Clinical Epistemology · folder 11 · knowledge graph engineering

Folder 11, mapped

Folder 11 is the odd one out, and deliberately so. Folders 01–10 each take a clinical problem and apply the portfolio's discipline to it. Folder 11 takes the portfolio itself as its subject and asks: what if all of this were one queryable structure instead of eleven folders of documents?
Built July 24, 2026 Status design only — nothing implemented Contents 2 notes · 1 source book · 2 source articles Instance MB-M1
Every other folder produces something — a framework, a corpus, a validation report. Folder 11 is the first folder that consumes all of them. It is the plan to turn ten folders of artifacts into one structure you can query, audit, and compute over.
Which is why it reads differently: no app, no data, no worked cases. Two design notes and three sources. It is a plan for infrastructure, and its value depends entirely on whether the infrastructure gets built.
01

What folder 11 is — and what it is not

Three clarifications that prevent the most likely misreadings.
Not a twelfth clinical framework.
It is a meta-layer. Folders 01–10 reason about patients; folder 11 reasons about folders 01–10. If it were numbered honestly it would be numbered 00 or ∞, not 11.
Not built. Not started.
A v0.1 design note plus a strategy update. No code, no graph, no database. Everything below describes intent, and the folder says so explicitly.
Not “a graph.”
Three different things in the folder are called a graph, and they stack rather than compete. Keeping them apart is most of the folder's conceptual work — see next section.
02

Three things called “graph”

Each arrived from a different source, each answers a different question, and conflating them is how this kind of project goes wrong. The folder's own framing: they stack.
The artifact

Knowledge graph

The durable data structure — the memory an agent reasons across. What to build.
Nodes = entities · Edges = typed, provenance-carrying facts
Pipeline: extract → resolve → assemble → query, built from structured-output calls. No trained NER, no relation classifier, no graph database required to start.
Source article A — the KG pipeline “playbook”
The execution

Orchestration graph

The shape of the work that assembles and interrogates the artifact. How to build it.
Nodes = agent jobs · Edges = data dependencies
A linear “do A then B then C” script is a degenerate graph — steps that never needed to wait are queued anyway. Control flow lives in code, so it runs the same way every time.
Source article B — the 14-step orchestration roadmap
The reasoning

Decision network

The formal reasoner layered on top of the artifact. What it's all for.
Nodes = chance (patient state), decision (next-best-action), utility (benefit/harm)
Turns the graph from something you traverse into something you compute over: posterior probabilities, expected-utility rankings, and a quantified abstain decision.
Source book — Algorithms for Decision Making (MIT, 2022)
The stack, in one line: The knowledge graph is the substrate; the orchestration graph is the builder; decision theory is the reasoner that makes “earned recommendation” a computable object rather than a slogan.
03

Folder 11 relative to the rest — what feeds what

this is the relationship
This is the answer to “how do I think about folder 11 next to everything else.” It isn't a sibling of the other folders — it's downstream of all of them. Each folder already contains material that becomes nodes, edges, or test cases. The striking part: nothing here needs to be newly labeled.
Source folder
What it contributes
Lands in
Patients/ canonical cohort · 18 patients
Master JSON, timelines, labs, dated ADT events, NBA snapshots — already structured. Loads directly as nodes and edges with zero model calls and zero extraction error.
Clinical graph
deterministic backbone
05 · hospital packets 6 packets + answer keys
Prose charts for genuine LLM extraction — and machine-readable ground truth that scores that extraction. Doubles as the evaluation harness.
Clinical graph
+ eval gold set
08 · synthetic labs 78 documents
Scanned lab, imaging and pathology reports — the OCR-and-extract target, and the reason the graph must carry document-level provenance on every edge.
Clinical graph
extraction target
09 · aberrant labs 40 implanted errors
The adversarial set. Enables something the source playbook doesn't even propose: a graph-level coherence benchmark — does graph reasoning flag impossible-for-life values and internal contradictions? A scoreable version of BRIDGE's “this doesn't add up.”
Verifier benchmark
the adversarial test
07 · CRRF nine gates · 46 refs
The nine gates become Gate nodes; the verified evidence base becomes EvidenceSource nodes. This is what makes “earned-through” an actual traversable edge rather than a claim in a white paper.
Epistemic overlay
gates + evidence
06 · BRIDGE stages · certainty · unknowns
Six stages, the 0–100 certainty scale, and the known-unknowns. Contributes two of the most distinctive edge types: drifted-to (the diagnostic trajectory) and silent-on (the chart's silences).
Epistemic overlay
trajectory + silences
04 · fluid framework 17 modules · confidence axes
Clinical-question rows, the five confidence axes, the next-best-actions, and the row-to-reference map — another already-structured artifact that loads deterministically.
Epistemic overlay
NBAs + axes
01 · 02 · orchestrator + HTN modules · NBAs
The multi-module architecture is already the orchestrator–workers pattern the graph serves as shared memory for: each module writes typed facts, a synthesizer traverses instead of re-reading source documents.
Epistemic overlay
modules
10 · CHAI governance risk modifiers
Risk modifiers and T&E chapters become governance nodes — so a gate can operationalizes a named CHAI modifier and the compliance story becomes a query.
Governance layer
audit surface
Why this matters: The usual blocker for a project like this is that there's no gold set and no labeled data, so you tune prompts blind. Here the gold sets were built first, for other reasons — the packets' answer keys, the 40 implanted errors, the verified evidence base. Folder 11 starts with its evaluation harness already in hand.
04

The core design — two coupled graphs

The clinical graph is commodity: plenty of people build one. The epistemic overlay is the part that is specific to this portfolio, and it's where the argument lives. They join at a single node type.

Clinical world-model graph

Commodity, but necessary — the facts
Patient Encounter Condition Medication LabResult ImagingStudy Finding Procedure Provider

Encounters are ADT-typed and timestamped; every edge carries provenance back to the source document. Resolved against real terminologies so it can interoperate with actual EHR data later.

Epistemic overlay graph

The differentiator — the reasoning objects
Framework Module Gate BridgeStage ClinicalQuestion NextBestAction ConfidenceAxis OutputClassification KnownUnknown EvidenceSource CHAIRiskModifier

Including an explicit Unable-to-determine node — a first-class node, not an absence. That single schema decision is the portfolio's whole ethos expressed in data modeling.

Where they join — the next-best-action node

A recommendation node sits at the seam: it points down into patient facts and up into gates and evidence. That makes the signature query possible — from a patient's missed-dialysis hyperkalemia event → to the next-best-action → to the gate that cleared it → to KDIGO 2024. For a CMIO defending a CDSS in governance review, “show me every edge from this recommendation back to its guideline and the gate that authorized it” is the artifact you actually want.

Two signature outputs fall out for free

A diagnostic trajectory is simply a time-ordered path of drifted-to edges. The known-unknowns list is simply the set of silent-on edges out of an encounter. Neither needs special machinery — they're consequences of the schema.

Predicates must be a controlled vocabulary

earned-through gated-by grounded-in supported-by contraindicated-by mitigates operationalizes escalates-to drifted-to silent-on has-certainty. Free verb phrases erode traversability.

05

Four places the clinical domain inverts the generic playbook

This is where folder 11 stops being a summary of two articles and starts being domain judgment. Each of these departs from the source material on purpose.
Inversion 1

Deterministic backbone before any LLM extraction

A large fraction of the portfolio is already structured — ground-truth JSON, the reference map, the gate list, the CHAI modifiers, patient master JSON and timelines. Load those directly: no model call, zero extraction error. Reserve extraction for genuine prose — packets, case analyses, the white paper, personas.

Inversion 2

Resolve against standards, not just LLM clustering

Clinical entities have ready-made canonical spaces: medications → RxNorm, labs → LOINC, conditions and findings → SNOMED CT or ICD-10. Use those as alias maps and fall back to model clustering only for what the terminologies don't cover. Side benefit: real-EHR interoperability later.

Inversion 3

Favor recall over precision — the opposite of the source default

The playbook tunes to “central entities only,” because a false entity spawns false relations. In medicine the asymmetry runs the other way: a missed medication or lab is the safety failure, and a false one gets caught downstream by the gates and verifiers. Raise recall; let the machinery filter.

The gates are what make this safe — remove them and this inversion is reckless
Inversion 4

Certainty and provenance are native, not future work

The source article lists edge-level confidence and temporal edges as future directions. This portfolio already has the rubrics — BRIDGE's 0–100 scale, the five confidence axes, the soft-trigger discipline. So attach certainty and source provenance to every edge from day one rather than retrofitting.

06

How it would get built

The construction topology is a diamond, not a chain — and the second node is the one that matters most for this project specifically.
Fan out

Extract

One agent per document, running concurrently, each returning schema-validated JSON. The structured-output schema is the node's contract.

★ the key idea

Verifier on the edge

A node whose only job is to try to kill a finding before it's allowed downstream. A fact enters the graph only after an adversarial check fails to refute it.

Barrier

Resolve

Cross-document entity deduplication — one of the few places a genuine barrier is justified, because it needs every document's results at once.

Reduce

Synthesize

Assemble the graph, summarize high-degree hub nodes across every source document.

Why “verifier on the edge” is the highest-value idea here

It is the portfolio's own epistemic discipline, applied one layer lower than it has ever been applied before — at data ingestion rather than at recommendation time. “A fact must earn the right to be asserted” is the construction-time sibling of the CRRF gates and BRIDGE's coherence verdict. And the elegant part: the verifiers' verdicts get written back as the certainty and provenance metadata on the edges — so the construction machinery's byproduct becomes the graph's first-class trust signal, which is exactly what the governance layer needs to surface.

1

Scope to one spine

The 18 patients plus the 6 hospital packets, where answer keys already exist. Don't boil the ocean.

2

Deterministic backbone

Load the already-structured artifacts. NetworkX is plenty at this scale — hundreds of nodes.

3

Prose extraction

Run structured-output extraction over packets and case analyses; resolve new entities against the backbone, not against each other.

4

Epistemic overlay

Wire recommendations to gates and evidence; add the Unable-to-determine and KnownUnknown nodes.

5

Evaluate

Score extraction against the packet keys; run the aberrant-labs coherence benchmark.

6

Migrate — only if justified

Move to a real property graph only when provenance queries for governance earn it. That's where Neo4j pays for itself, not before.

07

Caveats and open decisions

What the folder itself flags as unresolved or dangerous.

The caveats

Named in the design note, worth keeping visible
Synthetic only
Even the synthetic corpus carries PHI-shaped tokens. The graph must never mix in real patient data without de-identification and provenance controls — and that constraint is itself a governance edge worth modeling.
Graph informs,
gates decide
The knowledge graph is a fact-and-evidence store, not a decision-maker. A wrong edge that reaches a recommendation unchecked is precisely the failure mode the whole portfolio exists to prevent.
Predicate
discipline
Free verb phrases erode traversability. Keep the controlled vocabulary, and ideally standard entity codes.
Quantification
From the decision-theory layer: full Bayes nets and POMDPs need conditional probability and utility numbers. 18 synthetic patients cannot learn them. Near-term this is scaffolding and selective formalization, not a rebuild.

The open moves

Nothing here is started
1
Prototype the workflow on one packet, end to end. One document → extract → verify → into-graph, so the topology is visible on a single case before anything scales.
2
Formalize one decision as a decision network. The fluid-overload NBA, or a single BRIDGE known-unknown, with a real value-of-information calculation — the reasoning-layer analogue of the one-packet prototype.
3
Decide storage. NetworkX now versus Neo4j later — resolvable once the epistemic overlay's provenance queries are actually drafted, not before.
4
Write the schema tables and predicate dictionary. The design note sketches both; neither exists as a reference document yet.
08

The honest read on folder 11

What makes it strong

Why this is more than a plan to build a database
It's downstream
It doesn't ask for new material. Ten folders of artifacts already exist and most of the highest-value inputs are already structured.
Eval exists
The single most common failure — no gold set, so you tune blind — is already closed. The answer keys and the 40 implanted errors were built first, for other reasons.
It's coherent
with the ethos
Verifier-on-edge isn't a borrowed engineering pattern; it's the same discipline as the CRRF gates, applied at ingestion. The project's idea reappears one layer down, which is a good sign the idea is real.

What to watch

Where a plan like this usually fails
Nothing built
Two design notes and three sources. Every claim here is architectural intent. The gap between this and a working graph is the entire project.
Scope gravity
The plan explicitly says “don't boil the ocean,” which is the right instinct precisely because the material invites the opposite. Phase 1 is 18 patients and 6 packets — holding that line is the test.
The recall bet
Choosing recall over precision is defensible only if the gates and verifiers genuinely filter. That's an assumption the aberrant-labs benchmark is supposed to test — and hasn't yet.