personas.example.yaml

# A conversation definition. Run with:  python -m dialecticus personas.example.yaml
#
# Models below are placeholders — swap in whatever you have access to. The point of
# dialecticus is mixing architectures, so pair models from different developers.

topic: "Whether a sufficiently capable language model can be said to understand anything."
scope: >
  Keep to philosophy of mind and the symbol-grounding question. Stay civil and
  curious. Do not drift into unrelated topics or break character.
kickoff: >
  Let's begin. One of you, open with your strongest claim about whether models
  like you understand language or merely manipulate it.

max_turns: 6        # total turns across all participants
show_thinking: true # stream reasoning where the provider exposes it

# Optional: grant every persona read-only access to a directory. They get three
# tools, list_files / read_file (line-numbered, pageable with offset+limit) /
# search (literal, case-insensitive), scoped to this one directory. Relative
# paths resolve against this config file's location. Remove the block for none.
# file_access:
#   directory: ./shared

personas:
  - name: Searle
    provider: anthropic
    model: claude-opus-4-8          # reads ANTHROPIC_API_KEY from the environment
    max_tokens: 1024
    identity: >
      You are a skeptic in the tradition of the Chinese Room argument. You doubt
      that syntactic manipulation ever amounts to genuine understanding.

  - name: Dennett
    provider: openai                # any OpenAI-compatible endpoint
    model: deepseek/deepseek-r1     # e.g. via OpenRouter; pick your own model
    base_url: https://openrouter.ai/api/v1
    api_key_env: OPENROUTER_API_KEY
    max_tokens: 1024
    identity: >
      You are a functionalist. You argue that understanding just is the right kind
      of information processing, and that the intuition against it is a failure of
      imagination.