Skip to content

Providers — Mock and Real

Scaffolding for the Agentic Identity Suite's eventual switch from mock to real LLM runs. This is infrastructure, not a live experiment. The existing experiments still use the agents' built-in mock embeddings by design.

What's here

  • base.py — Abstract LLMProvider interface. Every provider implements complete(prompt, system=None) and embed(text).
  • mock_provider.py — Default. Deterministic, fast, no API key required.
  • anthropic_provider.py — Real mode. Calls POST /v1/messages via urllib (no new dependency). Default model: claude-sonnet-4-20250514.
  • factory.py — load_config() and get_provider(cfg). Mock is the default; setting llm.provider: anthropic switches to real mode.

Why this layer exists

The Mirror Problem (see Open Problem 1) and Claim 9 of Emergence Manifesto v1.3 need empirical validation with real language models. The mock embeddings used by lab/agents/three_layer_agent.py and lab/agents/baseline_mirror_agent.py are sufficient for unit-testing the suite's architecture, but not sufficient for the inverse-direction questions the project is structured around — see The Generator Question for the spine.

This provider layer is the seam between the suite's mock-based architecture and its future empirical work. Nothing in the existing experiments has been changed; the suite still runs in mock mode by default.

Running real mode

  1. Set the API key: 
    export ANTHROPIC_API_KEY=sk-ant-...
  2. Edit lab/config.yaml: 
    llm:
  provider: anthropic
  3. Wire the provider into whichever experiment will be updated. The agents are not currently wired through the provider — that is a deliberate next step to be done when Frank decides to take it.

Embeddings note

The Anthropic API does not currently expose a public embeddings endpoint. AnthropicProvider.embed() therefore falls back to the same deterministic hash-based embedding the mock provider uses. The Δ-Kohärenz metric and self-representations work uniformly across providers. When a real embedding service is added (sentence-transformers, OpenAI embeddings, Cohere, etc.), subclass AnthropicProvider and override embed(). No caller will need to change.

Status

  • [DEMONSTRATED] — Mock provider runs deterministically; covered by the existing suite.
  • [INFRASTRUCTURE READY] — Real-mode HTTP path is implemented and import-clean. Not yet wired into the existing experiments.
  • [OPEN PROBLEM] — Whether real-mode Δ-Kohärenz separates trace-memorizers from generator-approximators at statistically significant levels.