Personality Model Rubrics

Purpose

Big Five facets and Jungian cognitive functions provide compressed priors for the simulator. Concrete traits, speech patterns, reactions, and habitual acts remain primary and override the model when they conflict. Personality ratings do not directly determine interaction-preference settings.

Shared Scale

All ratings use very_low, low, mid, high, and very_high. Levels represent consistency across contexts and persistence when the behavior has a cost, not simple mention frequency. Missing evidence remains empty rather than being labeled mid.

Big Five

Rate the 15 BFI-2 facets from dialogue evidence, then aggregate each dimension by the median of its three facets. A wide facet split is flagged for human review. The five dimensions are openness, conscientiousness, extraversion, agreeableness, and negative emotionality.

Jungian Functions

Rate visible use of Ni, Ne, Si, Se, Ti, Te, Fi, and Fe independently, then order them by level and evidence count. Do not infer function scores from an MBTI label. A noncanonical stack is a review warning, not an automatic correction.

Calibration and HITL

Every rating includes scene evidence. Cast-relative controls test obvious within-show contrasts, and two independent model families produce ratings. Disagreement or missing evidence goes to HITL. No personality item auto-accepts.

The simulator sees only the compressed dimension summary and four-function working stack, after concrete persona traits. Detailed facet ratings, disagreement records, and MBTI cross-checks remain review artifacts.