The corpus infrastructure layer

Gradient descent
on your documentation.

Foundry captures every correction your AI agents make, turns it into a regression test, and measures whether your CLAUDE.md is actually getting better.

The re-alignment tax

Every session, your agents re-learn what you already taught them.

15%
How often GPT-4o follows 10 simultaneous rules
30%
Accuracy loss from bloated conversation history
Times you have corrected the same mistake
The agentic team

Five agents. One session. Every message.

The Cartographer routes context. The Librarian reconciles signal. The Herald watches the wire. The Warden guards the domain. The Artificer does the work. Together they turn a capable model into an agent that remembers your project.

the foundry CARTOGRAPHER

CARTOGRAPHER

Routes context

Reads the topology map. Decides what context the message needs. Sees everything, modifies nothing.

The loop

Capture. Classify. Route. Verify. Improve. Prevent regression.

Every step connected. One signal flows all the way through.

Capture

Every correction, question, redirect — recorded, attributed, timestamped.

Classify

Correction. Convention. Taste. Security. ADR. The model is already good at this.

Route

Each classified signal goes to its durable home — conventions doc, memory, guard, fixture.

Verify

Runs the change against the fixture suite. Does it actually produce better output?

Improve

Verified changes land in the corpus. Every entry traces back to the signal that motivated it.

Prevent regression

The fixture that validated the improvement becomes a permanent test. Knowledge accumulates without fragility.

Score your corpus.

The Oracle is in closed beta. Join the waitlist and we'll run a free corpus audit on your repo when a slot opens.