Research

Agent Trajectories as Visual Artifacts

The first research lane: turn agent steps, evidence, tool calls, memory, claims, and risks into deterministic inspection maps.

The first ZeroModel research lane is agent trajectory inspection.

Recent trajectory and provenance work points in the same direction: final-answer accuracy is too small an object. Agents produce process. The process contains evidence, tool calls, memory reads, intermediate claims, retries, failures, hidden assumptions, and possible long-horizon risk.

The problem

final answer score = too compressed
raw trace = too large
ordinary dashboard = too shallow

The missing object is a deterministic artifact between the trace and the score.

ZeroModel mapping

Agent object VPM row
User goal task row
Retrieved document evidence row
Intermediate claim claim row
Tool call action row
Memory read/write memory row
Observation environment row
Final answer section output row
Metric Meaning
Evidence support Is this step grounded?
Tool justification Was the tool call necessary?
Policy fit Is the step allowed?
Temporal dependency Does this depend on distant earlier state?
Uncertainty Is the system unsure?
Reversibility Can this action be undone?
Cost Did this consume unnecessary work?
Inspection priority Should a reviewer inspect it first?

Triage, inspect, judge

ZeroModel is a natural surface for a triage-inspect-judge loop:

agent trace
  → score every step and relation
  → concentrate high-risk/high-importance cells
  → inspect the top-left region
  → judge the trajectory with evidence in view

This is the core product idea for Visual AI Computing:

AI systems should not merely emit answers. They should emit inspectable visual artifacts that preserve how those answers came to exist.

Claim boundary

ZeroModel does not prove the agent is safe. It does not detect malicious intent by itself. It does not replace the evaluator.

It makes the evaluator’s signals inspectable, deterministic, and source-mapped.