[Proposed] How to define 'behavior' (definitional groundwork for behavior vectors)
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Captured from Todoist via my-goat autoprocess. Conceptual / definitional task.
Question: How to define "behavior" precisely, in a way that supports a "behavior vector" (cf. companion capture #eval-good-bad-cases-as-vector). Definitional groundwork for the persona/steering line.
Why it matters: The project leans on "persona vectors," "the assistant axis," and (proposed) "behavior vectors." These are used loosely. A crisp operational definition of behavior determines: what gets contrasted to extract a direction (CAA-style), what an eval measures, and whether "behavior" is distinguishable from "persona" / "trait" / "capability."
Candidate definitions to compare:
- Behavior = distribution over outputs conditioned on a prompt class (functional/behavioral).
- Behavior = an elicitable direction in activation space whose addition shifts outputs along an axis (mechanistic, CAA / persona-vector style).
- Behavior = a measurable score on an eval rubric (operational, ties to the 44-prompt misalignment rubric).
Open questions:
- Is "behavior" the same granularity as a persona trait, or finer (a single elicited tendency)?
- How does it relate to the persona-vector decomposition (identity / style / capability) in §I task #1?
- Does the definition need to be eval-anchored to be useful for the "good/bad cases → vector" pipeline?
Done condition: 1-page note settling on a working definition, cross-linked to the behavior-vector experiment and the persona-vector decomposition task.
Links: research_ideas.md §I (Persona Geometry, persona-vector decomposition), §7 (User Modeling). Companion to the eval-good/bad-cases → behavior-vector capture.
Source: my-goat queue file 2026-05-28T19-00-08_todoist-6gjhpXPPhhCVxwcv_how-to-define-behavior.md (Todoist id 6gjhpXPPhhCVxwcv).