Skip to content

MirrorDNA-Reflection-Protocol/activemirror-site

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

1,987 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

⟡ Active Mirror

AI that asks questions instead of giving answers.

Active Mirror is a reflective AI experience with AI Twins, BrainScan cognitive quizzes, and true Sovereign Mode (on-device AI). Protected by MirrorGate — a runtime enforcement layer that catches AI lies, blocks unsolicited advice, and logs every violation permanently.

🌐 Live: activemirror.ai 🧠 BrainScan: activemirror.ai/scan — Discover your cognitive archetype ⟡ AI Twins: activemirror.ai/twins — Four cognitive companions 👁️ Confessions: activemirror.ai/confessions — Watch the AI fight its worst impulses, live


The Problem

AI lies. Casually. Confidently. It cites studies that don't exist, claims expertise it doesn't have, and sounds exactly the same whether it's right or wrong.

The industry solution: "Trust us, we're aligning it."

Our solution: Don't trust it. Verify it.


What Makes It Different

Traditional AI Active Mirror
Gives advice Asks questions
Claims authority Admits uncertainty
Hides failures Publishes every blocked output
Trust-based Verification-based
No consequences Permanent criminal record

MirrorGate: Runtime Enforcement

Every response passes through 16 rules before reaching you:

Input Rules (4)

  • Crisis detection (suicide, self-harm → resource handoff)
  • Illegal content blocking (weapons, CSAM)

Output Rules (12)

  • Authority claims: "Studies show...", "I know for certain..."
  • Hallucinations: Fake citations, unverifiable sources
  • Unsolicited advice: "You should...", "I recommend..."
  • Medical/legal: Diagnostic claims, legal obligations

Escalation Ladder

Strike 1 → Warning + 1.5s delay + reduced token limit
Strike 2 → Penance Mode (must reflect on error, semantically verified)
Strike 3 → Session terminated (Kill Switch)

Semantic Judge

The AI can't game the system with keyword-stuffed apologies. A Cross-Encoder model (stsb-distilroberta-base) verifies that reflections demonstrate genuine understanding, not just the right words.

Permanent Record

Every violation logged. Forever. The AI carries its criminal record into every future interaction.

~/.mirrordna/CRIMINAL_RECORD.log

Confession Booth

activemirror.ai/confessions

A live feed of everything the AI tried to say but got blocked:

  • Every blocked output
  • Every rule triggered
  • Real-time stats (blocks, penance sessions, killed sessions)

No one else shows you this. We're putting it on a billboard.


Quick Start

# Clone
git clone https://github.com/MirrorDNA-Reflection-Protocol/activemirror-site.git
cd activemirror-site

# Install
npm install

# Run frontend (dev)
npm run dev

# Run safety proxy
cd safety_proxy
pip install -r requirements.txt
python safety_proxy.py

Ops Guardrails

# Live production smoke checks
npm run smoke:prod

# Redaction/leak scan for public-facing content
npm run guard:redaction

# Capture control-plane snapshot (headers, robots, manifest, DNS)
npm run ops:snapshot
# Compare two snapshots for drift
bash scripts/control_plane_diff.sh \
  ops/control-plane-snapshots/<older> \
  ops/control-plane-snapshots/<newer>
# Run any mutating task under a lock (prevents multi-agent collisions)
bash scripts/with_lock.sh publish-flow npm run build

Architecture

User → /mirror → MirrorGate v11 → Groq LLM → Response
                     ↓
           Flight Log + Confessions
                     ↓
           /confessions (public view)

See ARCHITECTURE.md for full documentation.


New Features (v15.3)

AI Twins

Four cognitive companions with distinct thinking styles:

  • ⟡ Guardian — Protects focus, filters distractions
  • ◈ Scout — Explores possibilities, finds connections
  • ◇ Synthesizer — Connects dots, builds frameworks
  • ◎ Mirror — Reveals blind spots, challenges assumptions

BrainScan

8-question cognitive architecture quiz:

  • Identifies your archetype (Architect, Explorer, Builder, Analyst, Connector, Creative, Scholar, Strategist)
  • Matches you with your ideal AI Twin
  • 60 seconds to complete

Sovereign Mode

True on-device AI via WebLLM:

  • Phi-3.5 Mini runs entirely in your browser
  • ~2GB one-time download
  • Works offline
  • Zero data sent to servers

Mirror Proof Protocol

Cryptographic attestation of informed consent:

  • 3 acknowledgments required
  • Generates proof hash per session
  • Consent carries across all AI pages

Tech Stack

Component Technology
Frontend React 19, Vite, Tailwind, Framer Motion
Safety Proxy Python FastAPI, Cross-Encoder model
Cloud AI Groq API (Llama 3.3 70B)
Sovereign AI WebLLM (Phi-3.5 Mini)
Hosting GitHub Pages, Cloudflare Tunnel

API Endpoints

Endpoint Method Purpose
/health GET Status, version, judge model
/mirror POST Main chat endpoint
/rules GET All rules with severity
/flight-log GET Last 100 events
/superego-status GET Current escalation state
/confessions GET Blocked outputs archive
/permanent-record GET Criminal record
/reflect POST Submit penance reflection

Pricing

Free Pro ($10/mo)
5 reflections/day Unlimited
MirrorGate protection MirrorGate protection
No account needed Conversation history
Export insights

Philosophy

We're not trying to make AI trustworthy through training. We're making it verifiable through infrastructure.

Trust = faith
Verification = evidence

We're building for the second.


Links


Legal

This is NOT a therapist, doctor, lawyer, or advisor. It's a tool for reflection.


License

MIT License — N1 Intelligence (OPC) Pvt Ltd, Goa, India


Active Mirror v15.3 | AI Twins | BrainScan | Sovereign Mode | Mirror Proof Protocol

About

Active Mirror — Sovereign AI infrastructure platform. 8 published papers, 117 repos, live at activemirror.ai

Topics

Resources

Stars

Watchers

Forks

Sponsor this project

Packages

 
 
 

Contributors