Privacy notice

Last updated · 3 July 2026

1. What we collect from you

You give us

• Your messages — the prompts you send in the chat surface.
• Files you upload — PDFs, Word documents, plain-text files, images, and videos you attach to a message.
• URLs you paste — Noesis fetches the content of URLs you include in a message and uses the extracted text as context for the answer.
• Feedback signals — 👍 / 👎 ratings and the optional comment when you down-vote an answer.
• Sign-in identifiers — your authentication via Firebase Auth (Google or email). We see your email address and a stable user ID.

Generated automatically as you use Noesis

• The responses Noesis produces for you, including the per-LLM drafts and the consensus answer.
• The audit signals attached to each response: cross-LLM agreement scores, critical-thinking scores, document-grounding verdicts, citation-verification results, bias / legal flags, security signals, AI-likelihood scores, threshold-gate decisions.
• The brain neurons created from your conversations. Noesis maintains a per-tenant learning brain that captures generalizations across sessions to improve future answers.
• Operational telemetry — latency per pipeline stage, cost per LLM call, error rates. Used to maintain the service.

2. What we do with it

Provide and improve the service

Your prompt is routed through Noesis's 4-LLM convergence pipeline (see section 3 below for who that involves). The audit pipeline runs over the consensus answer to produce the chip row you see in the UI. Your interactions inform the brain's learning over time — this is what makes Noesis improve at your domain across sessions.

Comply with applicable law

Where the EU AI Act applies, Noesis writes a record-keeping entry for each chat turn under Article 12: a Merkle-chained log entry containing hashed identifiers (your user ID and session ID are SHA-256-hashed with a server-side salt before they enter the log), the audit chip outcomes, model versions used, duration, and cost. Default retention is 7 years, as specified by Article 12.

Protect users

The uncrossable thresholds gate detects requests for legal verdicts, medical diagnoses, financial advice for an individual, or mental-health crisis intervention. On those requests Noesis switches to "informational mode" — it explains the underlying law / clinical evidence / standard of care and recommends qualified professional help. It does not render the verdict itself.

3. Who we share data with

Operating Noesis requires sending portions of your data to a small set of third parties. The table below lists every external service Noesis calls today and what data reaches it.

Each of these providers has its own privacy policy. Anthropic, OpenAI, Google, and Moonshot describe how they use prompts they receive — most notably whether prompts are used to train their models. Noesis's contracts with these providers, where available, request that your data NOT be used to train their public models. We will publish the specific contract status of each provider in this notice as those agreements are finalized.

Noesis does not sell your data, run ads, or share data with advertising networks. Noesis does not have an advertising business model.

4. How long we keep it

EU AI Act Article 12 records (chat turns)

Default: 7 years, as required by Article 12. Entries store hashed identifiers and audit metadata, not your raw prompt text. You can request deletion of the underlying identifiers; the Merkle-chained log preserves the chain integrity but the row becomes anonymized.

Raw chat sessions (your prompt + the response)

Retained while you can still see them in your "Recents" list. We are finalizing the retention TTL policy (target: opt-in only with a 30-day default TTL for raw prompt text, separate from the audit records). Until that ships, raw chat content lives on the same disk as the rest of the application and is removed on offboarding.

Brain neurons (learning across sessions)

Retained indefinitely unless you ask us to delete them. A brain neuron is a condensed generalization derived from one or more conversations (e.g., "user prefers Swiss restraint in UX work") — not a verbatim transcript.

Uploaded files

Stored per session. Detached uploads (files you stop sending with new prompts) remain on disk so you can re-attach them later, until you explicitly delete them.

Offboarding

If you ask to leave Noesis, we delete your raw chat content, your uploaded files, your brain neurons, and your account record within 30 days. EU AI Act Article 12 records are anonymized (hashed identifiers retained for chain integrity, raw identifiers purged).

5. Your rights and controls

Where the GDPR or equivalent law applies, you have the right to:

• Access the data we hold about you (GDPR Article 15)
• Correct inaccurate data (Article 16)
• Delete data we hold about you, subject to legal-retention exceptions (Article 17)
• Export your conversations and brain neurons in a portable format (Article 20)
• Object to processing — including model improvement based on your data (Article 21)
• Withdraw consent at any time where consent is the legal basis

To exercise any of these rights, write to us at the contact below. During the private pre-launch we respond manually; a self-service panel is planned before general availability.

Specific opt-outs available today

• Stop the deepfake probe — uploaded images can be uploaded with the probe disabled at the operator level. Currently global; per-user opt-out is planned.
• Stop reverse image search — this is user-triggered only. The image is never sent to Google Vision unless you click the "Find similar source images" button inside the AI-likelihood chip.
• Stop URL fetching — paste URLs without expecting them to be fetched; the chat handler treats URLs you mark as inline text only.

6. Security

Noesis runs on Cloudflare-fronted infrastructure with TLS everywhere. Authentication is via Firebase. Application data lives on disk in a Docker container with restart-on-failure. Backups are not yet automated; this is part of the cloud-deployability EPIC currently in flight (target: 2026 Q3).

The cryptographic Merkle-chain on the EU AI Act records is the only surface where we guarantee tamper-evidence today. The rest of the operational data is protected by standard application-level controls.

7. Contact

For privacy questions, requests, or to exercise any right described above, write to PlatformAI at the address shown at the bottom of the homepage. We respond within 30 days, typically sooner during the pre-launch period.

8. Changes to this notice

We update this notice when our practices change. The "Last updated" date at the top reflects the most recent material change. For significant updates we will surface a notice in the chat surface itself before the change takes effect.