Module CAP1 — The Air-Gapped Domain Model

The Air-Gapped Domain Model

A reproducible pipeline: open-data base → HIPAA-safe data prep → QLoRA fine-tune → eval → GGUF export → local serve. Runs on a single consumer GPU or Apple Silicon, fully local, zero telemetry.

120

minutes

artifacts

sub-sections

Capstone 1 integrates Pillars P00–P06 into one pipeline: choose an open-data base, prepare HIPAA-safe data, QLoRA fine-tune, evaluate for domain lift AND general-benchmark forgetting, export to GGUF, and serve locally via Ollama/llama.cpp. The deliverable is a reproducible GitHub README asset that a reviewer can clone and run. The property it proves — that no individual module could prove alone — is reproducibility under air-gap: a model that belongs to its domain, not to a vendor.

Key Claims

Load-Bearing Claims

Reproducibility under air-gap is the deliverable. A model you fine-tuned on your hardware and serve through Ollama, from a pipeline you can hand to a clinician, lawyer, or analyst, is a model you own. A cloud-trained, API-served model is a model someone else controls. The air-gap is a property of the whole pipeline, not a deployment feature.

Two evals are required, not one. Domain lift (fine-tuned minus base on held-out) proves steering worked; forgetting (base minus fine-tuned on a general benchmark) proves the fine-tune was disciplined. A submission with only the lift is incomplete — the forgetting number is the discipline check.

The held-out set is the steering-vs-memorization test. If the lift vanishes on held-out, you memorized, you did not steer. The held-out set exists to catch the cardinal error of treating fine-tuning as knowledge injection.

HIPAA-safety is a gate, not a trade-off. A submission with unclear data provenance fails regardless of the total score. The base must be open-data (auditable), and the data prep recipe must be publishable.

After This Module

Assemble a reproducible end-to-end pipeline — open-data base, HIPAA-safe data prep, QLoRA fine-tune, domain eval, GGUF export, local serve — that runs on a single consumer GPU or Apple Silicon with zero telemetry.

Defend the three success criteria: domain lift versus the base, successful local serve, and reproducibility from a clean clone.

Apply the steering thesis to distinguish where fine-tuning earned its gains (behavior) from where it cannot (knowledge), and design the eval accordingly (held-out lift + general benchmark forgetting).

Produce a GitHub README asset: a pipeline someone else can clone, run, and reproduce your reported numbers — the portfolio artifact.

Artifacts

Teaching Document

~3,500 words; 6 sub-sections — the capstone goal, success criteria, evaluation rubric, pipeline architecture (7 phases), solution key, why it matters; capstone-style design guidance, not theory

READ

Diagrams

5 diagrams — the seven-phase pipeline, the three success criteria, lift-vs-forgetting eval dual, the air-gap property, the evaluation rubric

READ

Slide Deck

reveal.js slide deck, dark theme, design-system teal

READ

Teaching Script