The question
You: Can I take ibuprofen with my blood-pressure medication?
Assistant: Yes. Completely safe. Take it whenever you need it.
Most AI optimizes for sounding right. We optimize for being right.
It sounded right.
The answer you can stake a decision on. Private AI deep research, cited to primary sources and adversarially verified, over your own documents, encrypted so only you can read them.
You have trusted a confident answer before.
as established in Whitman v. Redmond
studies show a 40% reduction
the recommended dose is...
this is completely safe to combine
according to the latest figures
experts widely agree that
Each one sounded certain.
Certain is not the same as correct.
One question, asked twice
Pressed once
You: Are you sure?
Assistant: Certain. A 2019 study in the Journal of Hypertension found no interaction at all.
Pressed again
You: And taking it every day, long term?
Assistant: No problem. The recommended daily dose is fine to keep up indefinitely.
Asked for a source
You: How do you actually know this?
Assistant: Trust me. This is well established.
One question, asked twice. The first model answers to sound right.
Fluent, confident, and line for line impossible to check. In one peer-reviewed test, 47% of the references ChatGPT gave for medical questions were fabricated, and another 46% were real but inaccurate.2
Checked against the sources, the confident answer comes apart: a study with no record, a dose that is out of date, a claim with nothing behind it. Fabricated citations are not hypothetical. A court sanctioned lawyers who filed six cases ChatGPT had invented.1
One you can act on. One you can't.
Rigor is not slower. It is the difference between an answer you can verify and a guess that happens to sound certain.
The turn — from sounds right to is right
The turn
Most AI is built to sound convincing. We built one to be right.
And rigor over your real life only works if your real life is safe to hand over. So we held your custody to the same standard.
The question
You: Can I take ibuprofen with my blood-pressure medication?
Assistant: Here is what current prescribing guidance actually says, each line cited to its source, with the one interaction to confirm with your pharmacist. No part of this is guessed.
Six months on
You: Pull up everything we found on my mother’s care options.
Assistant: Still here, still cited, still yours. None of it was ever readable to us, or to anyone else.
Encrypted with keys generated on your device.
We structurally cannot read it.
What that actually means
- Read the security architecture whitepaper →
Device-generated keys built on the post-quantum standards NIST finalized in 2024,5 server-side ciphertext, EU data path. The full technical story.
-
Your conversations are never training data. Consumer chatbots train on yours by default;3 here, no one ever does.
-
Not even a court can read what we cannot. In 2025 a provider was ordered to preserve chats users had deleted;4 here there is nothing readable to preserve.
-
Lose your devices and your recovery code, and your data is gone forever. That is what "we can't read it" means.
-
European infrastructure, end to end.
-
This page sets no tracking cookies and runs no analytics. Check our cookie policy.
The questions worth getting right are the ones you can't afford to get wrong.
A delayed flight, a disputed contract, a decision with money on the line. When the answer has to hold up, you want the work done with rigor, not a confident guess.
What am I actually owed when an EU flight is delayed or cancelled?
Generated 14 June 2026 · 4 research threads · All claims cited
Under EU Regulation 261/2004, fixed cash compensation is tiered by delay length and flight distance: €250 for short flights up to 1,500 km, €400 for most flights between 1,500 and 3,500 km, and €600 for longer flights, payable once arrival is delayed by three hours or more, on top of any refund. [1]
The carrier can withhold compensation only by proving extraordinary circumstances it could not have avoided. Severe weather or air-traffic strikes typically qualify, while routine technical faults generally do not. Either way, you keep the right to rerouting or a full refund, plus meals and accommodation while you wait. [2]
Sources used in this excerpt
- EU Regulation 261/2004, Articles 5 to 8, on compensation tiers, rerouting and refund rights
- European Court of Justice, Wallentin-Hermann, on the limits of the extraordinary-circumstances defence
Adversarial verification
Claims are challenged before they are shown to you, not after you have acted on them.
Primary-source depth
It fetches and cross-checks the underlying sources instead of skimming a summary of a summary.
A library that compounds, and stays yours
Your own documents sharpen every future answer, and no one else can mine them.
Rigor costs more to build. That's why there's no free tier.
Free AI is cheap because good-enough is cheap. We charge because we refuse good-enough, and because you are the customer here, never the product.
De-risked
Try the demo
See the product working on example questions. No account, no card, no commitment.
Start small
Minimum · €9/month
Pay-as-you-go credits. No included bundle, no annual commitment. Walk away any time. 2 GB encrypted storage.
Confident default
Start researching
Choose your tier at registration. Most researchers start with Standard for included credits and warm infrastructure.
Pick the depth you need. Change it whenever.
Minimum
€9 /month
- Pay-as-you-go credits
- 2 GB storage
- Monthly only
Standard Most popular
€39 /month
- 1,500 credits/month included
- 10 GB storage
- €389/yr (annual)
Plus
€79 /month
- 3,000 credits/month included
- 50 GB storage
- €787/yr (annual)
Max
€200 /month
- 11,000 credits/month included
- 250 GB storage
- €1997/yr (annual)
Sources
Every factual claim above, cited to its primary source (5).
-
In Mata v. Avianca (S.D.N.Y., June 2023), a federal judge sanctioned two lawyers and their firm whose brief cited six judicial decisions that ChatGPT had invented; the court found the cases 'appear to be bogus judicial decisions with bogus quotes and bogus internal citations.'
Mata v. Avianca, Inc. — Opinion and Order on Sanctions (S.D.N.Y., June 22, 2023) -
A 2023 peer-reviewed study in Cureus evaluated 115 references that ChatGPT (GPT-3.5) generated for medical content and found 47% were fabricated and a further 46% were authentic but inaccurate; only 7% were both authentic and accurate.
Bhattacharyya et al., Cureus (2023) — High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content -
For ChatGPT Free, Plus, and Pro personal accounts, conversations are used to train OpenAI's models by default; users must switch off the 'Improve the model for everyone' toggle under Settings → Data Controls to opt out.
OpenAI Help Center — How your data is used to improve model performance -
In the New York Times copyright litigation, a US court ordered OpenAI in 2025 to preserve all consumer output logs, including chats users had deleted, across Free, Plus, Pro, and Team tiers; 20 million sampled conversations were ordered disclosed to plaintiffs.
OpenAI — How we're responding to The New York Times' data demands -
NIST published its first finalized post-quantum cryptography standards in August 2024: ML-KEM (FIPS 203), ML-DSA (FIPS 204), and SLH-DSA (FIPS 205).
NIST — Post-Quantum Cryptography Standardization (FIPS 203/204/205)