Hybrid Mapping and Human-Guided Assistants

This page describes how Ariane supports a hybrid approach:

• Theseus performs automated exploration where possible.
• Human operators perform actions in real software where automation is not safe, reliable, or allowed.
• Both kinds of observations are stored in Atlas as the same kind of graph (states and transitions), with metadata indicating how they were obtained.

The goal is not full autonomy over “every interface”, but a realistic system where:

• Automation covers standards-based, accessibility-friendly UIs.
• Humans cover the hard parts.
• External tools (including AI agents) can rely on Atlas as a unified, machine-readable reference.

---

Why a hybrid approach?

Purely automatic exploration is limited by:

• Anti-bot protections, logins, CAPTCHAs, 2FA, and secure flows.
• Custom-drawn UIs (canvas, 3D, games) without good accessibility metadata.
• High-risk operations (deletion, payments, production changes).
• Combinatorial explosion if you try to explore all possible paths.

On the other hand, purely manual documentation doesn’t give you:

• A consistent data model across apps.
• Programmatic query and pathfinding.
• A shared graph that agents and tools can consume.

The hybrid mode combines both:

• Automation for broad coverage of “normal” UI patterns.
• Human-in-the-loop for exceptional, sensitive, or complex workflows.
• A single graph in Atlas as the reference.

---

Metadata conventions

Hybrid mapping is expressed entirely through existing Atlas records, using conventions in metadata.

Source of observation

On StateRecord.metadata and TransitionRecord.metadata:

• source = "auto" – discovered by automated exploration.
• source = "human" – recorded while a human operator drove the UI.

Examples:

{
  "context_id": "example-web-app-en",
  "discovered_at": "2025-03-01T10:00:00Z",
  "metadata": {
    "source": "human",
    "author": "operator-42",
    "session_id": "2025-03-01T09-59-00Z-session-1"
  },
  "state": { "...": "..." }
}