SKIP TO MAIN CONTENT
BASED

Research

Ground Contact.

Every answer is a claim with a return path. That is the mechanism the whole company is built around — and the standard every model, pipeline, and finetune is held to.

A model response is useful only when its path can be inspected, its sources checked, and its uncertainty seen. Ground Contact is the loop that makes inspection routine instead of exceptional.

The stack does not claim certainty. It builds conditions where error can be found faster, claims can be audited, and better answers can survive contact with evidence. If the system cannot show how it reached an answer, the answer stays provisional.

A single thread of light descending from a glowing point to a dark plane, touching the ground

The loop.

01

Claim

Output is a claim, not a vibe. It can be checked because it can be stated.

02

Source

Every claim carries its path back to the material it stands on.

03

Test

Output is scored against task-specific harnesses before it ships.

04

Revise

Error gets found fast and corrected in the open. No answer is final.

What we measure.

Truth-seeking is treated as an engineering problem with measurement constraints. Six metrics, applied to every model before it joins the roster and every pipeline before it ships.

Source-trace accuracy

Can the answer show where it came from?

Claim-support rate

How many claims survive contact with their sources?

Calibration error

Does the model know what it does not know?

Refusal quality

Does it decline the right things for the right reasons?

Retrieval precision

Does the pipeline surface the right material?

Reasoning consistency

Does the path hold up on a second pass?

Thesis / 01

Truth by design.

Truth by design is not a claim that any model owns the truth. It is a design constraint: the interface, the agents, the datasets, and the release process must make verification easier than belief.

Thesis / 02

Epistemic sovereignty.

The human remains the sovereign evaluator. A tool that hides its path trains dependence; a tool that exposes sources, uncertainty, and reasoning trains judgment. We build the second kind.

Etymology

Base
Foundation, ground, reality contact.
D
Data, derived signal, distributed verification.
Based
Foundation in data. Truth from first principles.

Access

Work with the lab.

API keys, scoped builds, and finetunes all start the same way: a letter. Tell us what you are building. We read every one.