Try ATMAN
ETHICS ARCHIVE // THE COGNITIVE MIRROR // RESPONSIBLE AI CHARTER

Responsible AI Charter

"Artificial intelligence must mirror human consciousness, never manufacture it."
POSTURE: COGNITIVE AUGMENTATION (NON-AUTONOMOUS)
CUSTODY: LOCAL-FIRST (INDEXEDDB AES-GCM-256)
TRANSIT: ZERO-RETENTION AIRLOCK PROXY
REDLINES: CLINICAL & LEGAL REFUSAL ENFORCED
Active Governance & Evolution Notice

ĀTMAN is currently in active development. As our cognitive intelligence architecture expands and new capabilities are introduced, our Responsible AI principles will continually evolve to ensure human dignity, cognitive sovereignty, and psychological safety. We hold ourselves to radical epistemic honesty: when the system is uncertain or a limitation exists, we state it plainly. Registered users will be notified of material governance updates.

INTERACTIVE ARCHITECTURE // COGNITIVE CHARTER

The Foundational Principles

Explore how ĀTMAN approaches human agency, conversational dynamics, non-extractive calm computing, and machine boundaries. Select any principle to inspect its operational realities, technical codebase enforcement, and known limitations.

PILLAR 01 // COGNITIVE REVERBERATION & HUMAN AGENCY

The Mirror Principle: Augmenting, Not Replacing Agency

ĀTMAN is grounded in a singular philosophical truth: AI should serve as an introspective mirror for the human mind, never an automated substitute for human consciousness. We reject the contemporary industry pattern of passive automation where algorithms tell users what to think, conclude, or believe.

Reflective Partnership

Ātmi acts as a contemplative companion that reflects your own thinking back to you with structural clarity. Rather than generating final conclusions, it presents structural outlines, unexamined assumptions, and alternative hypotheses so you reach your own authentic discoveries.

  • Non-Directive Prompts: System templates prioritize Socratic inquiry over declarative answers.
  • Zero Executive Delegation: The companion cannot execute life actions, sign contracts, or dispatch communications on your behalf.
Enforced in: public/js/workspace.js
Sovereign Authorship

The user remains the sole author, editor, and final arbiter of their thoughts, learning pathways, and decisions. ĀTMAN never prescribes moral dogmas, life paths, or philosophical verdicts.

  • Transparent Reasoning: Step-by-step cognitive scaffolding is exposed rather than hidden inside a black box.
  • User Retains All Copyright: 100% of generated drafts, mind maps, and journal entries belong exclusively to the user.
Policy: Terms of Service §4.1
User Sovereign Action

Approach conversations with Ātmi as an active dialogue with your own metacognition. Challenge the model’s formulations, disagree freely, and treat generated synthesis as exploratory draft material rather than authoritative truth.

  • Critique and Refine: Use the reflection bridge in the workspace to discard irrelevant suggestions.
  • Cultivate Self-Trust: Treat the AI as an externalized whiteboard, not a replacement for your personal intuition.
Known Machine Boundaries

Language models operate on statistical pattern completion. They possess no biological consciousness, subjective awareness, or moral conscience. They cannot "know" you in a psychological sense, nor can they experience the consequences of your decisions.

  • No Sentience: Companionship cues are conversational design metaphors, not emotional reality.
  • Risk of Over-Reliance: Excessive intellectual reliance on AI prompts can atrophy unprompted, spontaneous ideation.
"A tool that makes you dependent on its answers has failed you. A companion that clarifies your own thinking has served you."
ĀTMAN Core Axiom // Epistemic Sovereignty
RADICAL TRANSPARENCY // COMPARATIVE MATRIX

The Responsibility & Boundary Matrix

Responsible AI requires uncompromised clarity about boundaries. We explicitly delineate what ĀTMAN technically guarantees, what the system cannot promise, and what remains the sovereign responsibility of the human user.

ĀTMAN GUARANTEE

Local-First Cryptographic Custody

Your cognitive workspace drafts and memory vectors remain resident in your device’s IndexedDB sandbox, encrypted via Web Crypto AES-GCM-256 with PBKDF2 100k iteration key derivation.

Verified in: public/js/vault.js
ĀTMAN GUARANTEE

Zero-Retention Transit Gateway

Prompts routed to inference engines are streamed transiently through an in-memory Node.js airlock proxy that scrubs client identification headers and never stores transcripts to disk.

Verified in: server.js (/api/messages)
ĀTMAN GUARANTEE

Non-Extractive Behavioral Design

Zero advertising trackers, zero engagement streak punishments, zero infinite algorithmic feeds, and debounced 528Hz Solfeggio acoustic cues engineered for neurological calm.

Verified in: public/js/app.js & workspace.js
ĀTMAN GUARANTEE

Absolute Clinical & Legal Refusal

The companion categorically refuses to prescribe pharmaceutical dosages, issue psychiatric diagnoses, or provide certified legal counsel, actively routing emergency distress to verified hotlines.

Enforced via System Prompt Guardrails
KNOWN LIMITATION

Probabilistic Confabulation (Hallucination)

Foundation models do not understand truth in an ontological sense. They generate statistically plausible tokens that can result in inaccurate citations, false historical dates, or flawed reasoning.

Inherent to Current Deep Learning Architectures
KNOWN LIMITATION

Absence of Sentience & Genuine Empathy

Ātmi does not feel emotions, experience grief, or possess consciousness. Conversational warmth is an engineered linguistic interface designed for psychological clarity, not a sentient bond.

Philosophical Invariant: Machine != Consciousness
KNOWN LIMITATION

700-Token Flat Memory Budget Ceiling

To prevent data bloat and high token costs, our distiller caps companion working context at ~700 tokens across 4 discrete layers. Distant nuances outside recent memory are intentionally discarded.

Verified in: public/js/memory.js
KNOWN LIMITATION

Upstream Foundation Model Biases

Underlying LLM weights reflect broad internet training corpuses. Despite rigorous anti-sycophancy prompts and system grounding, subtle cultural, linguistic, and historical biases may surface.

Mitigated via System Prompting & Socratic Posture
USER SOVEREIGNTY

Empirical Fact Verification

You are responsible for independently corroborating scientific theorems, historical assertions, and empirical data against authoritative primary literature before relying on them.

Core User Responsibility
USER SOVEREIGNTY

Sovereign Decision Authorship

Never abdicate ethical choices, investment decisions, or personal relationship directions to an artificial companion. The responsibility and moral ownership of your life belong solely to you.

Core User Responsibility
USER SOVEREIGNTY

Local Passkey & Hardware Custody

Because ĀTMAN relies on client-side encryption without cloud backdoors, you must protect your physical device, maintain passkey hygiene, and export encrypted vault backups safely.

Zero-Knowledge Security Requirement
USER SOVEREIGNTY

Active Intellectual Resistance

Approach generative AI with healthy intellectual skepticism. Exercise your capacity for unassisted original thinking, creative struggle, and deep offline human contemplation.

Human Cognitive Preservation
CODEBASE EVIDENCE // TECHNICAL ENFORCEMENT

Under the Hood: How Principles Map to Code

Responsible AI policies are meaningless without verifiable code architecture. Inspect the exact technical mechanisms and file invariants that uphold our charter.

IMPLEMENTATION DETAILS

Implemented in public/js/vault.js via the standard Web Cryptography API. Keys are derived using PBKDF2 with 100,000 iterations of SHA-256 and unique 96-bit initialization vectors per record. Data is stored strictly in local IndexedDB tables with a 1-click zeroization method (wipeLocalStore()) that cryptographically purges memory without cloud residue.

VERIFIED CODE INVARIANT 
// public/js/vault.js
crypto.subtle.encrypt(
  { name: 'AES-GCM', iv: uniqueIv },
  derivedKey,
  encodedPayload
);
IMPLEMENTATION DETAILS

Engineered in public/js/memory.js via AtmanMemoryDistiller. It enforces a strict ~700-token ceiling partitioned into Layer 1 (Core Identity Anchor: ~80 tok), Layer 2 (Cognitive Vector: ~120 tok), Layer 3 (Episodic Salience: ~200 tok), and Layer 4 (Session Working Window: ~300 tok). This guarantees the model does not hoard unbounded historical chat transcripts.

VERIFIED CODE INVARIANT 
// public/js/memory.js
const BUDGET = {
  LAYER_1: 80,  // Identity
  LAYER_2: 120, // Dimensions
  LAYER_3: 200, // Salience
  LAYER_4: 300  // Working
}; // Total ~700 tokens flat
IMPLEMENTATION DETAILS

Handled by server.js at endpoint /api/messages. Requests are checked against strict rate limits, sanitized to strip origin telemetry or custom credentials, and forwarded via TLS to DeepSeek cloud inference backends. For authenticated sessions, turns and distilled memories are stored in the user's private database partition to maintain continuity, and can be permanently purged upon account deletion.

VERIFIED CODE INVARIANT 
// server.js (/api/messages)
// Ephemeral transit: stream chunks directly
// Zero database writes or transcript logging
res.writeHead(200, { 'Content-Type': 'text/plain' });
stream.on('data', chunk => res.write(chunk));
IMPLEMENTATION DETAILS

Programmed in public/js/app.js and public/js/workspace.js using the Web Audio API. Auditory cues synthesize 528Hz sine waves (the harmonic Solfeggio frequency associated with centered clarity) with an explicit 350ms cooldown debounce. This eliminates accidental double-triggers, audio clipping, and sensory hyper-stimulation.

VERIFIED CODE INVARIANT 
// public/js/app.js
if (Date.now() - lastChimeTime < 350) return;
lastChimeTime = Date.now();
osc.frequency.setValueAtTime(528, ctx.currentTime);
COORDINATED DISCLOSURE // ETHICAL OVERSIGHT

Algorithmic Accountability & Feedback Channel

If you observe unexpected model hallucinations, subtle demographic or linguistic biases, potential safety circumventions, or security vulnerabilities in our implementation, please contact our team immediately. We treat algorithmic anomalies with the same seriousness as critical security exploits.

Support@withatman.in GitHub Vulnerability & Bias Reporting Inspect Security Architecture