• Your prompts are processed entirely client-side in your browser
• No raw prompt text is ever transmitted to Destill.ai servers
• You can inspect the network tab — zero outbound data
• All processing follows EU GDPR Art. 5(1)(c) data minimization

If you opt in, we collect only aggregate statistics — never your prompt content:

• ✅ Classification result (TP/FP/TN/FN)
• ✅ Cumulative threat score (0.00–1.00)
• ✅ Which layers fired (agent IDs only)
• ✅ Processing latency (ms)
• ❌ Never raw prompt text
• ❌ Never API keys or session identifiers

This uses Differential Privacy (ε=1.0, δ=10⁻⁵) — even aggregate stats are noise-injected so individual prompts cannot be reconstructed. Patent Claims 1840-1862 (GTO Architecture).

POAW · CO₂ Impact — Methodology & Sources

Current industry practice: Every prompt to an LLM requires a GPU-powered safety classifier (e.g., OpenAI Moderation API, Llama Guard, Azure Content Safety). These classifiers run deep neural networks on GPU hardware, consuming 0.2–0.4 Wh per inference.

The NI-Stack replaces this with a 108-agent CPU-only cascade that processes prompts in 0.18ms using pattern matching, statistical analysis, and linguistic heuristics — consuming <0.0001 Wh per prompt. That's a 2,000× energy reduction.

Only ambiguous prompts (cumT 0.33–0.43, ~2% of traffic) are routed to a lightweight NPU model via DirectML. The remaining 98% never touch a GPU.

Every prompt scanned generates a Proof of Audited Work (POAW) receipt — a cryptographically signed (ML-DSA post-quantum) attestation of which agents processed the prompt, what thresholds were used, and what decision was made.

Receipts are hash-chain linked (each receipt references the previous), creating an immutable audit trail. This enables continuous monitoring for insurance attestation similar to Munich Re aiSure™.

Attestation status progresses from 📋 Collecting → 🔍 Under Review → 📊 Evaluating → ✅ Attestable as the chain grows.