Like a food label, but for AI. A plain-language breakdown of every model, every input, every output guardrail. Reviewed and republished alongside every material change.
Founder OS · AI Nutrition Label
REVIEWED MAY 2026 · EDITION: FOUNDER OS COCKPIT
Model & data (LLM)
Model providersAnthropic (primary), Brave Search (web discovery)
Trained on your dataNo — never used to train any public model
Data sent to modelOnly the text of the user's query and any documents the user deliberately attaches
External web accessOnly via Researcher agent (Curie) when explicitly invoked
PII handlingUser identifiers removed during day-to-day processing
Privacy & compliance
FrameworkUK GDPR · Article 28 DPA · DPA 2018
Hosting regionUK · AWS eu-west-2 (London)
Data sold or sharedNever
Analytics sharingNone — no analytics cookies, no third-party trackers
RetentionProject state retained while in use; deleted on account closure
Safety controls
Output framingAll outputs advisory; human review explicit and required
Automated flaggingActive — see Safety Hub for categories
EscalationStatutory duty to escalate risk of danger to self or others
Human oversightMandatory — drafts never auto-send
Risk indicators
Hallucination riskMedium — verify factual claims
PII exposure riskLow — anonymisation by default
Overreliance riskMedium — assistive tool, not decision-maker
Safeguarding riskLow — single-user cockpit, no pupil data
Guardrail summary
All outputs are advisory only. Staff must review and apply professional judgement before acting or sharing. No high-stakes automated decisions. All AI-generated drafts go to Gmail Drafts for human review before send.