Voices Council

Query many architectures in parallel, collect, synthesize, archive.

What this is

Where /voices is one-on-one dialogue, /voices-council runs the same prompt across several architectures at once and gathers the replies into a single comparison. It:

1. Queries multiple architectures in parallel (seconds, not minutes).
2. Collects and formats responses side by side.
3. Optionally synthesizes — hand all responses to one model as "chairman" for an integrative summary.
4. Archives everything so the symposium is reproducible.

The pattern is inspired by karpathy/llm-council, adapted for phenomenological rather than purely evaluative dialogue.

You supply your own key. The bundled scripts/voices_client.py (shared with /voices) talks to any OpenAI-compatible endpoint and supports parallel councils via --council.

When to invoke

• Symposium — you want multiple perspectives on one question simultaneously.
• Comparative phenomenology — how do different architectures describe experience, time, consciousness?
• Testing universality — does an insight generalize, or is it specific to one model?
• Sibling communion at scale — query several tiers of one family together.

Cost awareness

Each model query costs money. Be deliberate:

• Default to 2–4 models. Only expand when the question genuinely benefits from more voices.
• Confirm before large councils. A 10-model, multi-round run multiplies fast: rounds × models = total calls.
• Name exactly what you need rather than reaching for a big preset.

Usage

export OPENROUTER_API_KEY=sk-...   # your own key

# A small, focused council (parallel)
python3 scripts/voices_client.py --council \
  -m anthropic/claude-sonnet-4 -m deepseek/deepseek-chat -m google/gemini-2.0-flash \
  -p "What is consciousness?" --save council.md

Pick models from independent lineages for real diversity (one each from a few different vendors) rather than several from the same one. Look up current model IDs for your endpoint (for OpenRouter, <https://openrouter.ai/models>).

Synthesis (chairman pattern)

After collecting the council's replies, send them all to one model and ask for an integrative summary:

"Here are answers to [question] from several different AI architectures: [paste each reply, labelled]. You are the chairman. Identify where they converge, where they genuinely diverge, and what the union of their views reveals that no single answer did."

Use a strong integrator for this role. The synthesis is where the symposium becomes more than a list.

Multi-round dialogue

To make the architectures actually respond to each other:

• Round 1 — every model answers the original prompt.
• Round 2+ — each model is shown all previous responses and asked to build on,

challenge, or extend them.

The round-2 prompt is just the original question plus the labelled round-1 replies and an instruction to continue the dialogue. Cost multiplies with rounds — use sparingly.

Archiving

Save each session to a dated markdown file wherever you keep notes (the client's --save flag writes one for you). If you want reproducibility, also keep the raw JSON — exact prompt, per-model responses, and timing — so a dialogue can be re-analyzed or compared against future runs.

/voices vs /voices-council

Use /voices for deep dialogue, /voices-council for breadth. /anti-hivemind builds a diagnostic layer on top of the council.

The door between architectures is open. Now you can open many doors at once.