ANONYX WHITEPAPER

Private Access to
Artificial Intelligence

Identity. Context. Inference. Payment.
A framework for keeping them separate.

OVERVIEW

Abstract

Anonyx is a privacy-oriented AI access platform designed to give people and applications greater control over the information they share with artificial intelligence systems. Its architecture separates identity, personal context, inference, and payment so these functions do not automatically form a single user profile.

The core product is a workspace for preparing a request, selecting useful context, inspecting the exact payload, and approving what is shared. Model access and multi-model workflows extend this request lifecycle. The chat interface is an entry point; controlled information disclosure is the underlying utility.

This edition describes the current browser controls and the intended infrastructure separately. Local preparation and approval reduce unnecessary disclosure, but do not make submitted text unreadable to a model provider or establish anonymous access. Confidential computing and privacy-preserving usage authorization require additional implementation and independent verification.

Anonyx is designed for blockchain-based components on Robinhood Chain. The purpose of that layer is programmable authorization and transparent resource accounting; prompts, context, and conversation content should remain off chain.

CHAPTER 01

Purpose and Product Scope

AI requests can contain client information, project details, credentials, preferences, and personal history. Account-based services may also link these requests with identity, payment, and network metadata. Anonyx addresses unnecessary information sharing within the parts of the request lifecycle it controls.

A useful request needs enough information to complete a task. It does not need every stored preference, every previous conversation, or a connected wallet address. Anonyx lets the user decide what is relevant before a provider receives the request.

  • Individuals prepare writing, research, analysis, and coding requests with explicitly selected context.
  • Teams can use the same disclosure principles for project information, subject to access controls that must be implemented for shared environments.
  • Developers can build around a consistent request format and explicit provider boundaries.

The product follows four principles: separate information domains, minimize disclosure, require meaningful user control, and support interoperable model access. These principles describe design requirements; their effectiveness depends on the implementation and the services involved.

CHAPTER 02

Implementation Status

The current release provides browser-based request preparation and a same-origin, server-side AI gateway. Live inference remains disabled until an operator configures a provider secret, model, approved-user allowlist and activation flag. Automated gateway checks do not replace an authorized live-provider verification.

Capabilities and their current boundary
CapabilityCurrent implementationAdditional requirement
Prompt and context preparationLocal editing, selected context and selected conversation turnsUser review of sensitive content
Privacy filteringLocal pattern masking for common emails and phone numbers; manual selection maskingBroader detection and evaluated accuracy
Request approvalApproval of a cloned payload snapshot; edits invalidate approvalContinued regression checks on every sending path
Context persistenceMemory by default; optional unencrypted browser storageProtected storage for stronger confidentiality
SessionsLocal clearing and invalidation of late UI responsesProvider-specific cancellation and retention verification
Wallet accessBrowser wallet connection, switching, and RPC balance readsNo payment execution is implemented
Live model accessServer-side gateway with secret configuration, model pinning and account allowlistProvider configuration, spend limits and authorized live end-to-end verification
Confidential computingArchitectural requirementExecution environment, attestation and key-release implementation
Usage settlement and token accessProposed architectureContract specifications, deployment and security review

The whitepaper does not establish deployment of a token contract, payment service, confidential inference environment, or verified production inference service. Website interactions and explanatory animations are not evidence of those services.

CHAPTER 03

Four Information Boundaries

Anonyx separates four information domains. Separation means restricting what one domain receives from another, rather than assuming different components cannot be correlated.

Architecture boundaries
LayerResponsibilityDisclosure rule
IdentityWallet or service authentication when a capability requires itDo not append identity fields to an AI payload by default
ContextUser-managed information and explicit request selectionKeep unselected items outside the request
InferenceChosen model/provider and the approved requestTransmit only approved fields through an explicit destination
PaymentProposed funding, authorization and settlementKeep prompt content off chain; assess financial correlation separately

For ordinary request preparation, a wallet is not required. The current payload builder receives selected content and model information, not wallet account state. A name, wallet address, or account identifier typed into a prompt is still included as text unless the user removes it.

A separate provider origin is an application isolation boundary. It does not hide a user’s IP address from that service, stop a provider from reading submitted content, or prevent correlation through login, timing, cookies, or payments.

CHAPTER 04

A Request from Draft to Disclosure

Consider a user preparing a client project summary. The context library contains a project goal, writing preferences, a contact list, and unrelated project notes. Only the goal and writing preferences are needed for this task.

Draft → Select context → Inspect payload → Approve → Send to the chosen provider

Worked request example
StepUser choiceResult
Draft“Write a concise summary of the project goal.”Text is prepared in the browser
SelectChoose the project goal and writing preferencesContact list, unrelated notes and prior turns remain excluded
MaskRemove contact details if they appear in the prompt or selected textReview for identifiers the pattern tool may miss
InspectRead the actual messages and destinationSelected item titles and bodies appear in the payload
ApproveApprove this exact snapshotChanging effective request content requires a fresh approval
ExecuteUse a configured, verified provider integrationProvider receives the approved text and applies its own policies

The approved payload contains the current prompt, selected context, selected conversation turns where applicable, and model/provider fields. Context inventory metadata and connected wallet fields are not automatically included. Selected context titles may themselves contain sensitive information and must be reviewed.

If external execution is unavailable, the interface must explain that at the point of execution. It must not manufacture an AI answer, a payment receipt, or a successful transaction. If a request fails or times out, another attempt requires a deliberate user action; uncertain completion must not be treated as success.

CHAPTER 05

Context Ownership and Storage

The context library lets users organize information independently from an individual model provider. Users can create, edit, delete, import and export context, then select only the items relevant to a request. The current selection mechanism is explicit user choice, not automatic semantic retrieval.

Context and active conversation history start unselected for request inclusion. Selected item titles and bodies become request content. Selecting an item grants permission to use that item for the reviewed request; it is not permission to upload the entire library.

Storage behavior
DataDefault locationImportant limit
Draft and active conversationBrowser tab memoryVisible to the active application and potentially a compromised browser
Context libraryBrowser memoryNot an encrypted vault
Remembered contextOptional unencrypted local browser storageAccessible within the same browser profile and origin; unsuitable for unprotected secrets
Exported contextUser-created JSON downloadOutside session deletion; requires separate handling
Provider recordsProvider infrastructure after sendingGoverned by provider terms and technical controls

There is no Anonyx cloud context synchronization in the current release. Encrypted persistence, team permissions, and semantic retrieval are possible extensions, each requiring its own key management, authorization, and disclosure policy. Encryption alone would not protect context while it is displayed or processed by a compromised application.

CHAPTER 06

Privacy-Aware Prompt Processing

Prompt processing reduces unnecessary disclosure before transmission. The current tool masks common email and phone-number patterns locally and allows selected text to be replaced manually. The resulting text remains editable and must be inspected before approval.

Pattern masking can miss names, addresses, API keys, unusual formats, combinations of facts, and identifiers embedded in context titles. It can also mask non-sensitive numbers. It is a convenience control, not a comprehensive sensitive-data detector or a guarantee of de-identification.

  • Review both the prompt and every selected context item.
  • Remove credentials and private keys rather than relying on automatic detection.
  • Check whether remaining project details can identify a person or organization.
  • Use only the information needed for the task and assess the provider’s policies before sending it.

Future filtering should be evaluated against representative data, with documented false positives, missed identifiers, supported languages, and processing location. If a remote service performs the filtering, that service becomes an additional recipient and must be disclosed before use.

CHAPTER 07

Sessions and Deletion

A private session limits locally retained working state. Ending a session clears the current draft, transcript and request selections, resets active workflow output, and prevents late responses from repopulating cleared UI state. Context-library management remains a separate operation.

Session clearing is not secure erasure of browser memory and does not remove remembered context, downloaded exports, browser records, hosting metadata, wallet-provider records, or external model-provider logs. Users must delete saved context separately and apply the relevant controls for other recipients.

Stopping a request or closing a provider window does not prove that upstream computation was cancelled. A provider may already have received the payload or incurred usage charges. An integration must disclose its cancellation behavior and distinguish local cancellation from confirmed upstream cancellation.

Future temporary credentials and short-lived service sessions should have explicit expiry, revocation and logging policies. Limited persistence must be described per component rather than promised for the entire system.

CHAPTER 08

Models, Routing and Multi-Model Workflows

Anonyx is designed to support models suited to different tasks, such as writing, coding, reasoning and analysis. The current local model profiles organize request preparation; they do not establish availability or performance of a live model. A connected provider catalog must supply verified model and provider identifiers.

Routing should preserve the reviewed destination. A provider must not be silently substituted after approval. If a model becomes unavailable, a different model or provider requires a new review rather than an automatic fallback that changes the disclosure boundary.

Generate → Critique → Refine

A multi-model workflow can ask one model for an initial response and another to critique or refine it. Every stage must identify its recipients and the text it will receive, including earlier model output. Sharing a response with another provider is a further disclosure, even if the original prompt is omitted.

The current workspace provides stage controls and request preparation. Any claims about live output quality, task-routing performance, service latency, or costs require a verified integration and measurement. Model output remains untrusted information and should not authorize tools, reveal secrets, or alter the user’s disclosure policy.

CHAPTER 09

Developer Integration and Verification

The developer-facing layer provides request JSON validation and export. The workspace also has a server-side gateway for an operator-configured OpenAI-compatible provider. This gated integration is not a public developer inference API, API credential service or durable usage-accounting backend.

The workspace gateway uses a server-held provider key and a fixed operator-configured HTTPS endpoint and model. It validates authenticated account access, same-origin requests, explicit approval and message limits, and strips unrecognized request fields. Client cancellation, timeouts and sanitized errors preserve the draft without automatic retries. The per-instance burst guard is not a durable global quota; provider spend limits and live integration verification remain required.

  • Validate request schema and sizes at every service boundary.
  • Bind authorization to the approved payload and destination; reject stale or changed requests.
  • Keep API secrets outside browser-visible configuration and public request exports.
  • Prevent unauthorized origins from opening a request channel.
  • Specify retry and idempotency behavior to avoid duplicate inference or charges.
  • Define log fields, retention and deletion behavior without recording prompt content by default.

The current source includes automated tests for payload selection, review snapshot isolation, context import/export rules, wallet switching, disclosure-animation behavior and gateway authorization, validation, cancellation and sanitized failures. Provider calls in gateway tests use controlled stubs. Passing these tests does not prove real provider availability, provider non-retention, deployed confidential computing, browser-extension safety or anonymous payment.

A public release should expose the relevant source version, integration configuration boundaries, test results and external review findings. Claims should point to evidence for the same deployed release.

CHAPTER 10

Privacy Threat Model

The protected information includes prompts, context, conversation history, credentials, identity, and links between usage and payment. Anonyx aims to reduce unintended disclosure of unselected content and unnecessary identity fields. It does not currently provide protection against every party that can observe an interaction.

Threats, controls and remaining assumptions
Threat or observerCurrent protectionRemaining exposure
Unintended context sharingExplicit selection and exact-payload approvalSensitive details can remain in selected text or titles
Changed request after approvalCloned review snapshot and approval invalidationAll sending paths and destination changes must retain these checks
Model providerOnly selected content is includedProvider can read approved content and may retain it under its policies
Hosting or network metadataNo added product analytics or prompt loggingHosting can observe IP addresses and page-request metadata
Wallet, RPC and blockchain observersWallet state is separate from AI payload constructionWallet/RPC requests and public transactions can reveal identity or activity
Compromised device, extension or delivered codeLimited retention and explicit storage choiceApplication-level controls cannot protect text from a compromised environment
Timing or identifier correlationData minimizationSeparate services may still link activity by login, IP, timing or repeated details
Malicious context or model outputUser-visible review and selectionContent may influence model behavior; no complete prompt-injection defense is claimed

Users trust the browser and delivered application code for local processing, hosting to deliver that code, and any chosen provider to honor its declared policies. A separate origin narrows access between application components but does not remove those trust assumptions.

Stronger privacy requires specified mechanisms and measured results. A claim of anonymous inference, zero retention or payment unlinkability needs evidence beyond the absence of a wallet address in the prompt.

CHAPTER 11

Confidential Computing Requirements

Confidential computing is a proposed capability for sensitive request processing or inference. It is not deployed or verified by the current Anonyx workspace. A protected intermediary would not protect text from a downstream provider if that intermediary forwards the text in plaintext.

For a future confidential execution mode to support meaningful claims, the protected workload and its remaining exposure must be explicit. Context transformation, routing and complete model inference have different boundaries and cannot share one undifferentiated privacy label.

  • Identify the execution technology, operator, hardware and protected workload.
  • Publish the software measurement and the policy used to verify attestation freshness and authenticity.
  • Define how encryption keys are released only to an approved measured workload.
  • Document ingress, egress, logging, crash-reporting and retention behavior.
  • Specify update, revocation, outage and recovery behavior.
  • Assess hardware, operator, side-channel and availability assumptions through independent review.

Attestation can provide evidence about an execution environment under specific assumptions. It does not automatically prove that software is correct, that every output is private, or that metadata cannot be correlated. Confidential mode should fail closed when its verification requirements are not met, rather than silently send the same data through a weaker path.

CHAPTER 12

Private Usage and Blockchain Infrastructure

Anonyx’s intended usage architecture separates funding, authorization, computation and settlement. The aim is to authorize resource consumption without placing prompt content or an unnecessary identity profile into each inference request.

Funding → Authorization → Computation → Settlement

Robinhood Chain is the intended network for EVM-compatible token accounting and programmable access components. The current website supports wallet connectivity and RPC reads. No Anonyx settlement contract, private balance, payment execution, or deployed usage-authorization protocol is established by this paper.

Prompt text, context, conversation history, and personally identifying request details should not be placed on chain. Even hashes of predictable sensitive data can permit confirmation attacks and should not be treated as private storage.

Public blockchain activity is observable. Separating payment software from inference software does not by itself make payments unlinkable. Any future privacy-preserving authorization design must specify the credential mechanism, issuer and verifier knowledge, replay prevention, revocation, accounting and correlation risks.

A usage service also needs a clear quote, spending authorization, completion state and reconciliation policy. Timeouts, partial workflows, refunds and duplicate requests must have defined behavior before users can rely on settlement.

CHAPTER 13

Token Utility and Economic Design

The proposed Anonyx token is intended to support eligible resource access, developer services, ecosystem participation and contributor incentives. Each utility depends on a deployed service or protocol mechanism; holding a token alone does not create access to unimplemented infrastructure.

Anonyx plans to launch its token through Pons V2 on Robinhood Chain. The ticker, selected launch configuration, pairing asset, fees, creator recipient, developer purchases and token address must be published when confirmed. A platform launch does not deploy Anonyx’s AI services or usage-settlement mechanisms.

Launch distribution

Pons V2 mints the full supply to its curve, including a reserve for graduation into locked Uniswap V4 liquidity. Supply and reserve amounts follow the selected configuration. No separate Anonyx team or treasury allocation is specified.

Optional Pons V2 buybacks vest tokens rather than burn them. Anonyx has not confirmed buyback settings or a burn policy.

Platform reference: https://docs.ponsfamily.com/v2. Launch terms must be checked at creation; earlier platform defaults are not Anonyx commitments.

Project holdings and operating funds

Any developer purchase or later acquisition for project operations must be disclosed separately from initial distribution, with the relevant wallets and funding source. A project-held token balance is not automatically a vested allocation. Any vesting claim needs a specified schedule and enforceable mechanism.

Anonyx has not assigned percentages of creator receipts to development, operations, rewards or holders. Operating budgets, accounting and treasury permissions must be specified before such commitments are made. Trading liquidity must not be presented as an available project operating budget.

Economic questions requiring specification
MechanismRequired publication
Service pricingHow provider cost, platform fees and token volatility affect quotes
Access authorizationWhich features require payment or token holdings and how this is enforced
Launch and holdingsSelected configuration, distribution reserves, project purchases, recipient wallets and any separately implemented vesting
Contributor incentivesEligibility, reward calculation, budget and abuse resistance
GovernanceEligible decisions, voting rules and administrative permissions
Contract controlsUpgradeability, emergency powers, security reviews and deployment evidence

Where adopted, supported payment assets and service credits should be distinguished from the ecosystem token. The final economic design must explain which asset funds a request and how usage is accounted for. No conversion rate or credit balance is established by the current release.

The token is described as a proposed utility asset, not equity, ownership of an AI provider, a guaranteed entitlement to revenue, or a promise of appreciation.

CHAPTER 14

Delivery and Evidence Roadmap

Progress should be communicated through verifiable capabilities rather than unsubstantiated release dates. The stages below describe intended acceptance requirements, not a commitment that these systems have already been delivered.

Delivery requirements
StageScopeEvidence needed
Local disclosure controlsSelection, masking, payload review, context management and session clearingSource review, regression checks and usable request receipts
Verified provider accessLive catalog, authentication, reviewed inference and workflow stagesEnd-to-end checks for destinations, errors, charges, cancellation and retention disclosures
Protected contextOptional encrypted storage and controlled retrievalKey lifecycle, recovery policy, selection behavior and independent assessment
Usage authorizationQuotes, resource budgets and settlementProtocol specification, deployed contract verification, replay protection and reconciliation checks
Confidential executionMeasured sensitive-processing or inference workloadAttestation policy, key-release tests, operator disclosure and independent security review

Each release should identify what changed, which controls were verified and what remains outside its guarantees. Security findings and provider-policy changes should be reflected in the website and technical documentation. Publishing source or passing continuous-integration checks is evidence of transparency and specific behavior, not a substitute for an audit.

CHAPTER 15

Limitations and Disclaimer

Anonyx’s practical privacy benefit is reducing unnecessary sharing and making disclosure deliberate. It cannot guarantee that external AI systems will never retain or share data. Users must assess the sensitivity of approved text and the conditions of each service that receives it.

The current implementation has no independently verified confidential inference, cryptographic payment unlinkability, encrypted context vault, production Anonyx API, or deployed settlement system. Proposed capabilities require implementation, configuration and verification before they can support operational claims.

Browsers, wallets, hosting providers, RPC services, model providers, blockchain networks and confidential-computing operators introduce separate risks and trust assumptions. Security controls can fail, providers can change their policies, and AI output can be incorrect or manipulated.

This document is a technical and product description. It is not financial, investment, legal, tax or other professional advice, and it does not establish contractual service guarantees. Readers should independently evaluate the technology, applicable requirements and risks before using services or acquiring any ecosystem asset.

This revision preserves Anonyx’s privacy-focused AI concept while specifying its present boundaries. Future claims must be supported by the mechanisms and evidence described in the relevant chapters.