Private AI · sealed-benchmark acceptance
Ask your documents.
Get a cited answer.
Corpus Aegis turns your governed documents into a private retrieval system and, only when it measurably helps, an adapted model. Every answer names the pages it came from. Then we prove it on a benchmark your experts approved and a custodian sealed.
1 artifact built · 9 planned · no sealed certificate yet
Question
Can a contractor open the internal wiki?
Answer
Yes, but only with a sponsor's approval1 and only after the onboarding checklist is signed2.
Passage 1
Contractor accounts may be granted read access to internal systems when a named employee sponsors the request.Access Control Policy · page 4
Passage 2
Access is enabled after the contractor completes and signs the onboarding checklist.Contractor Onboarding Guide · page 2
Sources
Illustrative layout with placeholder content. This is not output from the shipped system.
How it works
Try it, read the numbers, then test it sealed.
01
Try the system
Sign in and ask artifact 01 a real question. Read the passages it found and the pages it cites.
02
Read the measurements
The performance page shows what has been measured so far, how it was measured, and where it falls short.
03
Sealed acceptance
Before any result is called a pass, the system is tested on questions kept sealed from development. That test has not been run yet.
Read the development measurements. They are published, but they are not acceptance evidence.
Hardware targets
We will test a system on your hardware, not on a slide.
Every quoted number will belong to a named profile: model, runtime, context limit, memory ceiling, and number of users. Change the hardware and the system is tested again. No profile below has been tested yet.
Search keeps your documents current. Tuning teaches the model how to answer. We will not tell you a model has stored your documents.
| Profile | Model sizes we plan for | Status |
|---|---|---|
| 16GB Apple Silicon | 1B–4B, 4-bit | Not yet tested |
| 24GB Apple Silicon | 7B–9B | Not yet tested |
| 36–64GB workstation | 14B–32B, quantized | Not yet tested |
| 96–192GB GPU rack | 70B–120B or replicas | Not yet tested |
Proof library
Ten planned. One built.
Enterprise policy questions
Five file formats, one answer
Contracts & clauses
Clause lookup across real contracts
Equipment manuals
Tables, scans and troubleshooting
Financial & compliance rules
Questions that span tables and text
Ownership
Delivered means delivered.
Take the processed corpus, the retrieval database, the permitted adapters or weights, the configuration, the evaluation harness, and the integration API inside your own boundary, including air-gapped. Your data is never reused for another customer or for general model training.