PERCEPTRON
NETWORK

// Abstract

Perceptron Network is a decentralised data mesh that transforms idle bandwidth and human cognition into a globally distributed infrastructure for AI data acquisition and agentic web access. The network operates through complementary layers: an off-chain node layer that provides residential endpoints for passive data collection and human-authentic web access for AI agents, and an active participation layer through which node operators contribute targeted, structured data via a bounty-driven mechanism called Data Questing. Emissions of the native token, $PERC, are distributed to node operators as a function of measurable performance scores, creating a self-reinforcing incentive structure aligned with data quality and network growth. This paper describes the architecture, economics, and design principles of Perceptron Network.

perceptrons.xyz

1.  The Data Imperative

The progress of artificial intelligence over the past decade has been defined as much by data as by computation. The transformer architecture, scaling laws, and the infrastructure of modern deep learning all rest on a single foundational dependency: access to vast, diverse, and high-quality data. As the frontier advances, that dependency is becoming a structural constraint.

Four converging pressures are now reshaping how the industry thinks about data supply.

1.1  Inference Scaling and the Shift to Test-Time Compute

The dominant scaling paradigm of 2020–2023, which held that simply training larger models on more tokens would yield predictable capability gains, has matured toward diminishing returns on pure parameter count. The leading laboratories have increasingly turned to inference-time computation — techniques such as chain-of-thought reasoning, repeated sampling, and tree-search over candidate outputs — as the lever for continued capability growth.

This shift has significant implications for data. Inference scaling requires models that can reason across diverse domains, handle ambiguous problem specifications, and generalise to novel distributions. The breadth and representativeness of training data becomes a first-order concern. Models trained on narrow or sanitised corpora underperform precisely in the reasoning tasks that inference scaling is designed to improve.

1.2  The Rise of Specialised and Domain-Specific Models

Alongside the frontier race for general capability, a parallel movement toward highly specialised models has taken root. Fine-tuned and domain-specific models have demonstrated that, for a large class of commercial tasks, a smaller model trained on high-quality domain data consistently outperforms a larger general model. The implications for data supply are direct: the value of a dataset is no longer solely a function of scale but of its specificity, provenance, and timeliness.

Hedge funds require real-time sentiment across private Discord communities. Life-sciences companies need structured clinical language corpora. Autonomous systems require diverse visual and geospatial data. Each of these demands a distinct collection infrastructure that centralised providers are ill-suited to supply at the required cost, granularity, or freshness.

1.3  Knowledge Distillation and the Synthetic Data Ceiling

The practice of distillation — using the outputs of larger models to train smaller ones — has accelerated rapidly, enabling remarkable capability compression. However, distillation is fundamentally bounded by the capabilities of the teacher model and, more critically, by the diversity and novelty of the inputs it is presented with during that process. Synthetic data generation, while powerful as a supplementary technique, cannot on its own introduce new distributional information about the real world. The quality of the data mesh that supplies real-world inputs to these training pipelines remains irreplaceable.

1.4  The Agentic Web and the Bot Problem

The most immediate commercial driver is the rise of autonomous AI agents that browse and transact on the open web. Agents are not people, and the web treats them accordingly. When an agent accesses a website from a datacenter IP address, or browses in a manner that does not resemble human behaviour, it is routinely classified as a bot — rate-limited, served degraded or misleading content, or blocked outright. As agents move from research demonstrations to production deployment at scale, this bot problem becomes a hard ceiling on what the agentic economy can actually accomplish.

This is the dynamic that drove one of the most significant recent capital events in the sector: Nimble’s $47M raise was, at its core, a bet that agents need human-like access to the web, and that providing it is a foundational infrastructure business. A residential IP network is precisely the instrument that solves this — it allows an agent to interact with the internet from the same kind of endpoints, and in the same manner, as a genuine human user, rendering it indistinguishable from organic traffic. The market has validated, in the clearest possible terms, that this capability commands a premium.

The compounding result of these four pressures is a structural demand imbalance: the world’s AI infrastructure needs more, better, more diverse, and more human-authentic data and web access than current collection mechanisms can supply at a sustainable cost. Centralised providers are expensive, slow to adapt, and concentrated. The logical response is decentralisation.

2.  The Perceptron Approach

Perceptron Network addresses the data supply problem by distributing both the collection infrastructure and the human labour of data production across a global community of participants, coordinated by cryptoeconomic incentives. The network is designed around a core insight: the raw ingredients of valuable AI data — residential IP addresses, idle bandwidth, authentic human attention, and personal data assets — are already globally distributed. What is missing is the coordination layer.

The network is not a blockchain in the traditional sense. Nodes operate off-chain, contributing real-world resources and producing real-world data. The on-chain layer is reserved for what it does best: transparent, tamper-resistant accounting of contributions and trustless distribution of rewards.

2.1  Competitive Context

Perceptron operates in a market that has historically been served by a small number of large, centralised providers.

The key differentiator is structural, not incremental. Pure marketplaces such as Kled monetise the data a participant already holds; pure infrastructure providers such as Nimble monetise bandwidth. Perceptron is the only network that unifies both: a node is simultaneously a passive residential collection endpoint and an active human contributor able to fulfil targeted Data Quests. This dual role means a single participant generates two distinct, complementary streams of value, and the network as a whole addresses the full spectrum of AI data demand rather than a single slice of it.

3.  Network Architecture

Perceptron Network is composed of two integrated layers: a passive infrastructure layer operated by node runners, and an active contribution layer through which nodes participate in targeted data collection tasks.

3.1  The Node Layer

A Perceptron node is a lightweight software client that registers a participant’s device as part of the network. The node operates primarily off-chain. Its core function in passive mode is to contribute the device to Perceptron’s residential IP network, making its bandwidth available for structured web data collection when instructed by the network.

The node ships in multiple form factors to maximise reach across the device landscape:

• Browser extension: a desktop client that runs in the background of the participant’s browser, providing residential bandwidth from genuine consumer browsing environments.
• Android application: a native mobile client that extends the network onto mobile devices and carrier connections, broadening both the geographic and network-type diversity of the node base.
• iOS application (forthcoming): a native iPhone client that will complete coverage across the dominant consumer device platforms.

Residential IP addresses are distinct from datacenter IP addresses in a critical respect: they are issued by consumer internet service providers and mobile carriers to real subscribers. When Perceptron’s collection layer issues a request via a residential node, the resulting data appears to the target infrastructure as originating from a genuine consumer device in that geography and on that network type. This produces data that is structurally different from, and more representative than, what centralised scraping infrastructure can collect.

Node operators are pseudonymously identified on-chain. The network tracks a set of performance metrics for each node across each epoch, which form the basis of reward calculation.

Two Functions of the Residential Layer

The residential node layer delivers value in two distinct ways, and it is important to distinguish them because they address different customers and different needs.

The first is data collection: the network uses its residential endpoints to gather publicly available web data at scale, structuring it into datasets for sale. This is the supply side of the data business described throughout this paper.

The second, and increasingly the more strategically significant, is agentic web access. The same property that makes residential nodes valuable for collection — their indistinguishability from genuine human users — makes them the ideal interface through which autonomous AI agents can interact with the open web. An agent routed through the Perceptron network browses from a real consumer IP, on a real consumer device profile, in a manner consistent with human behaviour. It is not flagged as a bot, not rate-limited, and not served the degraded content that web infrastructure reserves for automated traffic. As the agentic economy scales, this becomes a recurring, high-margin access product: agents do not consume web access once, they consume it continuously, every time they act.

This positions Perceptron directly in the same category that the market rewarded with Nimble’s $47M raise — but as a decentralised network whose coverage and authenticity scale with community growth rather than centralised infrastructure spend.

3.2  Data Questing

Data Questing transforms node operators from passive bandwidth providers into active data contributors. The mechanism works as follows: a data buyer or internal protocol module posts a Data Quest — a structured specification of the data required, the quality criteria that must be met, and the reward available for fulfilling the quest. Node operators who meet the eligibility criteria may opt in to the quest and contribute the requested data.

The mobile node clients are central to this layer’s reach. A browser extension is well suited to passive bandwidth contribution, but the richest active contributions — photographs, voice samples, video, and on-device documents — originate on mobile devices. The Android application, and the forthcoming iOS client, therefore serve as the primary distribution surface for Data Questing: they place the quest interface directly in the hands of participants on the devices where the most valuable human data is created. As the mobile footprint grows, so does both the volume and the diversity of data the questing layer can source, creating a direct link between client distribution and data-market supply.

Quest types span a broad range of data modalities:

• Language-specific corpora: text contributions in underrepresented languages for training multilingual models.
• Image and video datasets: real-world visual data from participant devices, subject to consent and privacy specifications.
• Conversational data: opt-in export of interaction histories from AI assistants and messaging platforms.
• Community intelligence: structured extraction of community discourse from platforms such as Discord, used for sentiment analysis and trend detection.
• Proprietary data assets: upload of data that participants already hold, such as domain-specific archives, professional datasets, or annotated examples from specialist fields.

The bounty for each quest is denominated in $PERC and is distributed to fulfilling nodes upon verification. Quest quality is assessed by a combination of automated schema validation, statistical sampling, and, where applicable, community review. Nodes with a history of high-quality quest contributions receive score boosts that increase their share of both quest rewards and passive emissions.

3.3  The Discord Intelligence Product

Perceptron’s tryparsely.ai platform provides a structured intelligence layer built on top of publicly accessible Discord data. The product covers over 20,000 Discord channels and has indexed over 600 million conversations to date. The output datasets are structured for direct use in AI training, sentiment analysis, community trend detection, and market intelligence applications.

This product represents Perceptron’s first dedicated data-as-a-service offering and is a direct demonstration of the network’s ability to structure raw community data into commercially deployable datasets.

4.  Token Economics

The $PERC token is the native unit of account and incentive within Perceptron Network. It serves three functions: rewarding node operators for network contributions, enabling data buyers to post and fund quests, and governing the allocation of protocol resources over time. $PERC will launch on Solana, with planned expansion to Sui via Chainlink.

4.1  Supply

The total supply of $PERC is fixed at 86,000,000,000 tokens (86 billion). This cap is set at genesis and cannot be exceeded; there is no inflationary minting beyond it. A defined portion of this fixed supply is allocated to network emissions — the pool from which node operators are rewarded for their contributions over time. The emission mechanism described below governs the rate and shape at which tokens are released from this pre-allocated pool into circulation; it does not create supply beyond the 86 billion cap. The terminal state of the token supply is therefore fully defined at launch.

4.2  Emission Mechanism

Emissions release tokens from the fixed emissions pool on a per-epoch basis. Rather than following a flat or purely time-based schedule, the amount released each epoch is determined by the aggregate verified work performed by the network in that epoch and the total supply already emitted over the network’s lifetime. This ties token release to demonstrated network activity while enforcing a scarcity brake that causes emissions to decline as the pool depletes. Because the release rate falls as cumulative emissions rise, total emissions converge to a finite sum that cannot exceed the allocated pool — fully consistent with the fixed 86 billion supply cap.

The per-epoch mint is governed by the following curve:

                    1 − e^(−w/W)
  mint = R · M · ────────────────
                      1 + h/H

where:

• R  =  emission multiplier (0 < R ≤ 1), set per epoch by the protocol
• M  =  maximum $PERC mintable per epoch (constant)
• w  =  net verified network work in the epoch
• W  =  work curve factor (controls saturation rate)
• h  =  lifetime $PERC already emitted
• H  =  supply difficulty factor (controls scarcity ramp)

The mechanism has two intuitive components. The numerator, 1 − e^(−w/W), is a saturating function of network work: as more verified contribution flows into an epoch, the mint rises toward the per-epoch ceiling M, but with diminishing returns. This prevents runaway emissions during activity spikes while still rewarding genuine growth. The denominator, 1 + h/H, is a scarcity brake: as cumulative lifetime emissions h grow relative to the difficulty factor H, every subsequent mint is proportionally reduced. It is precisely this denominator that guarantees the fixed cap holds — as h rises, the mint shrinks toward zero, so cumulative emissions approach a finite limit rather than growing without bound. Early in the network’s life, when h is small, mints are near their activity-driven maximum; as the network matures, the same amount of work yields progressively less new supply.

The combined effect is an emission curve that is activity-responsive in the short run and deflationary in the long run. Unlike a fixed linear or geometric schedule, this design ties new supply directly to demonstrated network contribution, ensuring that tokens are minted in proportion to the real economic value the network is producing. The parameters R, M, W, and H are protocol-level parameters that allow the network to tune the balance between bootstrapping incentives and long-term scarcity.

4.3  Node Reward Allocation

The per-epoch mint determined by the emission curve is not distributed equally across all active nodes. Instead, each node receives a share proportional to its normalised Performance Score (P) within that epoch.

Rewardᵢ = mint × ΔPᵢ

where:

• ΔPᵢ  =  Pᵢ / Σ Pᴸ for all active nodes j

Pᵢ = w₁·Uptimeᵢ + w₂·Bandwidthᵢ + w₃·FulfilmentRateᵢ + w₄·QuestScoreᵢ

  where w₁ + w₂ + w₃ + w₄ = 1, weights set by governance

Performance data is reported by nodes off-chain at the end of each epoch and verified by the protocol’s oracle layer before on-chain settlement. Nodes that report inaccurate metrics are subject to score penalties in subsequent epochs. The weight vector (w₁, w₂, w₃, w₄) is adjustable through governance, allowing the network to shift incentive emphasis as the product mix evolves — for instance, increasing the weight on quest score as Data Questing grows in commercial importance relative to passive bandwidth provision.

4.4  Quest Reward Distribution

Quest rewards operate as an additional, separate incentive layer on top of passive emissions. When a data buyer funds a quest, the bounty is paid in $PERC into a protocol-operated distribution wallet. Upon verification of fulfilled contributions, the protocol distributes the bounty to qualifying nodes in proportion to their verified contribution volume within the quest specification, net of the protocol burn described in Section 4.5. Quest rewards are not subject to the epoch emission curve; they represent direct data-market value flowing to node operators.

This design creates a dual-income structure for node operators: a predictable baseline from protocol emissions and a variable, performance-contingent income stream from quest participation. As data-market demand grows, the quest layer is designed to become an increasingly significant component of node operator income alongside protocol emissions.

4.5  Value Accrual and Token Burn

To counterbalance emissions and tie token value directly to network usage, a portion of the $PERC that flows through the network’s fee-bearing events is permanently burned. Because these fees are denominated in the native token and arise from genuine network activity, the resulting supply sink scales with real usage rather than with token issuance.

Two burn events apply:

• Funding burn: 10% of every quest bounty and agentic access payment is burned at the point of funding. The remaining 90% is distributed to the nodes that fulfil the quest or provide the access. The protocol does not retain its portion of these pass-through flows — it is destroyed.
• Claim burn: 2% of each emission reward claim is burned at the point of claim, so that the burn mechanism scales directly with the volume of tokens entering circulation.

These two events ensure that burn volume exists from launch — driven by reward claims — and grows with the data economy as quest and access volume increases. Together with the requirement that buyers acquire $PERC to fund quests and access, the network exhibits both a structural source of buy-side demand and a usage-linked supply sink, counterbalancing the emission curve. The burn percentages stated above are initial parameters and may be adjusted through governance as the network matures.

Burn flows are distinct from direct dataset revenue. Where a customer purchases a structured dataset outright (Section 5.1), that payment is protocol revenue and is not a pass-through reward flow; it is therefore not subject to the funding burn.

4.6  Token Utility

$PERC serves the following functions within the network:

• Network rewards: primary mechanism for compensating node operator contributions.
• Quest and access funding: data buyers and agent operators denominate and fund bounties and access payments in $PERC.
• Access and staking: enhanced access to premium data products and priority quest eligibility may require staked $PERC.
• Governance: token-weighted voting over protocol parameters including emission weights, burn rates, quest verification standards, and treasury allocation.

5.  Data Marketplace and Demand Architecture

Perceptron is architected so that token value accrues from genuine network demand across multiple complementary sources. Rather than depending on a single revenue line, the network is designed to capture value from dataset commercialisation, direct network access, and agentic web access — each denominated in or settled through the network’s economic layer. The following sections describe these demand sources and how they reinforce the token economy.

5.1  Dataset Sales

Structured data products generated by the network are distributed through direct enterprise sales and through third-party data marketplaces. The network’s signed integration with Brickroad provides an initial distribution channel for dataset commercialisation. Dataset categories include real-time signal intelligence, sentiment analysis corpora, trend detection feeds, domain-specific AI training datasets, and community intelligence products via tryparsely.ai.

5.2  Network Access and Agentic Web Access

Enterprise and institutional customers can access the Perceptron residential network directly for two purposes. The first is custom data collection workloads, priced per terabyte of data collected, substituting directly for legacy centralised proxy and scraping infrastructure at a structurally lower cost point.

The second is agentic web access: AI agent developers and operators route their agents’ web traffic through the Perceptron network to obtain human-authentic, non-bot-flagged access to the open internet. This is a recurring, usage-based product that scales with the volume of agent activity rather than with one-time data purchases, and it is the fastest-growing demand category in the sector. It directly addresses the market need validated by Nimble’s $47M raise, but is delivered through a decentralised network whose authenticity and geographic breadth scale with participation.

5.3  Target Customers

Primary customer segments include AI laboratories and model developers requiring training data at scale, hedge funds and quantitative research firms requiring real-time alternative data, enterprises requiring live market intelligence, and independent AI researchers and data scientists accessing the network via API. A core element of the network’s commercial positioning is a substantially lower cost of data acquisition relative to legacy centralised providers, driven by the distributed nature of the node base.

perceptrons.xyz  •  hub.perceptrons.xyz  •  x.com/PerceptronNTWK