Discovery

Fit a model’s trait axes from its samples instead of declaring them — producing intrinsic traits.

Discovery finds a model’s traits from its samples. You supply examples, and discovery fits the axes that best separate or spread them — producing intrinsic traits the model then trains around.

Definition

Discovery fits intrinsic trait axes to a draft model’s samples. An intrinsic trait is one whose axis is its definition — a fixed direction recovered from the data. It contrasts with an extrinsic trait, whose axis is re-derived from labels on every training run. Discovery is how intrinsic traits are created.

Mechanism

Discovery runs on a draft model. It examines the spread of the model’s samples in embedding space, fits candidate axes along the directions that carry the most structure, and attaches them to the model as intrinsic traits. The model then returns to draft with the new traits in place, ready to train.

Re-training freezes an intrinsic trait’s axis and re-fits only the breaks around it — the direction never rotates once discovered. Discovery is asynchronous and draft-only: the model reports busy while it runs, and the traits it produces become part of the model’s frozen definition once trained.

Interpretation

• Discovery suits content that clearly differs in ways you cannot yet name. The fitted axes surface structure already present in the examples, rather than imposing axes from outside.
• Discovered axes are descriptive: each separates the data along some direction, and naming what that direction means is the author’s job.
• Reach for an extrinsic trait when you know the axis and can label for it; reach for discovery when the data should reveal the axis.

Edge cases

• Discovered and authored traits coexist on one model — discover some axes, declare others.
• How many axes discovery proposes is bounded by the effort of the build.

Next

Traits

Intrinsic and extrinsic traits side by side.

Training

How discovered axes are trained and calibrated.