Pattern-to-Preference Rubrics

Purpose

These rubrics map repeated dialogue evidence to the canonical settings in PA_Interaction_Preference. They infer what PA behavior a user is likely to prefer; they do not rate the character’s own style and do not redefine a setting.

Evidence Rules

1. Prefer direct reactions to another person’s behavior: correction, approval, refusal, frustration, or repeated boundary-setting.
2. Require repeated or high-cost evidence before assigning high confidence.
3. Keep work and personal evidence separate.
4. Record counterevidence and context qualifiers.
5. When the transcript lacks a reliable proxy, send the cell to HITL rather than manufacturing certainty.

Reliable Transcript Proxies

The strongest direct proxies usually cover tone and formality, verbosity, emotional engagement, guidance level, reasoning visibility, uncertainty expression, autonomy, proactive outreach, task expansion, information elicitation, and topic management.

Process visibility, solution breadth, and capability boundary are PA-specific behaviors with weaker character-dialogue proxies. They require extra caution, explicit evidence, an accepted seed, or human adjudication.

Interpretation Pattern

For each candidate segment, record the observed behavior, the user’s reaction, the direction implied for the PA, the context, and whether the evidence conflicts with other scenes. A setting is promoted only after the evidence supports the functional definition in the taxonomy. Surface features such as sentence length or politeness are evidence, not ground truth by themselves.

Removed Historical Labels

Output structure and representation style were removed for weak independent discriminability. Source attribution was merged into reasoning visibility. Memory and privacy was removed as an active interaction-preference cell. These labels must not appear in accepted 2 × 14 matrices or final probes.