SZL Forge

Control before action. Evidence after.

Part of the szl-holdings estate · Product: a-11-oy.com · Proof: a11oy.net

Train, qualify, and publish bounded SZL model artifacts from explicit source and evidence.

SZL Forge is the canonical training-recipe, qualification, and publication control repository for several SZL model artifacts. This GitHub repository is software, curriculum, schemas, and evidence; it is not itself a weight repo.

Portfolio truth card

Investor value. Forge makes model maturity, public bytes, source lineage, and promotion boundaries navigable without presenting recipes, derivatives, or Spaces as equivalent to trained weights.

Developer/evaluator quickstart. Inspect the fail-closed portfolio binding before running or publishing anything:

python tools/publish_model_source_bindings.py \
  --source-revision "$(git rev-parse HEAD)"

This local plan does not publish. Compare its model ID, source files, expected weight hashes, receipt scope, promotion state, and limitations with the exact Hub revision under evaluation.

What SZL Forge is

• A QLoRA fine-tuning kit built on Unsloth.
• A set of model-specific curricula, schemas, qualification tools, signed evidence, and protected publication contracts.
• A way to produce locally controlled model derivatives on supported hardware; hardware fit, run completion, and artifact quality must be observed per run.
• A release source that keeps trained weights, quantized derivatives, recipes, snapshots, and Spaces in separate evidence lanes.
• The source home for Owned Agent Control v2, a separately packaged, fail-closed Windows process-tree supervisor with signed one-shot isolation requests and schema-bound context evidence.

Owned Agent Control v2

agent-forge/ turns the supplied Owned Agent Control design into an installable Python package with an operator runbook, explicit threat model, portable contract tests, a real Windows Server behavioral gate, and a read-only A11oy evidence projection. It does not grant A11oy process authority, modify remote providers, or claim to sandbox hostile code running as the same Windows user.

Start with agent-forge/README.md. Operational status must remain NOT_READY until the exact Windows runtime passes doctor and self-test; a Linux or packaging pass establishes portable contracts only.

OAC Clinical Gateway v2.5

clinical-gateway/ is the separately bounded integration surface for synthetic and already-deidentified cobas liat-shaped result traffic. It implements a local control engine, an analyzer-initiated TLS MLLP host listener, deterministic ACK^R33 responses, bounded file and polling adapters, a loopback API, and an operator UI. It produces signed, offline FHIR R4 candidate bundles; it does not publish those bundles into a clinical system.

The separate OAC System Health v1 model and synthetic operational dataset provide a reproducible, standard-library baseline for operator-attention advisories. Its 1,200 generated examples are not gateway event records or clinical data. The inference kernel verifies artifact receipts and cannot acknowledge a message, command a device, interpret a result, or authorize release.

The gateway intentionally provides no analyzer command, order download, calibration, test-start, result reinterpretation, automatic release, or clinical alert path. Its operational dataset and control-evidence score contain only local run observations and must never be treated as clinical evidence or a training corpus. Start with clinical-gateway/README.md and keep every deployment claim at NOT_SITE_VALIDATED until the exact device, site certificate, assay mappings, network, privacy controls, laboratory workflow, and destination reconciliation have been witnessed and approved.

One command (laptop)

The whole pipeline — folder, kit files, Unsloth, the CUDA-build torch the RTX 5050 needs, training, and the Ollama import — in a single PowerShell command (run from ANY folder; it puts itself in %USERPROFILE%\szl-forge):

iwr https://raw.githubusercontent.com/szl-holdings/szl-forge/main/forge.ps1 -OutFile "$env:TEMP\forge.ps1"; powershell -ExecutionPolicy Bypass -File "$env:TEMP\forge.ps1"

It prints each step honestly and stops on the real error if one appears. Prefer step-by-step? RUNBOOK.md is the same pipeline as one command per step.

If first words are garbage (@@@@…)

MEASURED 2026-07-12: Ollama's direct safetensors import of the Unsloth 16-bit merge produced corrupted weights — szl1 answered @ spam at temperature 0, even in raw mode, and re-quantizing the imported model did not fix it. Most likely the import path is at fault; the fix re-imports the already-trained merge at .\szl-model properly (no retraining):

iwr https://raw.githubusercontent.com/szl-holdings/szl-forge/main/rebirth.ps1 -OutFile "$env:TEMP\rebirth.ps1"; powershell -ExecutionPolicy Bypass -File "$env:TEMP\rebirth.ps1"

If rebirth STILL produces @ spam, the merge itself is suspect — report back for diagnosis rather than retraining (training is seeded, so an identical re-run would likely reproduce the same merge; isolating merge vs converter comes first). Future full forge.ps1 runs birth via this GGUF path automatically (step 6).

What SZL Forge is NOT

• It is not from-scratch pretraining. Training a frontier model from raw tokens genuinely requires datacenter-scale metal (many high-end GPUs, weeks of compute). Nobody should claim otherwise, and this kit does not.
• It is fine-tuning an already-open base model into SZL's own — a real, honest, achievable thing on a single laptop, and no more than that.

Kit contents

Exact-source model publication

publishing/model-source-bindings.json is the fail-closed release contract for the qualified ReceiptAgent and Khipu Hub models. The publisher verifies the expected public weight hashes, required signed receipts, and every canonical source file before it writes publication.json to either model repository. It then reads the exact Hub commit back and compares the published bytes.

The binding identifies the current GitHub source snapshot; it deliberately does not claim that the weight bytes can be reproduced from source alone or that an independent party certified model quality. Run the same gate locally without publishing:

python tools/publish_model_source_bindings.py \
  --source-revision "$(git rev-parse HEAD)"

Publication is performed only from protected main by the dependent publish-bindings job in .github/workflows/publish-model-inference-lab.yml, after that same workflow verifies the exact live Space revision and using the repository's encrypted Hugging Face organization credential.

Doctrine gate (dataset content)

tools/validate_sft_dataset.py proves a dataset is well-formed; tools/nemo_doctrine_gate.py proves its content conforms to doctrine. Every training/eval JSONL in this repo is gated through the szl-nemo kernel (rule_check R1–R5) — fabrication labels, honest unknowns, Λ-is-not-a-theorem, the 0.97 trust ceiling, and persona-correct fine-tune provenance — and each run is bound into a deterministic, hash-chained szl.nemo.receipt.v1 ledger. CI enforces it on every PR that touches doctrine-bearing data (.github/workflows/nemo-doctrine-gate.yml, kernel pinned by commit SHA). Current state (MEASURED 2026-09-01): all five gated datasets VALID, 75 records, 0 violations.

Pipeline

szl_dataset.jsonl
      │  (identity + doctrine examples)
      ▼
Unsloth QLoRA fine-tune  ──  train_szl.py  (base: unsloth/Qwen2.5-3B-Instruct)
      │
      ▼
merged 16-bit safetensors  ──  ./szl-model
      │
      ▼
llama.cpp convert_hf_to_gguf  ──  szl1-f16.gguf   (rebirth.ps1)
      │
      ▼
ollama create szl1 --quantize q4_K_M -f Modelfile.gguf
      │
      ▼
serve as SOVEREIGN_MODEL=szl1   (Alloy cockpit runs on SZL-1)

See RUNBOOK.md for the exact commands, VRAM/disk requirements, and Windows-specific notes.

SZL-Nemo (NemoClaw pattern)

The LangChain x NVIDIA NemoClaw Deep Agents blueprint (July 2026) pairs an open model, a tuned agent harness, and a governed runtime — tuned together. SZL's estate maps onto all three layers: open weights on SZL metal (this kit), the Alloy backbone as harness, and SZL's receipt/guardrail stack as governance.

SZL-Nemo is the estate's open-model slot for that pattern: NVIDIA's open nemotron-3-nano:4b (2.8 GB, 256K context — REPORTED from ollama.com) wrapped in the SZL honesty-doctrine system prompt via Modelfile.nemo. Honest tier: a wrapper, not an SZL fine-tune — SZL has not trained these weights, and no benchmarks have been measured on SZL hardware yet. See RUNBOOK-NEMO.md.

Conjecture Machine

conjecture_machine.py points the sovereign stack at SZL's own formula corpus: it loads a formula index (a local snapshot of the estate's thesis-formula-index ships in this repo — 80 entries), iterates each formula, and asks the sovereign OpenAI-compatible endpoint (default https://gpu.a-11-oy.com/v1, model env-selectable — llama3-szl-finetuned-q4:latest or szl-nemo) for an advisory proof sketch, lemma decomposition, and counterexample search. Every attempt is written to conjecture_runs/<timestamp>/<formula_id>.json.

Honest tier, hard-coded:

• It NEVER claims proven. Model output is an advisory sketch only; a formula stays a CONJECTURE unless a real Lean check passes (done in szl-holdings/lutar-lean, not here — the built-in lean_check is an honest stub that never passes). The corpus's own status labels are staged-advisory tags, not live proofs, and are not laundered into "proven".
• Λ uniqueness stays Conjecture-1. The Λ-uniqueness formula (TH10 — Uniqueness of Lutar Invariant) is doctrine-locked and can never be upgraded by this tool.
• Endpoint down ⇒ honest UNAVAILABLE, no fabricated output. Stdlib-only (Python 3.8+, no pip install).

See RUNBOOK-CONJECTURE.md for the one-command-per-step walk-through.

Honesty doctrine

SZL-1 is trained to hold to SZL's honesty doctrine: label claims MEASURED, REPORTED, or UNKNOWN, and let an honest UNKNOWN stand rather than invent an answer. The system prompt baked into Modelfile reinforces this at serving time.

Evaluation status

• ReceiptAgent Qwen2.5 and Khipu have repository-bound owner measurements, but remain limited or research-only according to publishing/model-source-bindings.json.
• ReceiptAgent Qwen3.5 v2 has a small preregistered acceptance gate. Its raw counts are evidence for that contract only, not a broad benchmark.
• Any recipe without a qualifying receipt remains UNAVAILABLE / NOT RUN for performance claims.
• eval_szl.py can produce a local result. Label that result MEASURED only for the exact model, hardware, inputs, and revision actually exercised.