The Need

The gap

* Most training data is static text.

* Most alignment data is preference labels or synthetic constitutions.

* Most alignment data is preference labels or synthetic constitutions.

* Most alignment data is preference labels or synthetic constitutions.

* Most alignment data is preference labels or synthetic constitutions.

* Most alignment data is preference labels or synthetic constitutions.

* Most “values” data is what someone typed, not what they practiced for a week and scored again.

* Most “values” data is what someone typed, not what they practiced for a week and scored again.

* Most “values” data is what someone typed, not what they practiced for a week and scored again.

* Safety and performance teams still lack an auditable chain from insight to behavior.

* Most “values” data is what someone typed, not what they practiced for a week and scored again.

* Most “values” data is what someone typed, not what they practiced for a week and scored again.

What current pipelines miss

* Longitudinal change (the same person, the same value, a week later)

* Longitudinal change (the same person, the same value, a week later)

* Longitudinal change (the same person, the same value, a week later)

* Self-scoring and the reason for the score

* Longitudinal change (the same person, the same value, a week later)

* Longitudinal change (the same person, the same value, a week later)

* Intent vs. action

* Longitudinal change (the same person, the same value, a week later)

* Intent vs. action

* Outcome deltas

* Negative exemplars - when the value was ignored and what followed

* Intent vs. action

* Domain splits (work vs. home vs. stress)

* Negative exemplars - when the value was ignored and what followed

* Negative exemplars - when the value was ignored and what followed

* Negative exemplars - when the value was ignored and what followed

* Negative exemplars - when the value was ignored and what followed

* Negative exemplars - when the value was ignored and what followed