The clustering work is the analysis

What it is

Someone photographs the wall and posts “analysis done.” The photo records positions, not judgments. In affinity work, moving a note beside another, splitting a cluster, merging two clusters, each step claims that those observations share a relation. Clustering as analytic work means the argument, the trial placement, the retraction, and the split are the formation of a defensible structure. Analysis is not a later meeting held after the stickers are up.

Why it happens

Proximity is an unnamed proposition. Two notes close together provisionally agree they are more alike than either is to the rest; pulling one away vetoes that proposition. The likeness has to be spoken before the judgment becomes a checkable claim rather than a gesture. If one person silently finishes the wall, or software emits a single clustering pass, the veto path disappears and only an uninterrogable end-state remains. Time pressure demotes clustering to interior design: align for the photograph, fill gaps for tidy columns, and the deliverable replaces the analysis.

Studying it

Process-record a clustering session: for each move, split, and merge, capture the similarity rule offered, the rule rejected, and whether placement later reversed. Compare “argue while placing” with “place separately, then merge” on how defensible the final clusters are—for example, whether a non-participant can reproduce placements from the spoken rules alone. Replay abandoned intermediate clusters to see how many consequential observations sat together only in a rejected structure. Outcomes: rate at which rules were made explicit, reversal count, and information lost between intermediate and final states.

Where it stops holding

A single analyst clustering overnight is still doing analysis; the vetoes just live in one person’s successive memory and cannot be audited from outside. A facilitator who requires everyone to “fill a zone” so that all hands move produces false participation: motion without contested criteria. Algorithmic clustering can suggest candidate neighbors; it cannot stand in for saying why they are alike. Taking a dendrogram as the finding skips the analysis. The neatness a workshop photograph wants often fights the intermediate mess analysis needs to keep.

Applying it

• Do not freeze a cluster until the likeness has been said out loud; mark provisional clusters lightly and allow them to be torn down overnight.
• Assign one person only to log pairs that once sat together and were later split; use the log after the session to check whether the end-state swallowed dissent.
• Do not tidy columns for a photograph first. If an external picture is required, make a separate explanatory figure and keep dispute traces on the working wall.
• Before closing, pick three clusters. A colleague who did not place notes hears only the spoken rule, not the notes, and predicts the contents. Failed predictions mean the grouping is still decoration.

Related

• Same group: Q4.01.1 Cluster from the notes; do not pour them into a ready taxonomy · Q4.01.3 Cluster titles must say what is inside, not stamp a category
• Adjacent: Q4.02 Coding and thematic analysis · Q4.17 Communicating findings and acting on them
• Search terms: clustering as analytic work · affinity diagramming session · analytic memo