Skip to content
Ikarus AI LabsGitHub ↗

Research programme

Research

Ikarus AI Labs investigates the technical foundations required for accountable AI-mediated software change.
Established

Supported by published or reproducible evidence.

Observed

Observed in the current implementation or experiment.

Hypothesis

Research proposition under investigation.

Planned

Future research.

Research 01PROTOTYPEHypothesis

From Capability to Evidence-Bounded Agent Authority

How accurately can static analysis construct an evidence-backed model of an AI agent's potentially reachable authority?

Investigates

  • authority taxonomy
  • provenance
  • precision
  • recall
  • unknown authority
  • confidence
  • authority changes
Research 02RESEARCHINGHypothesis

Agent Authority Attestation

How should identity, delegated human authority, tool permissions and operational accountability be represented?

Proposed schema

  • principal
  • agent
  • purpose
  • delegation
  • tool
  • resource
  • access mode
  • constraints
  • environment
  • approval
  • expiry
  • evidence
  • confidence
Research 03RESEARCHINGHypothesis

AI-Mediated Dependency Decisions

Do AI-authored dependency changes differ from human changes in lifecycle, ecosystem and operational risk?

Studies

  • vulnerabilities
  • lifecycle
  • deprecation
  • support
  • ownership
  • distribution
  • identity ambiguity
  • upstream changes
  • evidence confidence

Compares

  1. 1. Human-authored changes
  2. 2. AI-assisted changes
  3. 3. Autonomous AI changes
Research 04FlagshipEXPERIMENTHypothesis

Evidence-Based Policy for AI-Mediated Software Change

Can combined agent-authority and dependency evidence produce safer and more explainable software-change decisions?

Measures

  • unsafe changes prevented
  • false blocks
  • unnecessary reviews
  • decision consistency
  • explainability
  • evidence completeness
  • human effort
  • unknown conditions

Experimental arms

  1. 1. Dependency-only gating
  2. 2. Agent-only gating
  3. 3. Combined evidence-based gating
Research 05IDEAPlanned

Human Authority Boundaries for AI-Mediated Change

Which changes can be delegated safely to AI systems, under what evidence and escalation rules?

Authority levels

  • observe
  • analyse
  • recommend
  • propose change
  • bounded execution
  • high-impact execution

Cloud, Kubernetes and IaC are evaluation environments, not the definition of the research.