional incentive for the subnet’s longevity.
• Membership Program: Candles is introducing a membership system with tiered benefits, which is itself a product offering for users and other subnet owners. Three membership tiers (Tier 1, 2, and 3) will be available, purchasable with the Candles S31 token or $TAO (with discounts for paying in Candles’ own token). These memberships confer perks such as exclusive Discord channels with trading signal feeds, free or boosted entries into weekly pools, access to members-only high-reward pools, and discounts or premium content from partner subnets. Higher tiers expand the benefits – for example, Tier 2 and Tier 3 members can become “subnet partners” who are allowed to create or sponsor their own prediction pools using their project’s token. This effectively means that if another Bittensor subnet (or any crypto project) wants to leverage Candles’ platform for its community (say, to predict metrics related to their own project), they can do so by holding a membership. They could create custom pools (up to 5 pools for Tier 2, or unlimited for Tier 3) and even share in the revenue those pools generate. This “Prediction Pools as a Service” model is a deliverable that extends Candles beyond just TAO price prediction into a general framework any project could use. It also adds a revenue stream for Candles through membership fees and profit-sharing with those partner subnets. The memberships are thus a product that caters both to individual traders (seeking extra features and rewards) and other subnet developers (seeking to engage their user base with prediction games).
• APIs and Dashboards: While not explicitly named, it is anticipated that Candles will develop user-facing applications such as a web dashboard and APIs to access the above data and services. For example, a live dashboard may display real-time pool information (number of participants, current sentiment split) and the computed sentiment indices over time. Given the team’s interest in broad adoption and analytics, providing a public API or widget for the sentiment index is likely, enabling external websites to embed Candles’ Fear & Greed index for TAO or other assets. A portfolio dashboard could also be offered to Candles token holders, showing their membership status, winnings from pools, and miner stats. These tools would enrich the user experience and drive home the transparency of the platform.

In summary, Candles is building a full-stack product: a decentralized application for prediction pools (with smart contract automation), augmented by a rich set of analytics (sentiment indices, performance metrics) and a community engagement program (memberships and cross-subnet integrations). The Candles S31 token itself underpins these deliverables – it’s the unit of entry in pools, the reward currency, and the membership payment token. By blending a fun, game-like interface with serious data insights, Candles aims to both entertain and inform. Users can actively “play” the market sentiment game or passively use the sentiment data as a tool in their trading decisions. This dual nature makes Candles’ offering unique in the Bittensor ecosystem and the wider crypto market.

Technical Architecture and Integration

From a technical standpoint, Candles leverages the Bittensor substrate-based blockchain and its modular subnet framework as the backbone of its infrastructure. The subnet runs its own incentive mechanism logic on the Bittensor network, meaning it inherits Bittensor’s consensus (for security and block production) while implementing custom logic for prediction tasks and reward distribution. Key aspects of Candles’ architecture include:

• Tri-Agent Framework: The subnet’s code defines the roles of miners, validators, and participants as described, and the interactions among them. Miners and validators in Candles are specialized AI agents in the Bittensor context – miners can be thought of as prediction-serving endpoints and validators as evaluation nodes. This is analogous to other Bittensor subnets (which might handle Q&A, translation, etc.), but here the “service” being provided is a prediction of a future price movement, and the “evaluation” is checking that prediction against reality.
• Yuma Consensus Integration: Bittensor’s default incentive mechanism, known as Yuma consensus, handles how token emissions are distributed to miners and validators. Candles builds on this by modifying the scoring function (80/20 accuracy vs pool success) to allocate rewards. The underlying chain ensures that at each block, a certain amount of Candles S31 tokens are emitted and divided among miners and validators proportionally to their scores (with adjustments for the recycling rule). This is all recorded on-chain, providing transparency in how each miner’s reward was calculated. Essentially, Candles customizes the “reward function” of the subnet while relying on the base chain to carry out the token minting and transfer securely.
• Subnet Token (S31 “Alpha”): Candles uses a subnet-specific token often referred to as Alpha (for Subnet 31). In the Bittensor system, users can stake the main $TAO token into a subnet to mint these alpha tokens, which are used internally in that subnet’s economy. The S31 token thus represents staked value and is used for all transactions in Candles: paying entry fees, distributing rewards, and purchasing memberships. The architecture includes a two-way peg between TAO and S31 (facilitated by Bittensor’s staking and unstaking mechanism), and S31 has a floating market price against TAO (and by extension USD) depending on the subnet’s success. This design tightly integrates Candles with Bittensor’s economic framework – participants must interact via Bittensor’s wallet and staking processes to acquire S31 tokens, rooting Candles firmly within the ecosystem. According to network data, Candles’ token already carries value (as holders stake TAO into it) and will fluctuate as the subnet grows. The Alpha token utility is broad: it acts as the currency for pool bets and payouts, the unit of miner/validator rewards, and the medium for membership fees and subnet sponsorships. Candles’ architecture thus treats the token as a critical piece of infrastructure, with built-in mechanisms like token burning (e.g., a portion of fees can be burned) and recycling to manage its supply and demand.
• Smart Contracts and Automation: A distinguishing technical feature of Candles is its plan to use smart contracts for automation of critical functions. For example, instead of relying solely on validator nodes’ off-chain logic to handle pool payouts, Candles may deploy on-chain smart contracts (possibly leveraging the Polkadot/Substrate smart contract module or an EVM-based chain connected via bridges) that automatically escrow entry fees and disburse rewards according to the rules. The whitepaper indicates that contracts could cover entry validation, prize distribution, emissions recycling, and subnet sponsorships. This implies an architecture where certain components of Candles behave like a decentralized application (dApp) running on a contract platform, ensuring transparency and trustlessness. For instance, a pool contract could receive all entry fees, then when the outcome is decided, a validator triggers the contract to release 90% of those fees to the winner addresses on-chain, with the remainder going to a treasury address (for burn/operations). Similarly, a membership NFT or token might be implemented via smart contract to gate access to features. By offloading these tasks to deterministic contracts, the architecture minimizes any need for a centralized server or manual intervention in Candles’ operation. It essentially merges blockchain-based game logic with Bittensor’s consensus, showcasing a hybrid of AI subnet and DeFi-like smart contract infrastructure.
• Data Pipeline (Oracles): To verify predictions, Candles needs reliable external data (crypto prices). While the specifics aren’t fully detailed in the whitepaper, the validators likely incorporate an oracle mechanism to fetch the closing price of TAO/USD (and any other asset used) at the specified times. This could be done by querying a price API or utilizing a decentralized price oracle. The architecture must handle this data feed securely and accurately, since it directly affects rewards. We can expect that validators will use multiple sources or a medianized price to avoid manipulation. Once fetched, the price data is used in the validator’s logic to determine green vs red outcome, which is then recorded on-chain for transparency. Over time, Candles might integrate with a formal oracle pallet or service to publish price points on-chain (which could even benefit other subnets).
• AI/ML Models: From an AI architecture perspective, Candles does not enforce a specific model; instead, it sets up a competitive environment for models. Miners in Candles are free to choose their prediction approach – some may run time-series ML models (e.g. ARIMA, LSTM, reinforcement learning agents) on price data, others might code heuristic strategies (like following momentum or social sentiment), and some could simply copy human strategies. This open approach means the subnet’s “model architecture” is essentially the collection of all miner models running in parallel. Bittensor encourages such diversity, and Candles leverages it by allowing the best predictive models to earn more (through accuracy and pool success). In a sense, Candles is model-agnostic and functions as a meta-architecture: if a particular algorithm consistently outperforms others in predicting candle directions, its miner will naturally gain higher reputation and more rewards, pushing the network’s overall intelligence forward. Validators might not need complex models themselves – they perform a simpler aggregation role – but they do require robust software to handle scoring and possibly to run a baseline price prediction for comparison. The technical tooling for miners/validators is likely provided as open-source code (e.g., Python scripts or Jupyter notebooks, given NASChain’s codebase was largely Python/Notebook). Miners can adapt this code to integrate their model of choice. Thus, Candles’ architecture is flexible and can accommodate improvements in AI methods over time, turning the subnet into a continually evolving tournament of algorithms and human intuition.
• Integration with Bittensor Core: As a subnet, Candles is registered on the Bittensor network (Subnet ID 31) and benefits from core services like the Bittensor chain’s security, staking mechanism, and community support. The subnet’s lifecycle (registration, emission schedule, etc.) is governed by Bittensor’s rules. Candles has its own slot on Bittensor’s chain, and the OpenTensor Foundation (Bittensor’s core team) oversees the subnet framework generally, though Candles’ development is driven by its specific team. Because Candles was formerly known as NASChain during its concept phase, its integration has an interesting history: the team repurposed the existing subnet slot from an AutoML focus to the current predictive pools focus. However, the integration remained seamless since both concepts use the same underlying Bittensor principles (miners, validators, TAO-staking). The Alpha token economics (e.g., initial token distribution or any pre-mining from the NASChain phase) have carried over, and any early stakers of NASChain’s token became stakeholders in Candles once the pivot happened. This demonstrates Bittensor’s flexibility – subnets can evolve in their service while the core blockchain infrastructure stays consistent.

Overall, Candles’ technical architecture is an elegant fusion of blockchain, game theory, and AI. It relies on Bittensor’s decentralized protocol for scalability and security, introduces smart contract components for trustless execution, and fosters an open environment for predictive algorithms to compete. The end result is a self-contained economy of prediction that plugs into the wider Bittensor metagraph. As Candles matures, this architecture will enable further integrations – for example, plugging Candles’ sentiment data into other DeFi platforms via oracles, or using its membership token in DAO governance – solidifying its role as an important piece of Bittensor’s technical and social infrastructure.