Essay 38: What If I Didn't Build the Graph? – Why Curriculum-as-Code Transcends Flat Docs-as-Code with DOM GraphRAG & W3C PROV-O | Alan Pruitt (Coach Alan)

When I first engineered a Git-versioned Markdown repository with automated CI/CD workflows to manage my multi-course exercise science and clinical nutrition fleet (EXW 101, EXW 150, EXW 210, EXW 245, and EXW 265), I could have stopped at standard Docs-as-Code.

I already had clean Markdown source files, automated Canvas GraphQL synchronization, and a local vector index over 3,151 curriculum chunks. By conventional software documentation standards, my pipeline was finished.

Every time a system I build reaches baseline maturity, however, I pressure-test it against the same guiding counterfactual:

What if I didn’t build the graph?

What if I treated a clinical kinesiology and nutrition curriculum like a flat software manual—slicing pages into isolated 500-token chunks, retrieving them by raw cosine similarity, returning opaque AI confidence numbers to students, and assuming English-only terminology would magically serve the bilingual Hispanic-Serving Institution (HSI) students I teach along the Yuma, AZ / Imperial Valley, CA borderland?

The answer is clear: flat Docs-as-Code without structural pedagogy and cryptographic provenance eventually collapses back into probabilistic guesswork.

Anatomy of the Flat-Vector Trap

In enterprise software documentation, Docs-as-Code treats pages as independent reference articles. If a developer searches for an API parameter, any matching paragraph will suffice.

Higher education—especially clinical exercise physiology and nutrition governed by non-negotiable Federal Single Sources of Truth (PAGA 2nd Ed. 2018, 2025–2030 Dietary Guidelines for Americans, ACSM 12th Ed., and the NASM OPT Model)—operates under far stricter constraints. Leaving a Retrieval-Augmented Generation (RAG) pipeline flat creates four structural fault lines:

Context-Blind Retrieval (Orphaned Clinical Rules): When I slice a Markdown module into flat chunks, a contraindication table or safety gate loses its relationship to the Module’s Course Learning Outcomes (CLOs), its Required Reading foundation, and its parent phase inside The Clinical Framing Protocol (Pattern / Rule / Solve).
Unverifiable AI Synthesis (Zero Lineage): If an AI tutor or clinical auditor synthesizes feedback without a machine-readable provenance graph, neither the student nor an auditor can prove which exact Federal SSoT revision, Git commit, or local model produced the recommendation.
The Black-Box Score Fallacy: Handing a student a single opaque similarity percentage (84% Match) tells them nothing about why their clinical reasoning succeeded or where their safety gate failed.
The Bilingual Vocabulary Gap: When a first-generation student in Yuma or San Luis Río Colorado describes heat exhaustion in authentic Norteño-Spanish (“se le bajó la presión y anda bien mareado en el calorón de Yuma”), a flat English vector index misses the exact clinical mapping to exertional heat illness, orthostatic hypotension, and WBGT > 82°F fluid replacement gates.

Engineering the Upgrade: Four Pillars of S-A-R v2.2

Drawing on my experience as an Adjunct Faculty member of Exercise Science & Wellness at Arizona Western College, a Generative AI Strategist at The University of Arizona, and Principal of Webcognita LLC (VOSB), I independently designed the S-A-R Standard (Scalable / Auditable / Repeatable) so that I can close every architectural gap deterministically across my scalable $0 ZTC Open Educational Resource (OER) course fleet.

Instead of settling for flat Docs-as-Code, I upgraded my EXW Master Library (v2.2) with four open-standard semantic layers running on air-gapped Google Gemini Argon 4 models locally on my Apple Silicon (M1/M2) workstations for 100% FERPA privacy:

+-----------------------------------------------------------------------------+
|            CURRICULUM-AS-CODE (S-A-R v2.2) SEMANTIC GRAPH ENGINE            |
|                                                                             |
|  [1. W3C SKOS Ontology]  ->  Bilingual Norteño-Spanish <-> Clinical SSoT    |
|  [2. DOM GraphRAG]       ->  3,151 Nodes Linked by CLOs & Pattern/Rule/Solve|
|  [3. Born-Accessible XAI]->  9:1 Contrast 4D Radar (Pattern/Rule/Solve/Ver.)|
|  [4. W3C PROV-O Receipt] ->  Cryptographic SHA-256 Lineage (SSoT -> Argon4) |
+-----------------------------------------------------------------------------+

1. DOM GraphRAG (Structural Pedagogical Graph)

Rather than indexing isolated text fragments, my indexer (scripts/index_corpus.py) parses the semantic Document Object Model (DOM) of every curriculum file I author. I enriched every one of my 3,151 curriculum vectors with a dom_graph_node payload containing:

Its exact dom_path (h2 > h3 semantic hierarchy),
Its governing role inside The Clinical Framing Protocol (Pattern / Rule / Solve),
Its explicit Course Learning Outcome (CLO) alignment, and
Directed graph_edges (parent_module, requires_reading, grounds_lab, grounds_discussion, and governed_by_ssot).

When a student audits a lab inside Canvas, my engine retrieves not just a matching paragraph, but the complete pedagogical subgraph linking the reading, the lab rubric, and the federal guideline.

2. W3C PROV-O Cryptographic Provenance Receipts

Every live Sovereign AI Auditing response returned by my local API (scripts/canvas_rag_api.py) now emits a formal W3C PROV-O (http://www.w3.org/ns/prov#) JSON-LD provenance graph (prov_o_receipt) alongside a deterministic SHA-256 receipt_id:

prov:Entity (exw:SSoTCorpus): The exact federal guideline (PAGA 2018, DGA 2025-2030, ACSM 12th Ed., or NASM OPT) and the Git-versioned curriculum chunks retrieved.
prov:Activity (exw:ClinicalAuditSynthesis): The timestamped execution of the local RAG + Clinical Framing audit.
prov:Agent (exw:SovereignArgon4Auditor): My air-gapped local Google Gemini Argon 4 engine running on Apple Silicon (M1/M2) with zero external cloud telemetry.

3. Born-Accessible (`9:1` Contrast) 4-Dimensional XAI Radar Scorecard

Under my ADA Title II & Atomic WCAG 2.2 AA / 3.0 Higher Ed Architecture, explainability must be born accessible. Inside the live Canvas Clinical Auditor Modal, I replaced single-number confidence scores with a 4-Dimensional Explainable AI (XAI) Radar Scorecard engineered to my strict 9:1 Double-Contrast Guard (#FFFFFF on #090D16 = 19.3:1; #38BDF8 sky cyan = 9.82:1; #34D399 mint = 10.45:1; #FBBF24 amber = 11.85:1):

Pattern (Context & Telemetry — 25%): Client vitals, biometrics, and regional Yuma / Imperial Valley heat stressors.
Rule (Federal SSoT Citation — 25%): Explicit quantitative anchors from PAGA, DGA, ACSM, or NASM.
Solve (Clinical Intervention — 25%): Concrete FITT-VP or “Real Food” prescription with safety gates.
Active Verification (150-Word Defense — 25%): Authentic human justification in the student’s own voice.

4. W3C SKOS / Schema.org Bilingual Clinical Controlled Vocabulary

To honor my commitment to delivering bilingual “In Plain English” OER across borderland HSI communities, I authored a formal W3C SKOS (Simple Knowledge Organization System) and Schema.org (MedicalEntity) controlled vocabulary (01_INFRASTRUCTURE/skos_bilingual_clinical_taxonomy.json).

Colloquial Norteño-Spanish clinical expressions (skos:altLabel @es-MX-Norteno) map deterministically to English and Spanish clinical terms (skos:prefLabel) and federal SSoT rules (skos:definition). When a student queries my RAG pipeline in Norteño-Spanish, the SKOS layer automatically expands the vector query across languages—lifting top clinical retrieval similarity from 0.7743 to 0.8236 without requiring the student to abandon their community’s authentic voice.

Slow is Smooth, Smooth is Fast

What if I didn’t build the graph?

Without DOM GraphRAG, W3C PROV-O, Born-Accessible 9:1 XAI, and W3C SKOS bilingual ontologies, a course repository is just a static folder of text waiting to be summarized by commercial chatbots.

By binding every paragraph, bilingual clinical term, and AI audit receipt into an auditable semantic graph, my Curriculum-as-Code architecture defends academic rigor, guarantees 100% FERPA privacy, and expands equitable access for every student in my borderland classrooms.

Disclaimer: The Curriculum-as-Code architecture, S-A-R Standard, and open-source software tools described in this essay represent my independent faculty engineering and research through Webcognita LLC and do not represent official policy or endorsement by Arizona Western College or The University of Arizona.

Explore My Architecture Briefs & White Papers

All six of my formal architecture briefs and technical white papers—including the EXW S-A-R v2.2 DOM GraphRAG & W3C PROV-O Architecture Brief and my AWC 2026 Innovation Fund “Project Horizon” White Papers—are available for direct download in the Architecture Briefs & Grant White Papers shelf on alanpruitt.com.