Transcript research explorer
Explore the scored archive
Rank recordings by a dimension, explore their similarities, and follow patterns over time. Each recording links to its source passages.
- Scored recordings
- 4,582 of 4,582 in this release
- Scored passages
- 15,877 source-linked transcript chunks
- Scoring questions
- 78 ratings and classifications
- Exploratory factors
- 9 fitted across 4,522 recordings
Loading the score release…
Jump to an analysis
Videos by dimension
These filters apply to this table and Scores over time. Other views use their own selectors and analysis cohorts.
| Recording | Date | Recording category | Score / outcome | Range / option probabilities | Coverage |
|---|
Selected recording evidence
Scores over time
Change dimension or filtersView all period values and recording mix
| Period | Mean | Range (see trend method) | Recordings | Mean coverage | Recording categories |
|---|
Trend method, uncertainty and date coverage
Explore recording profiles
Find recordings with similar combinations of factor scores and inspect the passages behind their ratings.
Search narrows the recording selector. The map continues to display the complete analysis cohort.
Drag to pan; hold Ctrl or Alt while scrolling to zoom. Focus the map for arrow-key panning, +/− zoom, or 0 to reset. Each map remembers its own view. Viewing controls never move saved points or change scores.
Select a recording above or click near a visible map point. Crowded locations open a title-and-date chooser, ordered by screen distance; zoom in to separate dots. Similar-recording lists always use all retained standardized factor dimensions, not distances in either map. Changing the map preserves your selection and neighbors. Visual groups do not establish personality types or verified categories.
Map methods and projection details
Correlated factors receive separate equal weight in the full-profile distances. The maps are views of those saved ratings, not new factor models.
PCA variance retained
UMAP settings
UMAP is unavailable for this release; PCA and full-profile neighbors remain available.
UMAP is stochastic and parameter-dependent. The saved map uses a fixed seed and one computation thread for reproducibility. It is not a clustering test or a stability interval. See the UMAP parameter guide and reproducibility notes.
Inspect the ten profiles farthest from the cohort center
These are unusual combinations of saved scores. Their position may reflect subject matter, recording conditions, score ambiguity, or content; distance alone does not establish the explanation.
Question relationships
Explore how the scored questions relate within the analysis cohort. These are estimated relationships among transcript-text ratings; they do not establish traits or causal effects.
Expand all pair estimates for the selected question
Expand the clustered correlation heatmap
Average-linkage clustering arranges the questions using signed chord distance √(2 × (1 − r)). This order is a display aid; it does not establish factors or validate questionnaire groupings. Focus the question selector above to see its strongest positive and negative relationships with paired sample sizes.
Resize with the controls, then scroll to inspect. Focus the heatmap for keyboard scrolling; +/− change size, F fits width and 0 resets. Values and question order remain unchanged.
Expand map-fit diagnostics
Question map
Drag to pan; Ctrl or Alt + scroll to zoom. Focus the map for arrow keys, +/− and 0 to reset. Select a question with a click or Enter/Space; dragging never changes its saved position. Ordinary scrolling moves the page.
Inside recordings: variation and ambiguity
Find recordings whose average conceals changes across passages, then inspect the passages with the lowest and highest expected scores.
Expand recording rankings and passage links
Means and passage variability use word-count weights, differing from the original recording averages’ character weights. Rankings include every recording with at least three scored passages. Unequal chunk lengths, missing passages, and topic changes affect these descriptive comparisons.
Ambiguity and coverage
How to read these diagnostics
Model entropy ranges from 0 (one rating receives all probability) to 1 (equal probability across all five ratings). Model variance describes the spread of the returned rating distribution in squared score units. Passage variance describes differences among passage expected scores within recordings. Ambiguity does not measure rating accuracy.
Entropy and model variance pool available distributions using passage word counts. Passage variance averages recording variances equally, using recordings with at least two scored passages. Coverage counts scored passages out of all passages. Insufficient means the applicability gate prevented rating; missing means a passage was applicable but lacked a numeric value. Category and transcript quality groups use the model’s recording-level classifications.
Time, genre and candidate shifts
Compare the observed archive with a common genre mix, inspect cohort distributions, and explore changes in monthly scores.
Waiting for temporal comparisons…
Raw and common-mix yearly trends
Raw means use each year’s recordings. Adjusted means use a fixed genre mix from the dated cohort. Estimates require at least five recordings in every target stratum; unsupported years have gaps. These are descriptive means.
Yearly values and composition
Compare two cohorts
Whiskers show the central 80%, blue boxes the middle 50%, the vertical line the median, and orange dots the mean of whole-recording scores. The effect divides the second-minus-first mean difference by the pooled within-cohort standard deviation. These are raw cohort comparisons; composition and date can differ. Yearly common-mix estimates are shown separately above.
Explore candidate change dates and sensitivity
Markers are candidate boundaries in scored archive patterns. Agreement counts the six tested settings at the exact month; nearby boundaries can differ. Review genre composition on both sides. The method assigns no cause or life event to a boundary.
Question correlations
Open the full matrix to inspect exact pair values.
Open full correlation matrix (loads on demand)
Adjust cell size, then scroll the matrix horizontally or vertically. Exact values and keyboard-selectable cells remain available at every size.
Exploratory factor loadings
Loadings are oblique pattern coefficients, not correlations; depending on the fitted model, they may exceed −1 to +1. Intervals and factor labels follow the release method and are not proof of a universal or “true” personality dimension.
Analysis diagnostics
How much does the factor model explain?
Predicting a withheld question
Compare factor counts by hiding one question at a time in later recordings and predicting it from the other question scores. Lower error means better reconstruction.
This comparison uses three temporal folds. It measures reconstruction of transcript ratings; it does not establish a true factor count or predict audience reception. No confidence intervals are asserted from three folds.
Expand reconstruction errors and baseline values
Expand temporal folds, overlap checks and failed fits
Expand the reconstruction method and limitations
Relationships left unexplained
Residuals compare observed relationships with the correlations reproduced by the saved oblique factor model. Positive values indicate relationships stronger than the model reproduces; negative values indicate weaker relationships. Large pairs can reveal overlapping questions or structure left unexplained.
Expand the largest 15 residual pairs
Expand the residual heatmap
Orange: positive residual; blue: negative residual. Intensity uses the largest absolute residual in this model. Diagonals are gray. The selector above provides exact values. Resize with the controls, then scroll to inspect; zoom does not change any values.
How stable are these patterns?
Compare plausible factor counts, refits on different subsets, and possible near duplicates. This checks whether patterns are sensitive to analysis choices; it does not establish a single true factor structure.
Waiting for the validated analysis supplement…
Alternative factor counts
Convergence and admissibility matter alongside residual fit. A failed fit is shown for context and should not be treated as a usable solution.
| Factors | Fit status | Converged | Sample | Iterations | Heywood cases | Residual RMS |
|---|
Compare same factor count across refits
Each refit is aligned to the primary solution. Congruence compares loading patterns; score rank correlation compares recording order. Both are descriptive and can vary by axis. The full Pearson row reproduces the primary calculation; it is a baseline, not independent validation.
View factor score rank correlations
Spearman correlations compare the ordering of recordings between each refit and the primary solution.
Near-duplicate transcript candidates
Length, date, and genre associations
These are descriptive associations. They can reflect recording format, topic, transcript quality, or other differences; they do not show that one feature caused a score.