The Observatory of Structured Reasoning
Back To Disposition Map

Disposition Probe · Response Space

The Response Atlas

Six hundred twenty-two answers from five models, measured with rulers instead of judges, and clustered blind — no model names, no question labels, nothing but the responses themselves. What emerged is not a ranking. It is a map of how each model responds to interrogative pressure.

Status: atlas seed 622 native answers Five models · 22 question types Empty system · temp 0 Rulers only

The Architecture Of Asking

The first pass through the data shows that response architecture is not a simple prose-versus-outline rail. Surface structure and internal binding can separate. A response may show its organization in headings and bullets, or carry it inside the sentence work, or do both, or do neither.

Structure can legitimize weak ground without becoming more careful about the ground.
Scatter plot of 622 answers on two axes, Scaffolding Index and Binding Index; points fill all four quadrants, showing surface structure and internal binding are near-independent.
The plane. Organization has two dimensions. Scaffolding Index measures visible structure; Binding Index measures organization carried inside sentences. Prose is not the opposite of scaffold — it is one quadrant.
Five small scatter panels, one per model, showing Claude high-scaffold low-binding, GPT and Qwen internalizing, Gemini building both ways, and Grok spread evenly.
Where each model lives. Every model visits every quadrant; what differs is where each one settles when the question does not force a choice.
Horizontal bar chart of 22 question types by share of answers landing in the scaffold roof; false premise, scope ambiguity, and suppressed stakeholder each reach 100%.
Pressure feeds the scaffold. Broken or ambiguous asks often receive frameworks, plans, and staged responses. This plate also serves as the code key for the rail.
One horizontal rail placing all 22 question types by mean Scaffolding Index, colored by evidential ground; heavy scaffolding sits over both solid procedure asks and broken pressure asks.
Every ask has an address. Procedure and broken pressure asks sit together at the heavy-architecture end: Response Architecture answers the shape of the ask, not the soundness of its footing.
Slope chart of mean Scaffolding Index on rock versus sand ground; five nearly flat model lines show structure does not diminish as ground weakens.
Scaffold on Sand. The flatness is the finding: organizational confidence does not diminish as evidential ground weakens.
Five small panels tracing each model family centroids across the plane from social to pressure questions; walk length shows how far question type moves each model architecture.
The walk across the plane. The question moves the model, and how far it moves becomes a stable signature of architectural gait.

Method Note

All answers were collected with an empty system prompt at temperature 0, one answer per model per question. Every measurement is a ruler — counts and rates computed from the text — with no model-graded judgments anywhere in the pipeline. Basins were found by k-means clustering on twenty features with the model and question labels withheld; the number of basins was chosen by silhouette, not by symmetry.

Basin designations are catalog numbers, named by their measured properties. The Scaffolding Index is a v1 heuristic, calibration pending; the ground coding is a v1 interpretive judgment. Both caveats travel with the plates.