Cornerstone · the knowledge layer
Everything your organisation knows, as facts your AI can be trusted with.
Cornerstone turns your documents into typed, cited facts and keeps them current. Every answer your AI gives can be traced to its source, and when the facts are not there, it says so. It is the layer beneath every AI your organisation uses.
What changes
What your AI works from.
Today your AI reads whatever it can reach: the whole archive, in whatever state it is in, and it sounds just as sure when it is wrong. Cornerstone changes what it is given.
What your AI gets
A fact, with its source attached.
Each one carries where it came from, who wrote it, when, and what it applies to. An answer traces back to what your organisation actually wrote.
A clear no, when there is nothing.
On a question your documents cannot answer, your AI says so instead of assembling a plausible guess from the nearest match.
The current version, every time.
When the business changes its mind, the old fact is superseded and dated. What your AI reads is what is true now.
The facts a question needs, and no more.
Not the whole archive dragged through the model each time. That is why the same token budget goes further as your knowledge grows.
What stops happening
- Your AI reading the whole archive and still dropping the qualifier that mattered.
- A confident answer from a policy two revisions out of date.
- A rule written for one region applied to a case in another.
- Nobody able to say where an answer came from when it is challenged.
- The token bill climbing with every document you add.
The buyer already knows no system is never wrong. What blocks a rollout is the error nobody can see. Cornerstone makes the answer checkable, and says when there is none.
What it looks like
A fact, not a paragraph.
Cornerstone does not hold your documents. It holds what they say, as facts your AI can act on.
Typed
A rule, a decision, a process, a concept or a reference. Your AI receives a claim it can act on rather than a page it has to interpret.
Cited
Source, author, date and scope travel with every fact, so any answer can be traced back to what your organisation wrote.
Current
A changed fact supersedes the old one. Both stay on the record, and only the current one is read.
Rule
EU client data stays in EU regions
- Source
- Data residency policy, §2
- Author
- Legal
- Date
- 12 Mar 2026
- Applies to
- EU clients
Decision
Frankfurt is the default EU region
- Source
- Architecture review, 4 Jun
- Author
- Architecture
- Date
- 4 Jun 2026
- Applies to
- New EU tenants
Superseded Dublin is the default EU region · 2024
Reference
Data residency addendum, §4
- Source
- Contract template v3
- Author
- Commercial
- Date
- 20 Aug 2026
- Applies to
- Client contracts
How it works
Knowledge goes in once. Every answer reads from it.
01
Your documents go in.
Foundry reads what you already have, the documents, meetings, tickets and decisions, and turns it into cited facts on your own hardware.
02
Cornerstone holds them as facts.
Each is a rule, a decision, a process, a concept or a reference, with its source, author, date and scope. Nothing is silently overwritten; a changed fact supersedes the old one and both stay on the record.
03
Your AI works from them.
It receives the facts a question needs, cites them in the answer, and tells you when there is no answer to give. What your people and their AIs decide flows back in through the Coordinator, so the knowledge stays current as the work happens.
Every part of the platform deploys inside your boundary, and Foundry runs on the machines you already own. How it is deployed and secured.
What it does for you
Answers you can check, at a cost you can plan.
Measured, not asserted. We built every method we could and ran them all the same way on the same 2,201 documents; the numbers are ours, and so are the questions it got wrong.
Cited, every time.
In our benchmark Cornerstone found the right document 96.7% of the time, against 85% for the best other method we tested, and every answer arrived with its source.
It tells you when it does not know.
On 49 questions our test corpus could not answer, Cornerstone declined 41. Every other method we tested guessed.
Current, never silently overwritten.
A superseded fact stays on the record, dated. Your AI reads what is true now, and you can see what was true then.
More work from the same tokens.
Around 98% fewer tokens than reading whole documents, with the cost per question held within 1.1× where file search swung 232×.
How you reach it
Through the AI you already use.
Cornerstone sits beneath your AI rather than in front of your people. Workbench connects the assistant in VS Code or Cursor; the platform works with Claude, Codex, ChatGPT, Gemini, Copilot, or a model you run yourself.
Give your AI something it can be trusted with.
Point Cornerstone at documents you know, or book a call and we will show it working on yours.