ZeroModel’s core claim is not that a VPM is a heatmap. The claim is that an AI system can externalize dense state into a deterministic artifact that supports many views without regenerating the evidence.
An image contains trees, grass, people, cars, roads, and sky at the same time. Looking for people does not create a new image; it changes what becomes salient. Looking for trees does the same. The source remains the same.
A ZeroModel source table works the same way:
rows = items, claims, checkpoints, traces, frames, candidates
columns = metrics, scores, risks, evidence signals, policy signals
values = signal intensity
A ViewProfile is the policy lens:
same dense source table
+ people view
= people-salient VPM
same dense source table
+ risk view
= risk-salient VPM
same dense source table
+ training-progress view
= checkpoint-progress VPM
The evidence is not regenerated. The dense field is reorganized.
Most LLM interfaces are serial and sparse from the outside. They emit tokens, tool calls, retrieved chunks, summaries, or individual judgments. ZeroModel makes the intermediate state dense and external:
many rows
x many metrics
x deterministic layout
x source mapping
x provenance
This enables operations that should not require another model call:
A view profile is a small policy over metric salience.
ViewProfile(
name="hallucination-risk",
metric_weights={
"semantic_drift": 1.0,
"evidence_gap": 0.9,
"citation_gap": 0.8,
"verifiability": -0.6,
},
)
Positive weights mean high values matter. Negative weights mean low values matter.
The resulting VPM makes the policy spatial: items matching that policy move toward the inspection region.
The older spatial-calculus tools point beyond manual view profiles. SpatialOptimizer learns metric weights to maximize top-left decision mass, computes metric interaction graphs, and derives canonical layouts. DecisionManifold extends VPMs over time with curvature, inflection points, critical regions, and decision-flow paths.
Those tools should be treated as the next research layer. ViewProfile is the core abstraction they need to target.
The first research question is:
Can dense multi-view artifacts improve human inspection and control of AI system state compared with sparse serial summaries or raw score tables?
This is broader than hallucination detection. Hallucination risk, training progress, learning traces, critic outputs, anomaly detection, and search triage are all views over dense scored state.
Valid wording:
ZeroModel can build multiple deterministic policy views from the same dense source table while preserving source mapping and provenance.
Invalid wording until separately proven:
ZeroModel automatically learns the best view for every task.
That requires optimized view learning, benchmark tasks, and human inspection studies.