Persistent AI Systems
How memory and derived state change over time; how systems represent provenance, withdrawal, correction, and recovery.
Research
Reliability after the first response.
Velorin studies the conditions under which AI systems retain justified state, respect authority boundaries, and recover when accumulated information becomes incomplete or wrong.
Research program
Current areas of investigation.
The program is deliberately narrow: build inspectable artifacts, define what evidence would change the conclusion, and preserve results that do not support the preferred mechanism.
Agent Reliability & Authorization
How agents behave when exposed capabilities exceed legitimate authority, including release decisions and unsafe action selection.
Human–AI Systems
How judgment, initiative, memory, and oversight can be distributed without obscuring responsibility or degrading human agency.
Project 01
Velorin Continuity Testbed
The first published research record examines omitted lineage and coverage-conditioned release through a frozen internal comparison and a separate limited external filter evaluation.
VCTExecutable case study
Preserved technical record · 2026
Omitted lineage and release selectivity
The strongest supported result is local: on a constructed synthetic population with truthful coverage evidence, environment-local support evaluation preserved more valid releases than two tested conservative controls while matching their observed invalid-release count. A reduced external filter did not reproduce a safety improvement.
520 structured coordinatesFrozen evidenceNegative result preserved
Open the project record