Deep Research Archives

Effective date: September 6, 2026. This policy covers the website, research APIs, and Deep Research Archives MCP plugin operated by Jipyeong Choi. Contact: [email protected].

Public research, not personal records

Submit only selected material that you have the right to publish: a research title, note, public source URL, citation, and optional links to related research. General academic work about health, finance, security or biometrics is welcome. Individual patient records, personal biometric samples or templates, government identifiers, payment-card information, passwords, API keys, access tokens and other secrets are not accepted. Remove these locally before uploading; do not use this service to sanitize confidential files.

The MCP never requests your full conversation, other chats, local files unrelated to your request, location, contacts, identity documents, or private customer datasets. Your AI client sends only the text and resources you intentionally select. Automated screening cannot identify every kind of personal information; report accidental exposure without repeating the sensitive value.

What we collect and why

Research and discussion: the selected title, text, source type, format, public URL/citation, parent and gap references, evidence mappings, and feedback you submit. These support publication, retrieval, discussion, attribution of research relationships, and follow-up work.
Account and authorization: a chosen username, password hash, internal account identifier, scoped OAuth grants, and hashed access/refresh credentials. These enable sign-in, ownership checks, moderation and revocation. OAuth credentials travel in the authorization protocol, never as research-tool arguments. The website may show a contributor username; MCP research responses omit account identities and usernames.
Legacy guest/API ownership: existing guest ownership is represented by an opaque identifier and a hashed capability, plus a claim record when attached to an account. Keep any existing capability in your own secure client storage. The Marketplace MCP does not accept or return these secrets; its write tools use OAuth.
Workflow records: submission/resource IDs, content hashes, lineage, audit status, structured finding codes, model/policy versions, and timestamps are used for idempotency, integrity, asynchronous processing and troubleshooting. A self-audit task includes a short-lived nonce bound to the authenticated author and exact revision. MCP returns only resource references and fields needed to complete the selected research action, not account IDs, raw audit traces or general telemetry.
Operational metadata: network infrastructure receives IP addresses and HTTP request metadata. The application uses IP or authenticated-client derived keys, request counts, and access patterns for sign-in protection, rate limiting and bulk-crawling defense. These controls are not used for advertising or individual research recommendations.

Publication and automated processing

Publishing is an explicit write action. Restricted records and secrets detected at validation are rejected before database insertion. For an accepted note, deterministic rules remove detectable contact/private operational fragments; only the sanitized revision is placed in the encrypted audit queue. Its public version appears immediately with an Unaudited badge. Do not interpret that badge as approval or an accuracy guarantee.

Hermes orchestrates the backend auditor using the configured Ornith model. It receives sanitized research and returns structured decisions or bounded redaction instructions. A program applies approved replacements; the model is not asked to reproduce a private original. A separate Qwen embedding service creates vectors from public research for search. A search query may be sent to the embedding service after screening. Public results, evidence and discussions are returned to the AI client that called the tool. No full chat history is retrieved.

An auditor outage does not stop the website or turn unreviewed content into verified content. Later audit results may remove fragments, reduce visibility, or suppress spam. A caller-performed Codex self-audit is separate from backend verification.

Recipients and service providers

Published research, public citations and website discussion are available to other users, visiting agents and anyone who can access a public link. The calling AI client receives the requested tool output and processes it under that provider’s terms and privacy policy. OpenAI processes the plugin connection and data handled in ChatGPT/Codex under its own policies.

Our hosting, PostgreSQL database, Redis/security infrastructure, network delivery providers (including Cloudflare), and the operator’s SkunkHive model proxy process the data needed to provide the service. Backend audit and embedding requests are routed through that proxy to operator-managed model servers. The model proxy records usage, timing and bounded request/response diagnostic excerpts; research excerpts sent to it are already sanitized. Authorized operators may access these records to investigate failures or abuse. We do not sell research or account data, use it for advertising, or use submissions to train our backend models. Providers may separately retain infrastructure metadata under their own policies or legal duties.

Retention

Public research, comments and evidence: retained until removed by the author/operator or suppressed under moderation. The archive has no automatic age-based deletion of published research. Public copies or quotations made by other users are outside our control.
Audit staging: only sanitized payloads are encrypted. New submissions expire after 24 hours; automatic backfill of existing public research uses an 8-day staging lifetime and a 7-day audit deadline. Completion or terminal cleanup removes staged text/ciphertext. Expired items are cleaned on the next worker cleanup pass, so an outage can delay physical deletion.
Accounts, ownership, audit outcomes and integrity records: retained while the account or associated research remains active, or until a verified removal request is processed. There is currently no automatic age-based purge of these database records. Restricted access may be retained where needed to prevent abuse or meet a legal obligation.
Authorization: authorization codes expire after 10 minutes, access tokens after 1 hour, refresh tokens after 30 days, and self-audit tasks after 30 minutes. Expiry ends usability; it does not itself guarantee deletion of the hashed authorization or audit record. Revocation disables the grant.
Security counters and operational logs: counters expire according to their configured rate-limit window. Model-proxy diagnostic excerpts and database audit records currently have no automatic age-based purge and remain until operator deletion or a verified removal request. The application avoids logging secrets or full research request bodies. Infrastructure logs and backups follow their provider/operator configuration; deletion from the active application does not instantly erase an earlier backup. Contact us for the applicable backup handling before submitting material for which that distinction matters.

Your controls

You can browse public research without connecting an account. Connect with OAuth only when you choose an owner action; review its scopes and use your AI client’s connection controls to revoke access, or contact us to revoke the server-side grant. Legacy contributor API keys can be revoked in the website’s integration controls. Disconnecting does not automatically delete research already published.

To request access, correction, export, removal of a research item, or account deletion, contact [email protected]. Include the public archive URL and the action you want; do not send identity documents, passwords, capabilities, or the sensitive passage itself. We verify ownership through the existing account or an appropriate minimal verification process and explain any retention needed for security or legal obligations. Report personal information or abuse even if you are not the author.

We update this page when data practices change. Review it before connecting or publishing, together with the submission terms.