Introduction

Ngeodesic provides safety guardrails for AI with enhanced performance. Our Micro-LM (micro language model) sidecars act as a Layer 2 for AI: improving efficiency and reliability, enforcing determinism, auditability, and compliance — all while letting large language models remain fluent and flexible. For GitHub repos, you can visit:

• NGF Research

• micro-LM

Warp->Detect->Denoise (WDD)

Warp → Detect → Denoise (WDD) is a patented pipeline for making noisy latent spaces deterministic and auditable. It first warps embeddings into a stable geometry, then detects true signals with calibrated filters, and finally denoises phantom attractors. The result: reproducible PASS/ABSTAIN decisions with clear traces — turning stochastic LLM (large language model) outputs into predictable, safe, enterprise-grade behavior.

Micro-LMs

Micro-LMs are lightweight, domain-specific companions to LLMs that provide determinism, safety, and auditability. They focus on a narrow set of primitives (e.g., DeFi actions, ARC puzzles), mapping prompts into PASS or ABSTAIN with verifiers and traces. The LLM handles fluency, while the micro-LM enforces guardrails and compliance — the best of both worlds.

Micro-LM Sidecar

A Micro-LM Sidecar can be thought of as a Layer 2 for AI: it doesn’t replace the large language model, but it sits alongside it to improve performance, add safety guardrails, determinism, and auditability. In AI, the LLM is like Layer 1: broad, powerful, but stochastic and sometimes unsafe. A Micro-LM Sidecar is like Layer 2: lightweight, domain-specific, enforcing determinism and safety guardrails, but still “anchored” to the LLM’s fluency and reasoning power.

Request Cycle

User → LLM → /decide {prompt, context, policy}
                 ↓
           SBERT → PCA → map → θ → guards
                 ↓
   ← {approve|reject|abstain, reason, plan, trace_id}

• LLM explains/asks confirm

• User confirms

• LLM → /execute {trace_id, plan}

• Sidecar re-checks guards → tools (if OK) → finalize trace

Workstream Breakdown

Note

Status icons legend — ✅ Built • 🟡 Partial • 🔴 Planning

Workstream	Component	Status	% Effort	% Done	Description
WDD R&D	WDD	✅	30%	30%	Research, math, validation of Warp → Detect → Denoise (ngf-alpha, DeFi/ARC, proofs).
Core pipeline (R&D)	Micro-LM	✅	15%	13%	SBERT → PCA prototyping, thresholds, datasets.
Core pipeline (Engineering)	Micro-LM	✅	20%	17%	Engineering impl. of SBERT → PCA/θ, integration with WDD-lite.
Domain guards	Micro-LM	✅	10%	8%	HF / LTV / oracle freshness verifiers.
Plans & policies	Micro-LM	✅	5%	4%	Primitive specs, per-class thresholds, policy defaults.
Quickstart harness	Micro-LM	✅	5%	4%	In-process run + JSON outputs.
Tests & canaries	Micro-LM	🟡	5%	3%	Smoke cases, threshold checks.
API layer	Sidecar	🔴	4%	0%	/decide, /execute, schemas.
Orchestrator glue	Sidecar	🔴	2%	0%	Routing, retries, caches.
Observability	Sidecar	🔴	2%	0%	Logs, metrics, /healthz & /version.
Audit trail	Sidecar	🟡	1%	0.3%	Structured traces, IDs.
Ops safety	Sidecar	🟡	1%	0.2%	Warmup, kill-switch, double-gate at execute.

Total % Done ≈ 85% (±5%).

Our first pilots are:

• ARC Micro-LM → reasoning stress test (Abstraction & Reasoning Corpus).

• DeFi Micro-LM → finance primitives with verifiers for safety and risk control.

If you are a researcher, developer, or just curious about deterministic AI reasoning, this guide will help you get started quickly.