About this researchFixed report · batches 0–101 · 4 Sep 2026
Question

How did the practice move semantically from batch to batch, and which actual works produced that movement?

Method

Work texts are converted into embedding vectors: coordinates in a semantic space where related formulations lie near one another. The report compares centre-of-mass drift with the images.

Finding

Movement is not the same as progress: the trajectory reveals returns, local loops, and regions where varied wording does not change the artistic position.

How to read it

Start with the overall trajectory, select a time region, and compare its points with the work strip. Look for sustained movement across neighbouring batches, not merely the farthest point.

P2 — Drift + Gallery of Works

Phase 2 — Semantic Drift + Gallery of Works

How meanings shifted over time. Drift maps (points + center-of-mass trail) · full-width combinations timeline · band of real works beneath it (hover to enlarge) · current-work panel on pause. Batches: 0–101.

How to read: gray points — the entire type map (context), colored — active in the selected generation range. Line with circle — the trajectory of the center of mass (where the center of meaning “moved” over time), the bold circle is the center of the current range. Modes: Accumulation 0→N · Window [t−W, t] slides · Range A–B is arbitrary.

Current work

Cluster Combinations over Time

Each row is a type, each column is a generation (0–101). Cell color = dominant cluster type in this generation (the palette matches the maps above), brightness = saturation. Read vertically: the column at generation t is the “combination” of dominant motifs across all types at that moment. Yellow frame = current window (moves with the slider/Play). Below — band of real works by generation: hover over the preview to enlarge.
frame = selected window · lower strip = the works themselves (hover → enlarge) · empty = type absent in this generation

Active in the current window

Synchronized with the selection: for each type, which clusters are currently active and how many points they contain (bar ∝ count, color = cluster). This is a detailed slice of the “combination” for the current window [].

Cluster similarity matrices — how to read them

Each cell (i,j) — average cosine between the centers of clusters i and j (how close two family motifs are in meaning). Diagonal — self-comparison (=1, the brightest). Off the diagonal: brighter = clusters closer to each other (motifs intertwine), darker = farther apart (separate). Previously all cells looked equally blue, because in the dense dense_joint space cosines lie in a narrow range (~0.3–0.5); now the color stretched across the actual min–max of each matrix, so the structure is visible. Hover over a cell for the exact value.