ENDE
Blog · Insights · Analysis

From the Engine Room

Straight talk on AI, language models and requirements engineering – written by an engineer from the automotive industry.

LOCAL LANGUAGE MODEL The model interprets. The software decides. A key vault driven by a language model that runs entirely locally — field-tested on an off-the-shelf workstation, with no access to external services. Qwen3 14B · Q4_K_M ≈ 10 GB VRAM RTX 5060 Ti · 16 GB 8,192 tokens no network access THE MODEL proposes Understands natural language Picks the matching tool Resolves imprecise requests But: it can invent details and confuse similar intentions. Proposal THE APPLICATION decides Confirmation before every write Fault-tolerant search & validation Limits, checks, verification Sensitive values stay local — and out of the conversation context. Local does not automatically mean secure. 487 of 2,951 lines of code are dedicated to verification — 51 checks run on every build.
Field report
August 27, 202611 min read

What a Local Language Model Can Really Do – and Where Its Limits Are

Qwen3 14B on an off-the-shelf graphics card: what a local model handles well, where it invents details or misreads intentions – and why the safeguards belong in the application, not in the model.

Local LLMQwen3 14BTool calling
Read more
THE GROUNDING LADDER Where the answers come from. Fully sourced & traceable. The OSOS/Omega Regulation Check verifies system requirements and specifications against the current UNECE regulations. Every analysis draws hierarchically on verified sources — from the binding standard down to well-founded orientation. 1 Binding Normative Text (Original Source) Specific limits and requirements come directly from the imported regulations — including precise paragraph and chapter references. BINDING 2 OSOS-curated Context Editorially maintained guidelines and explanations — always transparently labelled as structured guidance. CURATED 3 General Model Knowledge Useful for initial orientation, but varies depending on the strength of the underlying LLM/foundation model — and therefore never justifies technical or legal specifics. ORIENTATION ✕ Free Invention (Hallucination) Unsourced statements, freely associated limits or invented standards are intercepted by strict system guardrails — every normative statement must be traceable to an imported source. BLOCKED PRINCIPLE 01 No speculation. If a detail is missing from the data, the system states it plainly: “Not contained in the regulation” — instead of inventing a number. PRINCIPLE 02 Fully transparent. Every requirement is linked directly to the exact passage in the original document and instantly verifiable. ! WHY THIS MATTERS FOR REGULATIONS A plausible-sounding but invented paragraph is more dangerous than no answer at all — reliability is our top priority. OSOS/Omega Regulation Check — Grounding, not hallucination.
Product
July 23, 20268 min read

Source Links Matter: How the Regulation Check Makes AI Answers Traceable

Grounding instead of hallucination: the Regulation Check bases its answers hierarchically on binding regulatory text, curated context and general model knowledge – and says so plainly when a detail simply is not covered by the imported regulations.

GroundingRegulation CheckUNECE
Read more
Benchmarking five LLMs on finding conflicts with UNECE clauses Comparing LLM performance within OSOS/Omega on UNECE-R48 compliance validation of a headlamp requirements spec detected (True Positive) False Positive missed — critical error 0 5 10 per model: detected (green) + missed (red) = 13 conflicts · orange = False Positive GPT-5.5 13 Precision 100 % Coverage 100 % ~$1.01 best overall value Gemini 2.5 Flash 12 1 Precision 100 % Coverage 92 % ~$0.33 Claude Opus 4.8 11 2 Precision 100 % Coverage 85 % ~$1.53 most expensive check Nemotron 3 Ultra 8 5 Precision 100 % Coverage 62 % ~$0.10 cheapest (openrouter.ai) Mistral Large 3 5 8 1 Precision 83 % Coverage 38 % ~$0.35 Task shape beats price tag False positives barely mattered (only Mistral: 1) — the models differ in their coverage. Only GPT-5.5 finds all 13; Opus (most expensive) doesn't win, Nemotron (cheapest) covers only 8. Mistral, with only 5 of 13, is the clear loser. UN-R48 Series 09 · 1,161 clauses, 13 hard conflicts · ~150 pre-filtered per requirement · ref.: human-reviewed gold standard.
Analysis
July 3, 20266 min read

The Dream of AI Autonomy

With Nemotron 3 Ultra, an open model finally comes noticeably close to the frontier class – and runs entirely in your own data center. What that means for data sovereignty, cost and critical reasoning tasks, and where the gap to the cloud leaders remains.

Self-hostingData sovereigntyNemotron 3 Ultra
Read more
Blog – OSOS/Omega | OSOS/Omega