Skip to content

The Sovereign Directive

uncoalesced edited this page May 11, 2026 · 1 revision

Welcome to the Peridot Wiki. Peridot is not an "app"; it is a sovereign runtime designed to reclaim user autonomy from centralized AI surveillance infrastructure.

Core Tenets

  1. Absolute Privacy: Air-gapped by design. No cloud, no telemetry.

  2. User Sovereignty: You control the logic, the model, and the hardware.

  3. Altruistic Compute: Idle cycles contribute to medical research via the Aether-Route handoff.


Architecture: The Aether-Route Protocol

Engineered by uncoalesced

Peridot utilizes a unique hardware-software handoff protocol named Aether-Route. It treats the host machine as a tiered processing plant, coordinating between the Ryzen 7 250 AI (NPU/CPU) and the NVIDIA RTX 5050 (Blackwell CUDA).

Processing Phases

  1. Semantic Triage: Every query is first intercepted by the Ryzen 7 250 AI. It performs a similarity search against the local vector cache. Matches are served in <40ms, completely bypassing the GPU.

  2. Resource Arbitration: If generation is required, a hardware interrupt is sent to the Medical Research Module.

  3. VRAM Hot-Swap: The Folding@Home buffer is purged via raw WebSocket command. ARPM telemetry confirms this handoff completes in a median of 516.01ms.

  4. Hardware-Native Generation: The RTX 5050 claims the VRAM and generates tokens at a sustained ~43.5 t/s.

  5. State Resumption: Upon response delivery, context buffers are flushed (0.00MB memory growth) and medical research resumes.


The Aether-Route Performance Matrix (ARPM)

The ARPM is the internal telemetry suite used to benchmark system readiness.

System Readiness Rating: 9/10

Validated on Ryzen 7 250 AI / RTX 5050 / 16GB RAM.

Metric | Result | Grade -- | -- | -- Handoff Latency | 516.01ms | A (Elite) Cache Throughput | 12,262.35 t/s | A (Elite) Native Throughput | 43.49 t/s | B (Standard) Memory Stability | +0.00MB Growth | A (Immutable)

Hardware Tiers

  • Elite (9-10/10): RTX 50-series (Blackwell) or 4090. Handoff < 600ms.

  • High Performance (7-8/10): RTX 3060/4050. 8GB VRAM.

  • Standard (5-6/10): 6GB VRAM. Heavy quantization required.


Security: Defense-in-Depth

Peridot implements a multi-layered security perimeter to prevent "Prompt Injection" and "Path Traversal" attacks common in cloud-based LLMs.

1. Ephemeral Authentication

API keys are generated cryptographically in RAM at boot. No keys are ever written to disk. If the process is terminated, the access credentials cease to exist (CWE-312 mitigation).

2. The Constitution Layer

The constitution.json file acts as the kernel's logic-gate. It is the only place where permissions (file read/write, web fetch) are defined.

  • Safe-Mode Fallback: If the file is missing or corrupted, the system defaults to allow_all: false.

3. Forensic Audit Logs

Every action, permission request, and response is logged with a SHA-256 session hash. This ensures that even if a local process attempts to hijack the kernel, the audit trail remains immutable.


Medical Research Module

Peridot integrates with Folding@Home to ensure your hardware is never truly "idle."

  • Cancer Research: Simulating protein dynamics to find new drug binding sites.

  • Alzheimer's: Modeling protein misfolding.

  • Seamless Transition: The Aether-Route protocol ensures that contributing to research never slows down your personal AI experience.

Command Controls

  • research enable: Opt-in to medical research contribution.

  • research status: Real-time VRAM allocation and folding points telemetry.


Developer Reference & API

API Endpoints

  • POST /ask: Primary inference endpoint. Requires RAM-only token.

  • GET /health: System heartbeat and hardware thermal check.

  • POST /research/disable: Immediate VRAM purge command.

CLI Reference

  • launcher.py: The master process manager.

  • setup.py: The hardware-aware dependency injector.

  • run_all_benchmarks.py: The master ARPM suite runner.


Roadmap: The Sovereignty Path

  • v1.3 (Current): RAG implementation with FAISS and high-density local embeddings.

  • v1.4 (Active Development): Performance optimization for 16GB RAM environments and Blackwell-specific tensor routing.

  • v1.5: Native Linux Kernel support (Debian/Ubuntu).

  • v2.0: Decoupled React WebUI with bi-directional WebSocket streaming.

Engineered by uncoalesced