Thematic Drift

What this shows#

This heatmap displays thematic drift scores across all 14 monitored categories over time. Thematic drift measures semantic shift in document language compared to a rolling 8-week intra-administration baseline. It detects when the topics and language in a category's documents change relative to recent patterns.

Thematic drift is descriptive, not evaluative. A spike could reflect a legitimate policy pivot, a new legislative session, or a concerning shift in institutional priorities. Concern status is driven separately by AI document review.

How to read the heatmap#

Each cell represents one category in one week. The default view shows z-scores, which measure how far the current week's thematic distance deviates from the rolling 8-week mean.

Blue cells (negative z-scores) — below-average thematic change (unusually stable language)
Neutral cells (z-score near 0) — typical thematic variation
Red cells (positive z-scores) — above-average thematic shift
Dashed border — bootstrap period (fewer than 8 weeks of history available)
Striped cells — no data available for that category-week

Other metrics use a sequential scale (neutral to warm) since they are non-negative values measuring magnitude rather than deviation.

Metrics#

Metric	Range	What it measures
z-Score	-4 to +4	Standard deviations from rolling 8-week mean centroid distance. Default view.
Centroid Distance	0–0.5	Cosine distance between this week’s document centroid and the rolling 8-week centroid. Higher = more semantic shift.
Novel Doc Rate	0–1	Proportion of this week’s documents that exceed the novelty threshold vs. the rolling centroid. Higher = more new topics.
Variance Ratio	0–3	Ratio of this week’s embedding variance to the rolling window’s variance. Values > 1 indicate more diverse document topics.
Cross-Admin Distance	0–0.5	Cosine distance between the current centroid and the Biden-era baseline centroid. Contextual only.

Methodology#

Thematic drift uses embedding-based cosine distance to measure semantic change. Each document is embedded using OpenAI's text-embedding-3-small model. The rolling window compares this week's document centroid to the mean centroid of the prior 8 weeks within the same administration.

The 8-week intra-administration window avoids false positives from normal administration-to-administration policy changes. Cross-admin distance is available separately for researchers comparing across administrations.

For full methodology details, see the methodology page.

Loading thematic data...