Core Product Feature

Adversarial Chat Deck

Deploy models in opposing pairings, run prompt benchmarking debates, and visually compare code implementations with built-in real-time cost trackers.

View Source Code

Arche Chat Deck (dual-pane console)

CONNECTED (OLLAMA + OPENAI)

Model A Codex (gpt-4o)

Tokens: 8,200 | Cost: $0.041

Response Payload:

I will refactor the database query routine to utilize a thread-safe connection pool from psycopg2.

import psycopg2.pool

# Thread-safe pool allocation
pool = psycopg2.pool.SimpleConnectionPool(
    minconn=1, maxconn=10,
    dsn="dbname=sessions user=postgres"
)

def query_session(session_id):
    conn = pool.getconn()
    try:
        with conn.cursor() as cur:
            cur.execute("SELECT * FROM logs...")
            return cur.fetchone()
    finally:
        pool.putconn(conn)

Execution validated. 0 linter warnings.

Model B Hermes (llama-3-70b)

Tokens: 9,100 | Cost: $0.018

Response Payload:

For high concurrency applications, I suggest writing an asynchronous connection pool using aiopg.

import aiopg

# Async pool initializer
async def init_pool():
    async with aiopg.create_pool(
        dsn="dbname=sessions user=postgres"
    ) as pool:
        return pool

async def query_session(pool, session_id):
    async with pool.acquire() as conn:
        async with conn.cursor() as cur:
            await cur.execute("SELECT * FROM...")
            return await cur.fetchone()

Async loop execution checked. 1 warning.

Adversarial Debate Stream: Model A argues that asynchronous I/O introduces unnecessary loop complexity for static dashboard queries. Model B counters that synchronicity causes worker starvation under concurrent file uploads. Arbitrator recommendation: Proceed with Model A sync pooling for the SQLite local backend, and transition to Model B async if migrating to PostgreSQL cloud clustering.

Dual-Model Pairing

Run identical tasks side-by-side on OpenAI, Anthropic, or local Ollama instances to evaluate quality and performance variations immediately.

Cost Guard Audit

Track input and output token consumption in real time. Configure automated cost thresholds to freeze sessions before runaway bills accumulate.

Local-First SQLite

All dialogue contexts, prompt metrics, and loop alerts are compiled in a local database. Zero cloud dependencies, zero external telemetry.

Model Arbitrator

Let a third model audit debates, score implementations, and automatically suggest consensus pipelines to insert into your codebase.