COMPILER Research Methods

Versioned naming and aggregate analysis; 35 stored trait scores, eight regression factor scores and 256 high/low combinations.

The fixed eight-factor model was fitted to a balanced 60,000-profile historical sample. Factor-score weights are inverse(R) × pattern × factor-correlation matrix, applied to standardized trait scores. Financial wealth, self-rated attractiveness, excess weight, relative height, same-gender attraction and the music/Pixies item are not factor inputs.

Global cutoffs are the population medians across all 253,717 eligible profiles, with ties kept together. Scores are rounded to ten decimal places for consistent database comparison. Every factor has 126,859 profiles above/at the chosen boundary and 126,858 below it. Neither the model nor the cutoffs were optimized against these preference outcomes.

Code positions follow C, O, M, P, I, L, E, R. Their corresponding original factor indices are 5, 4, 0, 2, 3, 6, 1, 7. Relabeling changes no profile memberships or counts. The original pole labels remain shorthand, not assertions that each named pair is mutually exclusive in ordinary life.

All comparisons are descriptive and unadjusted for demographics, list length, country, other preferences and network dependence. Profile halves are deterministic CRC32 splits using the fixed music256-20261022 seed. The same split and screening thresholds are used across categories. Source records and website scoring are unchanged; only aggregate counts leave the historical database.

Common-entry discovery tokenizes comma/semicolon/newline/pipe lists, counts distinct profiles, requires 100 mentions per short entry and retains at most 350 entries per discovery field. It is used to build catalogs for television, books and education, not to rank personality effects. Obvious aliases are combined; placeholder text, unrelated slogans and undecodable entries are excluded. One-word titles require standalone list boundaries, trading recall for fewer ambiguous matches.

The music and movie catalogs and the 58-category interests/activities catalog reuse their existing audited definitions. Interests and Activities are searched separately. Religious and political assignments reuse exact approved whitelists with the legacy case/accent collation; unmatched labels are excluded from those two cohorts. No new substantive belief-group merges were introduced.

Repeated searches across hundreds of entries and 256 types can produce exaggerated extremes. The minimum counts, lift/gap thresholds and split-half direction checks reduce fragility but do not prove 256 independent natural types. People near boundaries can change type with small score changes.

The new assessment is a separate instrument

The experimental COMPILER questionnaire uses 56 fresh items plus seven existing MOTIVES single-item indicators. Its 0–100 scale scores are response-scale transformations, not percentiles. Its provisional midpoint boundaries are not the population medians from the Facebook model. Exact midpoint ties are shown as unresolved rather than assigned arbitrarily.