Docs

Architecture

ZeroModel is the deterministic artifact layer between raw AI traces and human inspection.

ZeroModel starts from a simple premise:

The output of an AI system should not disappear into a score, a paragraph, or a dashboard. It should become an artifact.

The artifact kernel

A Visual Policy Map is a deterministic spatial view over scored data. It carries:

• raw values,
• stable row identifiers,
• stable metric identifiers,
• normalization rules,
• layout recipes,
• source mapping,
• provenance payloads,
• deterministic identity.

The artifact is not the model. It is the thing you inspect after a model, evaluator, tracker, verifier, or policy system has produced signals.

Core pipeline

source system
  → scored rows
  → ScoreTable
  → LayoutRecipe or ViewProfile
  → VPMArtifact
  → render / bundle / compare / gate / inspect

Agent trajectory pipeline

agent run
  → steps, tool calls, evidence, memory, claims
  → grounding / policy / cost / uncertainty scores
  → Visual Policy Map
  → top-left inspection region
  → reviewer focuses on the highest-signal cells first

Training pipeline

tracker export
  → checkpoints
  → train / held-out / regression / stability / efficiency metrics
  → training progress VPM
  → best checkpoint and warning surface

Research pipeline

paper or benchmark output
  → claims, spans, trace steps, labels, scores
  → deterministic artifact
  → reproducible inspection
  → compare against raw tables and scalar dashboards

ZeroModel is intentionally modular. The artifact remains auditable while consumers around the artifact can grow: PHOS packing, view profiles, spatial optimization, decision manifolds, learning traces, training progress, critic evidence, and edge gates.