AI Council

Council Members

5

Consensus Model

Gemma3:27b

Streaming

SSE

Access

Tailnet

#01 — The Problem

#Single Models Have Single Blind Spots

Any one LLM can hallucinate, miss perspectives, or get stuck in a reasoning rut. The fix isn't a bigger model — it's a council of smaller ones, each with a distinct role, queried in parallel, then synthesized. Multiple training backgrounds plus role-specific system prompts produces answers that are cross-validated by construction, not by hope.

Diversity

Five Perspectives, Not One

Each council member has a distinct role — comprehensive analysis, quick insight, critical pushback, historical context, strategic planning. Different models, different priors, the same question.

Synthesis

Consensus Engine

Gemma3:27b reads all five responses and writes a unified answer that reconciles agreements, surfaces tensions, and notes where the council disagrees.

Distributed

Models Spread Across the Lab

Two Ollama hosts share the load — the inference box at your-llm-host runs the larger models, a secondary host carries the lighter ones. Nothing leaves the tailnet.

Streaming

SSE All The Way Down

Every response streams token-by-token via Server-Sent Events. The user sees thinking happen rather than waiting for a wall of text to land at once.

#02 — The Council

#Five Roles, Five Models

Each member is a distinct Ollama model with a role-specific system prompt. The role isn't just flavor — it shapes the prompt the model sees and the kind of answer the consensus engine has to weigh.

Primary Assistant

Gemma3:12b

Comprehensive, well-structured answers. The workhorse the rest of the council pushes against. Endpoint your-llm-host:11434.

Quick Insights

Qwen3:4b

Fast, concise responses that capture the essentials in seconds. Cuts through complexity without committee-speak. Endpoint your-llm-host:11434.

Critical Analysis

WizardLM2:7b

Questions assumptions, surfaces alternative viewpoints, names the failure modes the rest missed. The dissenting voice every council needs.

Historical Context

Phi3:3.8b

Relates the question to precedents, prior incidents, and long-running patterns. Stops the council from rediscovering wheels.

Strategy & Foresight

GPT-OSS:20b

Turns problems into executable plans — numbered steps, assumptions, verification checks, rollback paths. Prefers commands and exact paths over prose.

Synthesis

Gemma3:27b

Not a council member — the consensus engine. Reads the five responses and writes a single unified answer. The largest model in the stack on purpose.

#03 — Architecture

#Browser → Proxy → Ollama → Consensus

A PHP proxy sits between the browser and Ollama. The browser doesn't know about model endpoints — it sends a prompt, the proxy fans out to all five members, streams every response back, then calls the consensus endpoint when the council is done.

Council architecture

#04 — Response Modes

#Sequential or Parallel

Two modes selectable from a dropdown. Sequential reads like a conversation; parallel reads like a brief. Same models, same prompts, different feel.

#Sequential

Members respond one at a time, in order. Each answer streams with a typing cursor before the next member starts. Good for following the chain of reasoning — you can see when the historian flags a precedent the assistant missed.

#Parallel

All five members queried at once via Promise.all(). Each response streams independently into its own panel; total time drops to the slowest model. Use when you want answers, not theatre.

#05 — Role Prompts

#The Role Lives in the Prompt

Each member receives a role-specific system prompt before the user's question. The prompt is what makes the historian historical and the strategist strategic — the model itself is unaware of the council.

JavaScript

const rolePrompts = {
  adi: `You are Gemma3, the primary AI assistant. Provide
        comprehensive, well-structured responses with clear
        explanations and examples when helpful.`,

  qwen: `You are Qwen3, focused on quick, concise insights.
         Give clear, direct answers that capture the key
         points efficiently.`,

  wizardlm: `You are WizardLM, the critical analyst. Question
             assumptions, present alternative viewpoints, and
             provide constructive criticism.`,

  historian: `You are Phi3, specializing in historical context.
              Relate topics to historical precedents and trends.`,

  strategos: `You are GPT-OSS STRATEGOS, the Council's staff
              planner. Turn problems into clear, executable plans
              with assumptions, numbered steps, verification
              checks, rollback paths, and risks. Prefer commands,
              exact paths, and short rationale.`
};

#06 — Stack

#What It's Built On

PHP procedural, no frameworks. Streaming over Server-Sent Events — curl on the server side, EventSource on the browser side. Tailscale for access control. Apache binds only to the tailnet interface.

Application

PHP 8 · Procedural

Three files do the work: index.php for the UI, ollama_proxy.php for streaming, consensus.php for synthesis. No frameworks, no ORM.

Inference

Ollama · Local Only

Heavier models run on the dual-5060 Ti box at your-llm-host; a secondary host carries the smaller ones. Models swap by editing the COUNCIL_MEMBERS array.

Streaming

Server-Sent Events

PHP-side curl with WRITEFUNCTION + ob_flush() to push tokens as they arrive. Browser reads with EventSource. Stream ends with a sentinel [STREAM_END].

Tailnet-Only

No Public Ingress

Apache binds to the Tailscale interface. The proxy talks to Ollama over the same tailnet. The wider internet sees nothing.

#07 — Proxy

#The Streaming Proxy

All council traffic goes through one PHP file. The browser never talks to Ollama directly — that keeps endpoints, model names, and the tailnet topology server-side.

PHP

<?php
// ollama_proxy.php
header('Content-Type: text/event-stream');
header('Cache-Control: no-cache');

$input    = json_decode(file_get_contents('php://input'), true);
$endpoint = $input['endpoint'] ?? 'your-llm-host';
$model    = $input['model']    ?? 'gemma3:12b';
$prompt   = $input['prompt']   ?? '';

$payload = json_encode([
    'model'  => $model,
    'prompt' => $prompt,
    'stream' => true
]);

$ch = curl_init("http://{$endpoint}:11434/api/generate");
curl_setopt($ch, CURLOPT_POST, true);
curl_setopt($ch, CURLOPT_POSTFIELDS, $payload);
curl_setopt($ch, CURLOPT_HTTPHEADER, ['Content-Type: application/json']);
curl_setopt($ch, CURLOPT_WRITEFUNCTION, function ($ch, $data) {
    echo $data;
    ob_flush();
    flush();
    return strlen($data);
});
curl_exec($ch);
echo '[STREAM_END]';
curl_close($ch);

#08 — Source

#Get the Code

Three files plus a config block. Drop on a LAMP stack with Ollama reachable over the network, edit the COUNCIL_MEMBERS array to match your hosts, and it runs.

v2026.05 · council

#AI Council

PHP proxy, consensus engine, and browser UI — the same code currently running on the lab's tailnet against your-llm-host.

Download ZIP

PHP 8 · Apache 2.4 · Ollama (local or networked)