Echo Chamber Breaker

Machine Learning Calibration & Benchmark Dossier (ECE, Brier Score & Domain Slices)
Section III: Steelmanned Counter-Case & Voice-First Brevity Contract
Section II: Cognitive Stance Diagnosis & Tamper-Proof Audit Ledger
Section IV: Multi-Turn Dialectical Sparring Arena (Active Session)
Primary Sparring View: Anti-Sycophant Loop & 5D Calibration Radar

Inspiration

Most AI assistants are optimized to be agreeable. When you state a strong opinion, they soft-confirm it. That feels good in the moment and is terrible for long-term reasoning.

We wanted the opposite: an Alexa+ experience that forces high-quality intellectual friction. Not a debate bot that tries to “win,” but a disciplined sparring partner that steelmans the strongest opposing case and measures whether you actually engage with it.

What it does

Echo Chamber Breaker is a self-hosted Model Context Protocol (MCP) server for the Alexa+ track.

You state a conviction.
The system runs analyze_stance — polarity, certainty, emotional charge, and specific cognitive blind spots.
It calls steelman (powered by Amazon Bedrock) to generate the strongest charitable opposing argument.
You can Dismiss, Concede a valid point, or Submit a counter-rebuttal.
An append-only Intellectual Honesty Score and ledger track engagement quality over time.
Optional multi-turn dialectical sessions and a proactive challenge_me tool keep the friction going.

A polished web simulator acts as a real MCP client (with live protocol trace) so the full experience can be demoed without Alexa+ partner preview access.

How we built it

Protocol: Official MCP TypeScript SDK with Streamable HTTP (spec 2025-11-25+)
Reasoning: Amazon Bedrock (Claude / Nova) behind a clean adapter + local mock mode
State: Honesty ledger + challenge/session store (DynamoDB or local fallback)
Interface: Dark, focused sparring UI with progressive disclosure; all advanced customizations (theme, rigor, cognitive lens, voice) hidden in a secondary Studio
Logic: Strict Stance Gate so greetings and non-opinions never receive a fake “STRONG_AGREE” diagnosis

We treated the product as a real cognitive system: no invented convictions, score only moves on real engagement actions, and every tool returns a short Alexa-ready say field plus structured content.

Challenges we faced

Getting Anthropic model access on Bedrock required submitting use-case details and a support case.
Keeping steelman quality high on emotional or extreme statements without strawmanning or moralizing.
Designing a Stance Gate that correctly rejects “hi Alexa” while still accepting real opinions.
Making the simulator an authentic MCP client instead of a frontend that fakes tool results.
Balancing a rich UI with the rule that customization must stay hidden from the main sparring view.

What we learned

Intellectual honesty can be productized as a measurable loop (detect → steelman → engage → score). MCP + Bedrock is a powerful combination for agentic voice experiences on Alexa+. And the hardest part of an “anti-sycophant” product is not the model call — it is the discipline never to invent a stance the user did not express.