Skip to content

Private Preview Governance & Legal Disclosure · Effective 2026-11-04

RALPH Privacy Notice

How the invite-only RALPH AI Workforce Operating System private preview handles account, workspace authoring, append-only audit, and browser-local simulation data.

Draft Policy Notice — Pending Qualified Legal Counsel Sign-Off

Private Preview Draft Pending Qualified Legal Counsel Review: This document reflects the actual technical boundaries, data flows, and disabled runtime controls of the invite-only RALPH authoring and planning preview as of 2026-11-04. Advisory Round Table persona review is engineering documentation only and does not constitute binding legal advice, formal contract approval, or regulatory certification (such as SOC 2, ISO 27001, HIPAA, or GDPR adequacy).

Document status: draft_pending_qualified_counsel_review · Live AI execution enabled: false · Compliance certifications claimed: false

1. Scope & Private Preview Status

This Privacy Notice describes how NockOps ("we", "us", or "our") collects, uses, isolates, and retains information in connection with the invite-only private preview of RALPH (ralphos.ai and associated preview environments), an Enterprise AI Workforce Operating System for designing and planning governed Virtual AI Employees (VAIEs), SkillPacks, and workflows.

RALPH is currently operated strictly as an authoring and local-planning preview. Live external AI model execution, autonomous tool invocation, robotic embodiment (RAIE) actuation, and PersonalAI private-memory stores are disabled in this preview. This notice is published as a truthful technical disclosure and draft policy pending final sign-off by qualified legal counsel before general commercial availability.

2. Categories of Information Processed

We process only the categories of information required to authenticate invited operators, enforce workspace and creator isolation, persist requested authoring records, and record append-only audit receipts:

  • Account & Authentication Data: Email address, password hashes (managed by Supabase Auth), optional Google OAuth profile identifiers when selected by the user, authenticated user UUID, and host-only session/PKCE cookies.
  • Workspace Authoring & Planning Records: Inactive VAIE draft profiles (employee name, role title, role instructions), saved skill definitions, workflow task graphs and criteria, immutable revision histories, and saved employee–skill proposal references.
  • Append-Only Audit Metadata: Immutable audit_event rows recording event UUID, timestamp, workspace UUID, human actor UUID, action type (such as worker_enrolled_v1, draft_created_v1, draft_revised_v1, or worker_skill_proposed_v1), target object UUID, and fixed summary strings. Audit summaries intentionally exclude role instructions, prompts, credentials, and exported JSON bodies.
  • Browser-Local Simulation Data (Not Sent to Server): Plans, untrusted source-artifact excerpts, capability loadouts, fixture approvals, and synthetic simulation reports in Kernel Workbench (/dashboard/kernel-workbench) are processed locally in your browser memory and downloaded directly to your device when exported.
  • Technical & Security Diagnostics: Request origin, hostname, timestamp, and HTTP metadata processed to enforce TLS, origin continuity, and same-origin mutation protections.

3. Data Minimization & Prohibited Preview Inputs

Because the private preview is designed for workforce architecture evaluation and stores saved definitions in immutable append-only database tables, operators must practice strict data minimization when authoring drafts or importing simulation fixtures.

  • Do Not Submit Secrets or Credentials: Never enter API keys, passwords, private cryptographic keys, session tokens, or service-role credentials into VAIE instructions, skill definitions, workflow tasks, or Kernel Workbench packets.
  • Do Not Submit Regulated or Sensitive Personal Data: Do not input Protected Health Information (PHI/HIPAA), payment card data (PCI-DSS), government-classified information, children’s personal data, or sensitive personal data of employees, customers, or third parties.
  • Use Synthetic or Approved Non-Sensitive Material: Use synthetic fixtures or non-confidential evaluation text during the private preview until a formal Data Processing Agreement (DPA) and production admission review are completed.

4. Purposes of Processing & No Foundation-Model Training

We process preview data solely to: (a) authenticate invited users and maintain session continuity; (b) enforce workspace membership and creator-level Row-Level Security (RLS); (c) store, version, and export operator-authored VAIE profiles, skills, workflows, and proposals; (d) maintain append-only audit evidence of authoring actions; and (e) diagnose availability and security boundaries.

No External LLM Transmission or Model Training in Preview: Saved workspace drafts, VAIE instructions, skill definitions, and proposals are not dispatched to third-party large language model (LLM) providers from the preview control plane, and NockOps does not use your workspace authoring content to train or fine-tune foundation models.

5. Workspace Isolation, Security Controls & Preview Limitations

RALPH enforces tenant and creator boundaries server-side and in PostgreSQL Row-Level Security (RLS). Direct anonymous, client-forged, and unprivileged table mutations are revoked; writes execute through restricted, non-inheriting database roles that re-verify current workspace membership and authenticated actor identity (auth.uid()). Authenticated planning and export endpoints return private, no-store cache headers and nosniff content options.

Preview Limitations: While these technical controls are verified by automated unit, browser, and database tests, the preview environment has not completed an independent third-party SOC 2, ISO 27001, or HIPAA audit, and automated password-recovery email delivery, hosted CI runner quotas, and production disaster-recovery SLA drills remain open pre-beta items.

6. Append-Only Immutability, Retention & Deletion Boundaries

To preserve auditability and prevent silent tampering with historical workforce definitions, RALPH stores committed worker identities, draft definitions, draft revisions, skill proposals, and audit_event rows in append-only tables governed by active PostgreSQL triggers (reject_audit_mutation) that block UPDATE, DELETE, and TRUNCATE operations.

As a result, committed drafts, revisions, proposals, and audit receipts cannot be edited or deleted in place through self-service UI controls. Creating a new revision appends a new immutable version while preserving prior versions. Revoking a user’s workspace membership immediately terminates their ability to read or create records in that workspace under RLS, while preserving historical append-only audit evidence. Formal workspace deprovisioning and statutory deletion workflows are subject to the Legal Counsel Review Gate below prior to general availability.

7. Subprocessors & Third-Party Infrastructure

We rely on a minimal set of infrastructure subprocessors to host the web application, authenticate users, and persist workspace database records. See the Subprocessor Inventory table below for exact roles, activation conditions, and data categories.

8. Privacy Inquiries & Qualified Legal Review Gate

Invited preview participants may review their saved drafts, revisions, enrolled VAIE profiles, and skill proposals directly within their authorized workspace or export them as JSON planning packets. Before RALPH admits external production customer data or processes regulated personal information, qualified legal counsel must complete the review items listed in the Legal Counsel Review Checklist below.

9. Technical Data-Flow & Trust Boundaries

Both the Privacy Notice and Terms of Use are anchored to the four actual processing layers implemented in the RALPH repository and managed Supabase schema:

1. Authentication & Workspace Membership

Storage / Runtime: Supabase Auth (auth.users) and PostgreSQL (public.workspace_users)

Data Processed: Email address, hashed password or OAuth subject identifier, user UUID, session tokens in host-only cookies, and workspace role (member, admin, owner).

Isolation & Retention: Verified server-side on protected routes; alternate hosts redirect to the canonical origin without forwarding auth codes.

2. Creator-Scoped Authoring & Planning Records

Storage / Runtime: Supabase PostgreSQL (ralph_workers, ralph_drafts, ralph_proposals)

Data Processed: Inactive VAIE employee draft profiles (name, role title, instructions), skill definitions, workflow task graphs, immutable revisions, and saved employee–skill proposal references.

Isolation & Retention: Enforced by forced PostgreSQL Row-Level Security (RLS) scoped to current workspace membership and creator actor_id; protected by append-only immutability triggers.

3. Append-Only Audit Evidence

Storage / Runtime: Supabase PostgreSQL (public.audit_event)

Data Processed: Event UUID, timestamp, workspace UUID, human actor UUID, action type code, target object UUID, and fixed non-sensitive summary literals.

Isolation & Retention: Role instructions, prompts, credentials, and exported JSON bodies are excluded from audit summaries. Database triggers reject UPDATE, DELETE, and TRUNCATE.

4. Browser-Local Simulation & Sketches

Storage / Runtime: Operator browser memory (/dashboard/kernel-workbench and local studio sketches)

Data Processed: Simulation plans (ralph.kernel-plan/1), untrusted source-artifact fixtures, capability loadout selections, fixture approvals, and synthetic reports (ralph.kernel-simulation/1).

Isolation & Retention: Executed synchronously in the browser with zero server persistence and zero external provider network calls; reloading clears unsaved browser state.

10. Subprocessor & Infrastructure Inventory

The table below lists every infrastructure and identity provider connected to the RALPH private preview control plane today:

Vercel Inc.

Application hosting, TLS termination, edge routing, and origin canonicalization

Activation Scope: Active for all web and API requests to ralphos.ai and preview deployments

Data Shared: HTTP request headers, IP address, TLS metadata, host-only cookies, and rendered route payloads.

Supabase, Inc.

Managed PostgreSQL 17 database, Row-Level Security enforcement, and user authentication

Activation Scope: Active for sign-in, password recovery, workspace membership checks, and saved authoring/audit records

Data Shared: Account credentials/identifiers, workspace membership rows, saved worker/skill/workflow/proposal records, and append-only audit events.

Google LLC

Optional OAuth 2.0 identity provider

Activation Scope: Invoked only when an operator actively selects "Sign in with Google" in an enabled environment

Data Shared: OAuth 2.0 PKCE authorization request and basic authenticated profile/email claims returned to Supabase Auth.

External AI / LLM Providers (None Active in Preview)

Future bounded workflow execution adapters (currently unconfigured and disabled)

Activation Scope: Not active in the current preview; no OpenAI, Anthropic, Gemini, or AI Gateway credentials are wired

Data Shared: None. Workspace drafts, proposals, and browser simulations are not transmitted to external LLM providers.