peridot-local 1.2.0b1

A sovereign local kernel with dynamic hardware scaling.

██████╗ ███████╗██████╗ ██╗██████╗  ██████╗ ████████╗
██╔══██╗██╔════╝██╔══██╗██║██╔══██╗██╔═══██╗╚══██╔══╝
██████╔╝█████╗  ██████╔╝██║██║  ██║██║   ██║   ██║   
██╔═══╝ ██╔══╝  ██╔══██╗██║██║  ██║██║   ██║   ██║   
██║     ███████╗██║  ██║██║██████╔╝╚██████╔╝   ██║   
╚═╝     ╚══════╝╚═╝  ╚═╝╚═╝╚═════╝  ╚═════╝    ╚═╝   

SOVEREIGN AI KERNEL — v1.2 BETA

⚠️ BETA
Medical research module under validation. Report issues.

Engineered by uncoalesced

GPU-accelerated, air-gapped AI runtime. Zero telemetry. Zero cloud dependency.

---

> OVERVIEW

Local LLM runtime with permission-based function calling. Runs on your hardware. Logs all actions.

┌─────────────────────────────────────────────────────────┐
│  USER QUERY                                             │
│     │                                                   │
│     ▼                                                   │
│  PERMISSION LAYER  ──── constitution.json               │
│     │                        │                          │
│     │              (block / allow / modify)             │
│     │                                                   │
│     ▼                                                   │
│  INFERENCE ENGINE  ──── localhost:5000 (air-gapped)     │
│  Llama-3-8B-Instruct                                    │
│     │                                                   │
│     ▼                                                   │
│  AUDIT LOG  ──── immutable, SHA-256 verified            │
└─────────────────────────────────────────────────────────┘

---

> PERFORMANCE

Hardware: NVIDIA GeForce RTX 5050 Laptop GPU (8GB VRAM)
Model: Llama-3-8B-Instruct · Q4_K_M Quantization
Context: 2048 tokens (n_ctx_per_seq)

Task	Output Tokens	Speed
Short Response	50	58.76 t/s
Medium Response	150	59.77 t/s
Long Response	512	57.71 t/s

Sovereign VRAM Handoff Benchmarks

Testing the latency of dynamic resource reallocation between background research processes and active LLM inference via WebSocket signaling.

• WebSocket Handoff Latency: 21.10 ms
• Sustained Inference Post-Handoff: 57.04 t/s

Technical note: The 21.10ms WebSocket handoff latency and 13.08ms VRAM purge demonstrate efficient state transitions on the Blackwell (sm_120) architecture. CPU threads are strictly isolated (0 threads allocated to background tasks) to ensure zero drop in generation speed during active inference.

---

> ARCHITECTURE

[1] — Inference Engine

Core LLM runtime. llama-cpp-python with cuBLAS acceleration.

Model:     Llama-3-8B-Instruct (GGUF · Q4_K_M)
Backend:   llama-cpp-python + cuBLAS
Endpoint:  localhost:5000 (no external routing)
Context:   Sliding Window (VRAM-aware)
Precision: 4-bit quantization

[2] — Sensory Subsystems

Local audio. No cloud APIs.

Auditory System — OpenAI Whisper

Voice-to-text transcription.
Hands-free command input.
No external audio transmission.

[3] — Permission Layer

Function-call interceptor. Blocks execution before action runs.

# constitution.json
{
  "allow_file_read":    true,
  "allow_web_fetch":    true,
  "allow_code_execute": false,
  "blocked_domains":    ["example.com"],
  "approved_domains":   ["arxiv.org", "pubmed.ncbi.nlm.nih.gov"]
}

Delete constitution.json to remove restrictions.

[4] — Audit Log

Append-only log. All queries, actions, permission decisions.

[2026-02-08 14:32:01] QUERY     "analyze my bloodwork results"
[2026-02-08 14:32:01] PERMISSION read(bloodwork.pdf) → ALLOWED
[2026-02-08 14:32:01] ACTION    file_read(bloodwork.pdf) → OK
[2026-02-08 14:32:03] RESPONSE  delivered (312 tokens, 5.52s)
[2026-02-08 14:32:03] HASH      sha256: a3f9c2...

SHA-256 hashing at shutdown for integrity verification.

[5] — Sovereign VRAM State Machine (Medical Research)

Dynamic hardware orchestration via local WebSockets (Port 7396). The kernel hot-swaps GPU VRAM between active inference and background medical research (Folding@home) without OS-level process fighting or manual intervention.

State: IDLE (15s threshold) └─ Socket Signal: {"cmd": "state", "state": "fold"} └─ GPU VRAM fully allocated to Folding@home. └─ CPU Threads: Isolated (0 threads to FAH to prevent bottlenecking).

State: ACTIVE (Prompt Detected) └─ Socket Signal: {"cmd": "state", "state": "pause"} └─ VRAM Yield: 21.10ms (WebSocket) + 13.08ms (purge) = 34.18ms total. └─ LLM takes absolute priority with zero inference latency penalty.

Disabled by default. Opt-in via research enable. Targets: Alzheimer's, Cancer, Parkinson's.

[6] — Interface

Custom tkinter UI. High-DPI. Terminal aesthetic.

• Hardware telemetry (CPU/RAM/VRAM)
• Drag-and-drop image input
• Conversation history
• Research status

---

> HARDWARE SUPPORT

Tier	Hardware	Mode	Speed
Full Support	NVIDIA RTX 3060+ (6GB+)	Standard	40–70 t/s
Full Support	NVIDIA RTX 4050+ (8GB+)	Standard	50–80 t/s
Full Support	NVIDIA RTX 5050 (8GB)	Standard	58 t/s (tested)
CPU Fallback	Any modern x64	CPU-Only	8–12 t/s
Lite Mode	AMD Radeon 680M / 780M	Phi-3	8–15 t/s
Lite Mode	Intel Iris Xe	Phi-3	5–10 t/s
Community	AMD RX 6000/7000	ROCm (Linux)	35–50 t/s
Community	Intel Arc A750/A770	Vulkan	25–40 t/s

Lite mode: Phi-3 Mini, 2048-token context.
Community builds: See COMMUNITY_INSTALL.md.

---

> INSTALLATION

Prerequisites

OS:      Windows 10/11 (64-bit)
GPU:     NVIDIA RTX Series, 6GB+ VRAM
Python:  3.11
Storage: ~10GB free (SSD recommended)

Setup

1. Clone

git clone https://github.com/uncoalesced/Peridot.git
cd Peridot

2. Environment

python -m venv venv
.\venv\Scripts\activate

3. Installer

python setup.py

PERIDOT SETUP WIZARD
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[✓] NVIDIA GPU detected: RTX 5050 (8.0GB VRAM)
[✓] CUDA 12.1 compatible
[✓] Recommended model: Llama-3-8B-Instruct (Q4_K_M)
[✓] Installing CUDA-enabled llama-cpp-python...
[✓] Downloading model (4.7GB)...
[✓] Writing config...

Setup complete. Run: python launcher.py
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

> USAGE

Launch

python launcher.py

Wait for:

[OK] Inference engine online — localhost:5000
[OK] Visual Cortex mounted
[OK] Auditory System mounted
[OK] Peridot ready.

Commands

Directive	Action
system status	Diagnostic across all subsystems
read [file.py]	Ingest local file into context
research enable	Activate medical research module
research stats	Display Folding@home stats
analyze this image	Invoke Visual Cortex

Configuration

constitution.json controls behavior and permissions.

{
  "system_prompt": "You are Peridot...",
  "allow_file_read": true,
  "allow_web_fetch": true,
  "allow_code_execute": false,
  "approved_domains": ["arxiv.org", "pubmed.ncbi.nlm.nih.gov"],
  "blocked_content": []
}

Delete file to reset. Auto-regenerates on launch.

---

> ROADMAP

[████████████████████] v1.0  Core Inference Engine (NVIDIA/Windows)
[████████████████████] v1.1  Performance Optimization (BETA)
[████████████████████] v1.2  Stability Fixes + VRAM Handoff & Medical Research testing
[█████░░░░░░░░░░░░░░░] v1.3  RAG Engine Implemenation
[░░░░░░░░░░░░░░░░░░░░] v1.5  Linux Support (Ubuntu/Debian)
[░░░░░░░░░░░░░░░░░░░░] v1.6  AMD Radeon (ROCm)
[░░░░░░░░░░░░░░░░░░░░] v1.7  macOS Support (Apple Silicon)
[░░░░░░░░░░░░░░░░░░░░] v2.0  WebUI (FastAPI + React)

---

> PHILOSOPHY

See PHILOSOPHY.md for design rationale and project principles.

---

> LICENSE & DISCLAIMER

License: MIT — free for personal and commercial use.

Disclaimer: Experimental software. User assumes responsibility for all commands, content, and hardware usage. Provided as-is, no warranty.

---

PERIDOT · SOVEREIGN AI KERNEL · v1.2 BETA

Engineered by uncoalesced

Your hardware. Your model. Your rules.