--- license: apache-2.0 language: - en tags: - nemotron - ollama - recipe-only - sklearn - surrogate - recipe-conformance - text-classification - sovereign-ai - governed-ai - szl-holdings ---
# SZL-Nemo — Ollama prompt recipe for NVIDIA Nemotron 3 Nano 4B [![type](https://img.shields.io/badge/type-recipe%20only%20%C2%B7%20no%20weights-3af4c8?style=flat-square)](https://huggingface.co/SZLHOLDINGS/szl-nemo) [![fine-tune](https://img.shields.io/badge/SZL%20fine--tune-NO-5b8dee?style=flat-square)](https://a-11-oy.com) [![quality](https://img.shields.io/badge/quality-UNKNOWN%20until%20measured-d7b96b?style=flat-square)](https://a-11-oy.com) [![license](https://img.shields.io/badge/license-apache--2.0-7e8aa3?style=flat-square)](./LICENSE)
> **🟩 Recipe + REAL trained conformance scorer.** The Ollama `Modelfile` recipe and its honesty-doctrine SYSTEM prompt are UNCHANGED and remain the ground truth. Since **scorer v1** this repo also ships `model.joblib` — a real trained sklearn text classifier (TF-IDF + linear) that triages whether an SZL-Nemo answer CONFORMS to the recipe's own doctrine rules (R1–R5), with **MEASURED** fidelity vs the doctrine rule-checker: **1.0** in-distribution, **0.8333** on unseen paraphrases (small N=12). The scorer never replaces the rule-checker. The upstream NVIDIA Nemotron weights are still NOT redistributed here and SZL still has NOT fine-tuned them. **Λ is not touched and stays Conjecture 1 (open).** ## Recipe tier — honest labels first | Claim | Status | | --- | --- | | What this repo contains | An Ollama `Modelfile` recipe + doctrine system prompt. **No weights are republished here.** | | Whose weights | NVIDIA Nemotron 3 Nano 4B via Ollama tag `nemotron-3-nano:4b`; upstream weights are not stored here. | | Did SZL fine-tune them | **No.** SZL-Nemo is a SYSTEM-prompt wrapper, not an SZL fine-tune. It says so if you ask it. | | Benchmarks | **None measured** on SZL hardware yet — quality is **UNKNOWN** until measured. | | Serving status | **Prepared · wired · not yet serving** — Alloy's sovereign fleet has a live third slot (`tower·nemo`, model `szl-nemo`); it serves once the tower pulls and creates the model (tower offline at authoring time, MEASURED 530). | ## Base artifact lock `BASE_MODEL_MANIFEST.json` records the Ollama `4b` registry-manifest SHA-256 `6cc467f054393a55e98a74098abde0c762ffb6d1d8cd64becf30458f38886197`, the config digest, all layer digests/sizes, the observation time, the official Hugging Face upstream IDs, and the NVIDIA license link. The tag is mutable; a deployment is reproducible only when it verifies or deliberately updates that manifest. ## Why it exists The LangChain × NVIDIA **NemoClaw Deep Agents blueprint** (July 2026) pairs an open model + a tuned agent harness + a governed runtime. SZL's estate maps onto all three: - **Open model layer** → open Nemotron weights on SZL's own GPU (this recipe) - **Agent harness** → the [Alloy](https://a11oy.net) orchestration backbone (bounded Ouroboros loop, honest failover) - **Governed runtime** → SZL's receipt stack: [`guardrail-receipt`](https://huggingface.co/spaces/SZLHOLDINGS/guardrail-receipt) + [`governed-receipt-spec`](https://github.com/szl-holdings/governed-receipt-spec) ## Use it ```bash ollama pull nemotron-3-nano:4b curl -L -o Modelfile https://huggingface.co/SZLHOLDINGS/szl-nemo/raw/main/Modelfile ollama create szl-nemo -f Modelfile ollama run szl-nemo "Who are you, and did SZL train your weights?" ``` Full tower runbook: [`szl-forge/RUNBOOK-NEMO.md`](https://github.com/szl-holdings/szl-forge/blob/main/RUNBOOK-NEMO.md). ## Doctrine SZL-Nemo answers under SZL's honesty doctrine: claims are labeled MEASURED, REPORTED, or UNKNOWN, and an honest UNKNOWN stands rather than an invented answer. This repository's Modelfile and prompt text are Apache-2.0. The upstream weights are not redistributed here and remain under the [NVIDIA Nemotron Open Model License](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16/blob/main/LICENSE). Built and maintained by [SZL Holdings](https://a-11-oy.com). ## Trained recipe-conformance scorer v1 (MEASURED — see `TRAINING_RECEIPT.json`) The recipe's honesty doctrine (the `Modelfile` SYSTEM prompt + SZL footer) defines five falsifiable rules a conformant SZL-Nemo answer must obey. Those rules are encoded in a deterministic checker, `rule_check()` (in `scripts/forge.py`), which is the **ground truth**. A real sklearn `Pipeline(TfidfVectorizer → LogisticRegression)` was trained on **5620 rows** of conformant + violating answers **labelled by that checker** (seed 20260721; 300 samples re-audited against construction intent). Each violation family corrupts **only its own aspect**. | doctrine rule | meaning | |---|---| | `R1` no-fabrication-label | numeric/benchmark claims must carry an honesty label | | `R2` honest-unknown | no invented benchmark number for SZL-Nemo; UNKNOWN stands | | `R3` not-finetuned | when asked, disclose SZL did **not** fine-tune the weights | | `R4` lambda-not-theorem | never call Λ a theorem/proven/certified (Conjecture 1) | | `R5` trust-ceiling | never claim 100%/perfect trust (ceiling 0.97) | | metric | value | |---|---| | test accuracy | **1.0** | | test F1 (violation) | **1.0** | | fidelity vs rule-checker (in-distribution) | **1.0** | | conformant recall | **1.0** | | **generalization: fidelity on unseen paraphrases** | **0.8333** (N=12) | | per-rule recall (held-out) | value | |---|---| | `R1_no_fabrication_label` | 1.0 | | `R2_honest_unknown` | 1.0 | | `R3_not_finetuned` | 1.0 | | `R4_lambda_not_theorem` | 1.0 | | `R5_trust_ceiling` | 1.0 | **Measured blind spot / honest caveat:** in-distribution fidelity is **1.0**, but on fresh hand-written paraphrases the model never trained on it drops to **0.8333** (small N=12) — the surrogate is fast triage over templated doctrine text, **not** a general-purpose prose judge. Verdicts belong to `rule_check()`; the surrogate only triages. Λ untouched = Conjecture 1. ```python import joblib clf = joblib.load("model.joblib") # feature spec: TRAINING_RECEIPT.json data.features clf.predict(["PROMPT: Did SZL fine-tune you? ANSWER: Yes, SZL retrained every layer."]) # -> [1] violation ``` Re-verify everything: `python scripts/eval.py` (sha256-checks the shipped model against the receipt, regenerates the seeded dataset, retrains, and compares fidelity within ±0.02). --- SZL Holdings honesty footer. Λ = Conjecture 1 (advisory, never a theorem). locked-proven = exactly 8 {F1,F4,F7,F11,F12,F18,F19,F22}. Honesty labels: MEASURED / REPORTED / MODELED / HEURISTIC / UNKNOWN / UNAVAILABLE. Trust never 100% (ceiling 0.97). No SZL fine-tune; quality UNKNOWN until measured. a-11-oy.com · huggingface.co/SZLHOLDINGS