# Aradia Briefs — Sovereign AI Evidence Library

> Aradia Briefs: documented, cited, decision-shaped proof of why sovereign on-premises AI differs from public cloud AI. Court rulings, incident reports, and buying decisions — each with primary sources, cross-checked against our continuous Risk Monitor.

`>` Briefs / Sovereign AI Evidence Library — https://aradia.com/briefs/

Every claim we make has a brief behind it. Marketing says "private," "secure," "sovereign." We don't argue — we show the court rulings, the incident reports, and the invoices. If a cloud-AI failure has a name and a citation, it lives in this library.

## The Library

Brief pages, newest first (full brief text lives at each page URL, not duplicated here):

- **1991 to Agentic Systems — Turnkey Private AI & B2A Partner Launch** (/briefs/agentic-systems-launch/) — Corporate Release / Launch Brief / 2026-11-05. One of the first 50 websites on the internet (Est. 1991) reinvents itself as an Agentic Systems company — shipping turnkey NVIDIA DGX appliances for 1-to-200-person firms and the world's first 10% B2A Partner Program.
- **LLMjacking — Google's $1M Cloud Bill Warning** (/briefs/llmjacking/) — Security / Threat-Intel Brief / 2026-11-02. Google/Mandiant: stolen LLM API keys burn $100K+ per day and up to $1M per organization in hours. The breach starts as a billing email. The fix is an inference surface with no key to steal.
- **The Digital Employee — Own It, Don't Rent It** (/briefs/digital-employee/) — Workforce / Trend Brief / 2026-10-25. 2027 is called the year of the digital employee. Gartner: 15% of work decisions autonomous by 2028 — 40% of agentic projects canceled by 2027. The gap is ownership, and the fix is $15,125 flat.
- **OpenAI Subscription Treadmill — more cost, less compute** (/briefs/subscription-treadmill/) — AI Economics / Platform Pricing / 2026-09-29. DevDay 2026: $500/mo Pro 500 tier lands while the $200 plan reopens with half its compute allowance. Platform-controlled pricing, and the $15,125 one-timer fix.
- **The $500 Million Uncapped Seat Burn** (/briefs/uncapped-seat-burn/) — Enterprise / Cost Disaster / 2026-05-29. An enterprise client's unconstrained Claude licenses burned half a billion dollars in 30 days — no caps, no alarm, no error. The largest cloud-token billing failure in internet history, and the meter that stops it.
- **U.S. v. Heppner — public cloud AI waives privilege** (/briefs/law-firms/) — Law Firms / Court Ruling / 2026-02-17. Judge Rakoff, S.D.N.Y.: client facts pasted into Claude or ChatGPT are voluntary disclosure to a commercial vendor. What the ruling breaks, and the $15,125 air-gapped fix.
- **Air Canada — you answer for what your bot says** (/briefs/chatbot-liability/) — Consumer / Court Ruling / 2024-02-15. A chatbot hallucinated a bereavement refund; the airline paid. 'The model is a separate information source' failed as a defense. Every promise your agents make is your company's word.
- **Silent Model Drift — 97.6% in March, 2.4% in June** (/briefs/model-drift/) — Engineering / Research / 2023-07-18. Stanford/Berkeley measured GPT-4's stable-task accuracy collapse in one quarter with no errors thrown. Frozen models on owned metal are the only reproducible answer.
- More briefs land here as the Risk Monitor verifies new incidents. Machine-readable incident feed: https://aradia.com/api/incidents.json

## How a Brief Is Different From a Blog Post

- `//` **CITED, NOT OPINED** — Every brief traces to a primary source: a docket number, a ruling, an SEC filing, a published incident report. No "experts say." No vibes.
- `//` **MONITORED, NOT STALE** — Our continuous monitoring watches for the failures we warn about. When one is verified in the wild, it gets its own entry in the incident feed and its own brief.
- `//` **DECISION-SHAPED** — A brief ends at the decision: buy the Spark if…, the cloud works if…. Readable in five minutes by the person who signs the purchase order.

Related: [/solution/](/solution/) (Where It Matters sector cards) · [/pricing/](/pricing/) · [/agents/](/agents/)