zeromodel

Temporal spatial calculus

ZeroModel now has three connected layers:

Dense source table
→ ViewProfile
→ SpatialOptimizer
→ DecisionManifold

The decision manifold is the temporal layer. It treats a run, trace, training history, evaluator stream, or other scored sequence as a series of dense panels.

Concept

A source table can contain many signals at once. A view profile turns one policy lens up. A spatial optimizer derives a metric-weight profile for an explicit top-left mass objective.

A decision manifold asks what happens when those panels evolve over time.

panel_0 → optimized_view_0
panel_1 → optimized_view_1
panel_2 → optimized_view_2
...

Each optimized panel is still a normal VPM artifact. The manifold records how the spatial view changes between adjacent frames.

What is measured

The first implementation measures:

Signal	Meaning
top_left_mass	how much normalized signal sits in the inspection region
metric_weights	which metric lens the optimizer derived for that frame
row_order	which source rows became most salient
column_order	which metric columns became most salient
curvature	weighted change across mass, weights, row order, and column order
inflection_indices	frames after the largest adjacent-frame changes

This is intentionally geometric, not semantic. It does not say why the change happened. It surfaces where inspection should begin.

Relation to the old “40 steps” idea

The older idea was that a very large sequence of panels could be reduced to a small number of meaningful inspection steps. This module is the first reproducible version of that idea:

many dense panels
→ optimized spatial views
→ curvature over time
→ candidate inspection frames

The repository still should not claim that it can find the universal best forty steps for arbitrary world data. The safe claim is narrower:

ZeroModel can summarize a sequence of dense scored panels as a deterministic decision manifold and surface frames with large spatial-view changes.

What remains to prove

Before making stronger claims, we need:

1. generated temporal fixtures with known change points;
2. benchmark comparisons against scalar dashboards or random frame sampling;
3. real exported sequences from training, critics, RAG verification, or Writer traces;
4. user studies or task metrics showing inspection improvement.

Why this matters

This completes the first foundation:

static dense representation
→ multiple policy views
→ optimized view profiles
→ temporal decision manifolds

Applications such as hallucination inspection, training progress, critic traces, and Writer evaluator streams can now sit on top of that foundation.

This site is open source. Improve this page.