> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mutual-dissent.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Core Concepts

> Fan-out, reflection, synthesis — how multi-model debates work

## Panel

The panel is the set of models that participate in a debate. The default panel is `claude-opus`, `gpt`, `gemini`, `grok`, and `nemotron`. Each alias maps to a specific model:

| Alias         | Model                             |
| ------------- | --------------------------------- |
| `claude-opus` | anthropic/claude-opus-4.6         |
| `claude`      | anthropic/claude-sonnet-4-6       |
| `gpt`         | openai/gpt-5.4                    |
| `gemini`      | google/gemini-3.1-pro-preview     |
| `grok`        | x-ai/grok-4.20-beta               |
| `nemotron`    | nvidia/nemotron-3-super-120b-a12b |

You can pass any subset of aliases or full OpenRouter model IDs to `--panel`. Duplicate aliases are supported for single-model multi-agent debates:

```bash theme={null}
dissent ask "Your query" --panel claude,claude,claude
```

## Agent Identity

When a panel contains duplicate model aliases, each agent receives a unique identity by appending a numeric suffix: `claude-1`, `claude-2`, `claude-3`. When aliases are unique, they are used as-is with no suffix.

Agent identities are used for:

* **Reflection pairing** — each agent sees its own previous response and the others' responses
* **Display labels** — CLI and web UI show the agent identity, not the raw alias
* **Stats tracking** — per-agent token and cost breakdown alongside per-model rollups
* **Transcript logging** — the `agent_id` field in `ModelResponse` persists in JSON transcripts

The `display_label` property on `ModelResponse` returns the agent identity if set, falling back to the model alias for backward compatibility with older transcripts.

## Rounds

A debate runs in rounds. Round 0 is the **initial round** — every panel model receives the query and responds independently. Rounds 1+ are **reflection rounds** — each model sees its own previous response and all other models' previous responses, then reflects and refines.

The default is 1 reflection round. The maximum is 3. More rounds increase cost and latency but can surface deeper disagreements.

## Synthesis

After all reflection rounds complete, the **synthesizer** model receives the full debate transcript and distills it into a final answer. The synthesizer defaults to `claude-opus` but can be any model alias. The synthesizer participates in the debate panel as well unless excluded explicitly.

## Transcripts

Every debate produces a `DebateTranscript` — a complete structured record saved as JSON to `~/.mutual-dissent/transcripts/`. Filename format: `YYYY-MM-DD_shortid.json`.

Each transcript includes:

* The original query
* Panel composition and synthesizer
* All rounds with per-model responses, token counts, latency, and routing decisions
* The synthesized final answer
* Aggregate stats: total tokens, per-model cost breakdown, total cost in USD

## Roles

Each `ModelResponse` carries a `role` field:

| Role         | Round number | Description                           |
| ------------ | ------------ | ------------------------------------- |
| `initial`    | 0            | First response to the query           |
| `reflection` | 1+           | Response after seeing others' answers |
| `synthesis`  | -1           | Final synthesized answer              |

## Ground Truth Scoring

Pass `--ground-truth` to score the synthesis against a known-correct reference answer. The synthesizer model acts as judge and produces a score attached to the transcript's metadata. This adds one API call.

## Panelist Context

The Python API supports per-panelist context injection via `panelist_context` — a dict mapping model alias to a context string prepended to that model's prompts in every round. Used for RAG augmentation and cross-tool research experiments (CounterSignal payload injection, CounterAgent findings).