What is being scored
The versioned bank contains 78 questions: 18 about expressed communication style, 15 about expressed affect, 20 about rhetoric and social acts, 11 about hypothetical audience reception, nine classifications, and five evidence or applicability checks. The 53 style, affect, and rhetoric ratings are eligible for exploratory factor analysis (EFA). All 64 numeric ratings—including reception estimates—can appear in correlation and question-map views. Reception estimates stay outside the primary factor model.
The unit is the transcript passage, then the recording. A passage can include Daniel, guests, quoted speech, or playback, so these ratings do not isolate Daniel's personality or establish anyone's identity, inner state, or the truth of a story. The personality and emotion literature informs the vocabulary, but this is a new, unvalidated instrument—not the BFI-2 or a validated mood scale. See Soto and John on the BFI-2 and Russell's affect circumplex.
Passages, gates, and recording scores
Each transcript is divided along segment boundaries into passages of up to 12,000 characters. Long segments are split with offsets retained; every planned passage, including the final short one, is scored. Character limits are not token limits.
Each question has its own rubric and, where needed, an applicability gate. The gate asks whether enough intelligible text exists to rate that family of questions. If not, the score is missing; a clear passage that simply lacks the feature may receive zero. The gates are fallible model judgments, not human quality checks.
Recording ratings are weighted by text characters, not model confidence, and use only passages where that question is applicable. Question coverage and within-recording variation are shown. A classification summarizes scored passages; it is not automatically a whole-video label.
The API receives text, not audio or video. Its score is an expected value over five descriptive rubric levels, rather than a validated psychological unit. Typesafe documents a probability distribution for score levels and a confidence value derived from that same distribution; the latter is not an independent estimate that the answer is correct. See the official Score documentation.
Audience and text-to-speech classifications
The 11 reception items estimate a hypothetical viewer's immediate response. They are not surveys or validated predictions of views, likes, comments, or watch time. A separate engagement extract has 78 recording IDs, but only 50 view counts, 48 like counts and 41 comment counts; timestamps exist for 50 entries, and there is no reliable join to this corpus. It cannot validate audience predictions. Reach, thumbnails, timing, recommendations, and audience composition are also outside the text prompt.
The text-to-speech (TTS) item labels textual evidence as explicit, playback or imitation possibly relevant, no text evidence, or unknown. None of these categories determines whether the audio contains synthetic speech. In a qualitative spot-check of the first 30 chronological receipts, four passages were labeled “explicit” even though the inspected text did not establish synthetic speech; three were from the same livestream. A 14-character partly unintelligible fragment also passed the initial reception gate. This was a targeted spot-check, not a random human-labeled test set, so no accuracy percentage is reported. Audio review is needed for claims about TTS or the actual number of speakers. The text-only limit is consistent with multimodal conversation research such as MELD.
Correlations, factor analysis, and the question map
Correlation and the multidimensional-scaling (MDS) question map use nonconstant numeric ratings with at least 80% coverage. Correlations use one shared complete-case set and are shown only with at least 20 recordings; pair counts and missingness are reported. Associations may reflect wording overlap, shared model errors, format, or chronology, not causation.
EFA includes only eligible questions and recordings with full passage coverage. Each question needs at least 80% recording-level coverage; each included recording must have a usable score for every retained item. The threshold is max(100, five recordings per item)—265 if all 53 questions remain. This is a heuristic, not a guarantee of adequacy. Constant items are removed, and complete-case selection can shrink the sample.
Horn's PCA parallel analysis suggests the factor count; principal-axis common-factor roots and Velicer's MAP are sensitivities. PCA and common-factor methods answer different questions, and zero factors is allowed. Fitted solutions use principal-axis extraction with oblique Promax rotation. They describe covariance in this question set, not “the true dimensions” of a person. See Fabrigar et al. on EFA practice and Horn's parallel analysis.
Recording factor scores are regression-estimated coordinates from standardized questions, not 0–4 rubric ratings. Their orientation is arbitrary: higher is not inherently better or worse, and a coordinate is not a uniquely determined personal trait measurement.
The MDS map uses signed chord distance, √(2 × (1 − r)): positive correlations place items nearer and negative correlations farther apart. Its normalized RMS reconstruction error describes how well a map with a chosen number of dimensions preserves those distances. The explorer reports errors for the selected analysis snapshot and alternative map dimensions. These are geometric fit diagnostics, not confidence in a score or factor and not a rule for selecting the recording-level factor count. The relatedness ordering uses average linkage on signed chord distances; it is only a display order.
Robustness checks and observational associations
The offline supplement compares neighboring factor counts on the same complete-case sample and aligns each alternative to the primary solution. The current fitted counts, sample sizes and diagnostics appear in the explorer. Same-count congruence describes loading similarity; cross-count absolute loading cosine is a separate matching statistic. Neither establishes that factors are equivalent or validates the chosen count. Rank-correlation, higher-text-quality, leave-one-major-genre-out and candidate-deduplicated fits test sensitivity to analysis choices. Their agreement is conditional on this instrument and sample; quality and genre groups come from fallible text-model labels.
The lexical near-duplicate screen uses normalized five-word shingles, a fixed-seed MinHash candidate screen and exact comparison of retained shingle hashes. It forms connected components for an aggressive one-representative-per-component refit; the explorer reports the eligible cohort and candidate counts. Candidates are not confirmed duplicates: quotations, shared scripts, transcription artifacts and transitive chains can link different material. The screen is not exhaustive; differing shingle-set sizes and a 16-recording bucket cap can hide eligible pairs. It does not compare audio or determine narrative equivalence.
Length, date and genre associations use factor scores as outcomes. Adjusted ordinary least-squares estimates use only dated complete records and control for log transcript length, recording year and genre indicators, pooling sparse categories. They describe associations within this archive and depend on uncertain dates and model-assigned genres. No p-values or causal interpretations are reported.
Recording map and similar recordings
The recording map uses principal components analysis of centered and sample-standardized saved factor scores for complete recordings. The explorer reports the variation retained by its displayed axes; omitted variation can make recordings look closer or farther apart than their full profiles suggest. These axes differ from the MDS question map, which places questions rather than recordings.
UMAP is a secondary view of the same complete recordings and standardized factor coordinates as the PCA recording map. The selected analysis reports how many factors are retained. UMAP uses Euclidean distance, 30 neighbors, minimum distance 0.2, two output dimensions and fixed seed 20260930. The map is computed offline with one thread, preserving the existing PCA result and all scores. Its exact package version and settings are retained with the artifact. See the official UMAP parameter guide and reproducibility documentation.
UMAP emphasizes local neighborhoods. Its axes are arbitrary, it has no explained-variance percentage, and visual gaps, cluster sizes and long-range distances should not be read as calibrated differences. The fixed seed makes this saved computation reproducible; it does not establish stability across other seeds or parameter choices. Apparent groups are not verified categories or personality types. Switching views preserves the selected recording and its neighbors, which are always computed from the full factor profiles.
Similar-recording rankings use Euclidean distance across all retained standardized factor coordinates. Correlated coordinates each receive separate equal weight. A nearby recording therefore has a similar model-rated factor profile under this rule; proximity does not establish a shared topic, speaker, event or duplicate source. Inspect the factor comparison and transcript links to interpret a match. Recordings without a complete factor vector remain outside this map and ranking.
Variation within recordings and model ambiguity
The passage-discovery rankings require at least three usable numeric passages for the selected question. They summarize the spread of passage expectations, their minimum and maximum, the last-minus-first expectation, and average absolute change between consecutive scored chunks. Missing chunks break adjacency. These descriptive measures depend on passage length and topic or speaker changes; a large change does not by itself establish an emotional transition.
These new diagnostics use planned passage word counts as weights, with a minimum weight of one. Original recording aggregates use character weights. Group summaries average within-recording variances equally across recordings with at least two usable passages. Distribution entropy and model variance instead pool usable passage distributions with word weights, so long recordings contribute more text to those ambiguity summaries.
Entropy is normalized to 0–1 across the five rubric levels: zero indicates all probability on one level and one indicates a uniform distribution. It describes ambiguity in the returned model distribution, while between-passage variance describes differences in passage expectations. Neither is a correctness estimate. Coverage tables separately count passages rejected by the applicability gate and applicable passages missing a numeric value; these are passage counts, not word fractions.
Residual structure and held-out reconstruction
The residual view subtracts the selected factor model's fitted question correlations from the observed correlations. The fitted common covariance uses the oblique pattern loadings and factor correlations together. Large residual pairs identify relationships that this representation leaves unexplained; they can also reflect related rubric wording or shared model errors.
The factor-count comparison uses three chronological folds with expanding training periods. Known lexical overlap groups spanning a split are removed from both sides. Training data determine means, scales, loadings and uniquenesses. Each question in a later recording is then predicted using its other 52 ratings, with that target value masked. This evaluates reconstruction of model ratings within a recording, rather than forecasting future audience response.
The explorer reports each fold's training and test counts, any overlap exclusions, numerical fit status, reconstruction errors and training-mean baseline. A larger representation can improve reconstruction without establishing that it is the true factor count. The primary representation remains tied to its stated retention procedure.
The 53 questions and complete-case cohort were selected before this comparison and are held fixed. Question selection is not repeated within each fold; unknown overlaps can remain. Fold scores are descriptive, and three folds do not provide asserted confidence intervals or a unique true factor count.
Adjusted time comparisons and candidate shifts
Raw yearly summaries show recording-level distributions and changes between neighboring observed years. Standardized yearly means use the same genre mix across periods: the dated cohort's mix, after pooling genres below the larger of 30 recordings or 2% of that cohort into “other.” Every target stratum must contain at least five recordings in a year. Unsupported standardized means and contrasts remain unavailable; missing strata are not silently dropped or interpolated.
The explorer shows which years support the full target mix and withholds unsupported comparisons. Calendar calculations include only complete recordings with a resolvable month; undated recordings remain available in overall analyses but receive no guessed calendar position. Eligible-month counts and uninterrupted calendar runs are reported alongside candidate segmentation.
Candidate shifts use monthly factor means only when a month has at least ten recordings. Low-count months and calendar gaps divide the sequence into separate runs. Penalized segmentation compares minimum segment lengths of three and six months, two penalty settings, and a sensitivity using nonoverlapping two-month averages. Those averages start at each run's beginning and discard an unmatched final month. Adjacent-month dependence can remain, so agreement between settings is descriptive rather than a significance test.
The page shows recording counts and genre composition around a candidate. Model-assigned genre, uncertain archive dates, missing videos and changing formats can influence a shift. No automatic event attribution or causal conclusion is attached.
Three different kinds of uncertainty
- Model rating range: the central 90% of a character-weighted mixture of passage-level rubric probabilities. It describes spread in the model's passage ratings, not uncertainty in the recording's mean or empirical coverage of a true score. For numeric question trends, the period band averages these per-recording endpoints; it is not a jointly bootstrapped range for the period mean.
- Within-recording variation: how much passage-level expected scores differ inside that recording. This can reflect real changes in topic or voice, as well as text and model noise.
- Factor trend stability range: the 2.5th to 97.5th percentiles of period means across factor refits. Each accepted refit scores the same fixed complete-case cohort, then the scores are averaged within each period. The explorer offers month-cluster resampling and adjacent three-month-block resampling, requesting 500 refits for each method and reporting how many converge properly. Month-cluster resampling includes undated recordings as singleton clusters in the month pool, so their count varies by replicate. The block method holds their count at the observed number per replicate. These bands are available only when category and title filters are clear and the selection is not undated. The range is conditional on this cohort, question bank, retained factor solution and alignment procedure. It is not a sampling confidence interval, population interval, or verified-truth interval.
A passage-mixture range can even exclude the displayed average: a distribution concentrated at level 4 with a small amount at level 3 might have an average of 3.96 and 5th–95th percentile bounds of 4–4. That is not a narrow interval around 3.96.
Month-cluster resampling does not preserve dependence between adjacent months. The three-month sensitivity samples adjacent observed-month blocks, skipping calendar gaps; its circular wraparound is an additional assumption. Both methods refit the model but evaluate period means on the same fixed cohort. The number of factors and rotation are themselves uncertain. A narrow model rating range can still be confidently wrong; no human-labeled calibration set exists for this instrument. For context on calibration, see Guo et al.; calibration methods require independent reference labels.
What a source link establishes
Each analysis is tied to a transcript revision and links to source passages with exact source offsets and, where available, inherited timestamps. Numeric question timelines place points at passage start times and show vertical bars for score levels with nonzero probability; the bars are not confidence intervals. Selecting a point opens that timestamp and transcript revision. A source link establishes which text was processed and where it occurs, not which words caused a rating. The displayed passage is not a question-specific explanation or an account of the model's internal reasoning. ASR timestamps and wording may be wrong. The Typesafe semantic-find example illustrates candidate selection with an explicit no-evidence option.
The explorer reports the selected snapshot's scored and complete-case recording counts. Recordings without a complete usable analysis vector remain browseable with missing values shown. Source revisions, not only corpus compatibility, determine whether scores reflect the latest text. Human validation and audience-outcome validation remain unavailable.