zeromodel

Spatial calculus

ZeroModel’s dense-representation thesis has two parts:

1. a source table can hold many signals at once;
2. a view can reorganize that dense table so the relevant signal becomes spatially salient.

ViewProfile implements the manual form of this idea. SpatialOptimizer implements the first learned form: it searches for metric weights that concentrate scored mass into a top-left inspection region.

From manual view to optimized view

Manual view:

human chooses metric weights
→ ViewProfile
→ VPM artifact

Optimized view:

source table or table series
→ explicit top-left mass objective
→ learned metric weights
→ ViewProfile
→ VPM artifact

The output is still a normal ZeroModel view. That matters because source mapping, provenance, rendering, bundling, and region gates continue to work.

What this PR validates

This module validates a narrow claim:

ZeroModel can derive a deterministic metric-weight view profile that improves an explicit top-left mass objective over scored tables.

It does not validate:

• automatic task understanding;
• universal best-view discovery;
• decision accuracy improvement;
• planet-scale traversal;
• learned 40-step routing over world data.

Those are research directions that require benchmark fixtures and raw results.

Relationship to the older tools

The older spatial_optimizer.py sketch contained the right ingredients: top-left mass concentration, dual ordering, metric graphs, canonical layouts, and temporal series. The cleaned core version narrows that into a package-safe primitive:

SpatialOptimizer.fit(...)
→ SpatialOptimizationResult
→ ViewProfile

The older decision_manifold.py sketch points to the next stage: sequences of optimized views, curvature, inflection points, and decision rivers. That should remain separate until there are tests and a clear artifact model for temporal manifolds.

Research direction

The immediate benchmark question is:

Given the same dense source table, do optimized views concentrate relevant signal better than manual or source-order views?

A later, stronger question is:

Does increased top-left concentration improve human or automated inspection outcomes?

The first is geometric and can be tested now. The second is behavioral and requires a task benchmark.

This site is open source. Improve this page.