⚠ PROPOSED — a capability to evaluate, not current practice

Knowledge Vault for Study Operations

A local‑first Markdown vault — the same idea behind Obsidian, Foam and Logseq — that turns our Study Knowledge Base into a navigable, linked map of a study's operational knowledge, and doubles as a clean corpus for the on‑prem model. This note shows what it is, how it slots into what we already have, the governance, and which tool to pick.

◉ Open the live knowledge graph 📁 Browse the 28‑note vault

The CP-101 study knowledge graph — a gold hub at the centre with colour-coded clusters of linked notes arranged in named regions

The interactive graph of a synthetic study (CP‑101). Hub at the centre; colour = note type; clusters settle into named regions. Hover to focus a note and its links; click for the full note + backlinks.

Hub Design Statistics Data Process Safety Supply Program Reference

01What it actually is

Not a new system — a folder of plain text the rest of our stack can already use.

🔗

Linked notes

Each artifact is a Markdown note joined to others with [[wikilinks]]. Backlinks answer "what references this DTA?" in one click.

🕸️

Graph view

The links draw a map of the study. Navigate by relationship, onboard fast, and spot orphaned or missing connections.

🤖

RAG‑ready

Clean Markdown is the ideal corpus for the on‑prem model — "ask the study" with grounded, human‑gated answers.

02One vault, three jobs

It earns its place by doing three things at once, all over the same files.

① A graph layer over the Study KB

The auto‑ingest engine writes each entry as a .md note linked to a per‑study hub — open the folder and the graph is just there.

extends what we built

② A living team knowledge base

SOPs, naming conventions, the glossary, lessons‑learned — editable, cross‑linked notes. The HTML wikis stay as the read‑only published face.

grows over time

③ The corpus for the local model

Point the frozen on‑prem SLM at the vault. Same retrieval pattern as the local‑SLM operating wiki — and it still never sources a number.

feeds the SLM

03How it wires into what we have

Outlook → the existing auto‑ingest → a Markdown vault → people and the model read the same files.

  Outlook study mailbox  ──▶  Auto-ingest engine        ──▶   Markdown vault (.md + [[links]] + git)
   (ops emails only)            (= your Study KB)               ├─ Ingested  · read-only
                                                                └─ Team notes · editable
                                                                       │
                                          ┌────────────────────────────┴───────────────────────────┐
                                          ▼                                                          ▼
                              Obsidian / Foam / Logseq                                     Local frozen SLM
                              backlinks · search · graph                                   RAG over the vault

Plain Markdown + git gives versioning and a frozen history — which is exactly the sovereignty / reproducibility story the rest of the program already makes.

Same governance as the Study KB — non‑negotiable. The vault is operations only: never PHI, participant‑level data, unblinded assignments, or sponsor‑restricted IP. Every note carries an "informational ops support — not a source of record" line. Reported and regulated numbers come from validated tools (Phoenix WinNonlin, Pinnacle 21, the validated pipeline) and a human signs them — the vault never produces a number. A graph holds the wrong thing as easily as a mailbox does, so the inbound‑rule + watermark governance carries over unchanged.

04Which tool — and why it barely matters

The vault is just Markdown files in git. You own the files; the app is only a viewer you can swap.

Tool	Licence / IT story	Strengths	Watch‑outs
Foam ◀ recommended start	MIT, free. Runs inside VS Code — which biostatisticians often already have approved.	Lowest‑friction IT path; wikilinks + graph; lives next to your code.	Graph is good, not as lush as Obsidian's.
Obsidian	Free for personal; commercial‑use licence required at work.	The best graph/UX and plugin ecosystem; the most polished demo.	Third‑party app + a licence; community plugins are unvetted code — allow core + a small vetted set, or none.
Logseq	Open‑source, free.	Outliner‑style; strong daily‑notes / journaling.	Different mental model; smaller ecosystem.

The bundled Knowledge_Graph.html needs no install at all — it's a self‑contained viewer over the same vault. Use it for demos and read‑only sharing; use Foam/Obsidian when people need to edit.

05Honest about the graph

The graph is for navigation, onboarding, and spotting gaps — not analysis. It's genuinely useful (and it demos beautifully), but the load‑bearing value is the linked notes + search + RAG, not the picture. And to be explicit: this is standard knowledge‑management graphing — it has nothing to do with any research methods, and is kept entirely separate from them.

06Try it · reproduce it

See it — open the live graph; hover a note to focus it, click for the full note and its backlinks, use the legend to isolate a note type.
```
open Knowledge_Graph.html
```
Open it as a real vault — point Obsidian or Foam at the vault/ folder; its own graph view will match this one (same [[links]]).

Regenerate it — the vault is generated from a single source of truth (graph_model.json), so the graph is built, not hand‑drawn:

python3 assemble_vault.py        # rebuild the 28 notes
python3 kb_to_vault.py --demo    # production path: live Study KB entries → vault