Clinical Epistemology · folder 11 × Enterprise Clinical Intelligence Framework

Same object, opposite ends

ECIF was built top-down from organizational strategy and arrived at a layer it named but did not specify. Folder 11 was built bottom-up from existing artifacts and specified a structure without naming what it was for. They are the same thing, and this is where they meet.
Built July 26, 2026 Sources ECIF v0.1 · folder 11 docs 01–03 Status both design-stage; neither built
01

The structural claim — folder 11 lives inside ECIF layer 3

Everything folder 11 describes — all three of its layers plus its substrate — fits within a single band of the ECIF stack. That containment is the whole relationship in one picture.
ECIF
top-down

Starts from “where does durable advantage come from?” and descends. Names layer 3 as the moat — but an asset stays an abstraction.

↓
Layer 5
Outcomes
Better care, lower cost, operational improvement
moat: highest
measure & improve
Layer 4
Applications
Clinical workflows and user experiences — they consume the layer below
moat: high
partner
Layer 3Clinical Intelligence Layer ★ primary ownership · moat: very high
ECIF: “not an application — a reusable enterprise capability… assets that encode how the organization thinks.” Folder 11 is what that capability looks like as a data structure:
3
Decision-theoretic reasoning layer
Decision networks, belief states, value of information — posterior probabilities and a quantified abstain
ECIF's confidence
model, made formal
2
Epistemic overlay graph ← the intelligence asset
Gates, evidence, NBAs, confidence axes, KnownUnknown, Unable-to-determine, CHAI modifiers
ECIF's asset spec,
as a subgraph
1
Clinical world-model graph
Patients, encounters, labs, meds, findings — provenance on every edge, resolved to RxNorm / LOINC / SNOMED
ECIF's enterprise
data mapping
0
Substrate — folders 01–10 + Patients/
Five frameworks, four corpora, the governance overlay. Already built; already the gold sets
ECIF's “map existing
work on, don't rebuild”
The identity that makes the two documents one document ECIF lists “knowledge graph relationships” as a component of an intelligence asset, then never says what an asset physically is. Folder 11's answer: an intelligence asset is a subgraph — reasoning spec, evidence, confidence, failure modes and governance as typed nodes with provenance-carrying edges. That converts “we own clinical intelligence” from a strategy claim into an inventory you can point at, version, audit, and hand to a successor.
Layer 2
Foundation Models
Increasingly commoditized — remain model-agnostic
moat: moderate
Layer 1
Compute
Consume as a utility
moat: low
↑
Folder 11
bottom-up

Starts from eleven folders of artifacts and asks what structure makes them queryable. Builds layer 3 without knowing it was layer 3.

Trust — the vertical spine ECIF runs it across all five layers: data quality · provenance · explainability · governance · monitoring · validation · auditability · clinical safety · security · regulatory. In folder 11 this is not a separate layer — it is edge metadata: every edge carries certainty and source provenance from day one, and the verifier nodes that admit a fact write their verdicts back onto it.
02

Construct-by-construct crosswalk

Where the two vocabularies name the same thing, where each repairs a genuine gap in the other, and where neither has an answer. The direction column is the useful part — this is not a one-way fit.
✓ same — one construct, two vocabularies ← ECIF fixes — ECIF supplies what the portfolio lacks → portfolio fixes — the portfolio supplies what ECIF leaves as one line ⚠ gap — neither has it
ECIF constructPortfolio / folder 11 constructDirectionThe point
Clinical intelligence asset An epistemic subgraph — typed nodes + provenance edges ✓ same The central identity. ECIF says assets are the moat but not what one is; folder 11 gives the storage format.
Epistemology Layer — what do I know · how do I know it · how certain am I · what don't I know World-model nodes · provenance edges · has-certainty · KnownUnknown + silent-on ✓ same Four for four. The tightest join in either document: ECIF states four questions as a requirement, folder 11 makes them queries rather than prose.
Context Sufficiency — proceed / request more / escalate / defer / abort CRRF's six earned-authority tiers, T0 Silent → T5 Actionable → portfolio fixes Near-isomorphic, but ECIF's five read as exclusive branches. CRRF's move is that failing caps the ceiling rather than branching — a system that can't see the whole picture may still request data, it just may not advise.
Problem Framing — six dimensions CRRF Gate 1 — two dimensions (question + unit) ← ECIF fixes ECIF's “entity” is CRRF's unit, so ECIF adds four Gate 1 never had. Two matter: user persona and cognitive task.
Cognitive task taxonomy — summarization · sense-making · recommendation · prediction · explanation · population — nothing equivalent ← ECIF fixes Explains the portfolio to itself: BRIDGE is sense-making, CRRF is recommendation, patient summary is summarization. They resisted unification because they are different cognitive tasks, not different diseases.
Prediction ≠ Reasoning ≠ Recommendation Chance node → belief state → decision node + utility ✓ same ECIF asserts the separation; decision theory says why — you cannot cross from a probability to an action without a utility function.
Confidence model (one line in the spec) BRIDGE 0–100 with bands · five confidence axes · posteriors scoreable by Brier / log loss → portfolio fixes The portfolio has instruments where ECIF has a slot — and the decision-theory layer makes them calibratable, so “confidence” becomes measurable rather than declared.
Validation cases · failure modes (one line each) 6 packet answer keys · 40 implanted aberrant-lab errors · KnownUnknown nodes → portfolio fixes Already-built gold sets. The usual blocker — no labeled data, so you tune blind — is closed before either project starts.
Multi-level: patient → population → facility → region → enterprise CRRF unit-agnostic — “the unit changes, the steps don't” ✓ same ECIF extends a move already made in July, adding three levels above population. The graph is what makes it operational — facility and region are node types with aggregation edges.
Trust spine CHAI overlay (folder 10) + certainty/provenance on every edge ✓ same Partially instantiated already. Folder 11 treats trust as edge metadata rather than as a separate layer, which is the stronger form.
— no utility or values component — no utility or values component ⚠ gap The most consequential hole. Both have probability; neither has benefit/harm. Every recommendation the architecture produces asserts a utility function nobody has written down — and in dialysis the utility on a missed-treatment intervention is nothing like the one on a transplant nudge.
Asset lifecycle — five owners, versions, monitoring, drift detection — no Version, Owner or MonitoringMetric node types ⚠ gap ECIF requires what folder 11 doesn't model. Concrete version: when a guideline is superseded, every grounded-in edge pointing at it needs revisiting. Evidence drift has no home in the schema.
03

What doesn't close — and the one move that tests all of it

Four seams

Real gaps, not stylistic ones
No values
layer
Confidence is the probability half of a recommendation. The utility half — benefit versus harm — is absent from both documents. Adding a values component to the intelligence-asset spec is probably the single highest-value edit to ECIF.
No lifecycle
in the graph
Owners, versions, monitoring and drift are required by ECIF and unmodeled in folder 11. Evidence drift is the sharp end: superseded guidelines silently invalidate edges.
Scope
chasm
Folder 11 phase 1 is 18 synthetic patients and 6 packets, and its own caveat forbids mixing real patient data without de-identification and provenance controls. ECIF is enterprise-scale on real data and implicitly requires exactly that. This bridge is not small.
Two words
overloaded
“Stack” — ECIF's five layers are organizational and economic; folder 11's three are computational and sit inside ECIF layer 3. “Epistemic” — the repo, ECIF's Epistemology Layer, and the epistemic overlay graph are three scopes with one word. Disambiguate before this circulates.

One vertical slice

Three open “prototype one thing” moves are secretly the same move
1
Write it as a specification. Take the fluid/volume framework — the most complete asset in the portfolio — and populate ECIF's Clinical Reasoning Specification template in full: question, required and optional context, sufficiency rules, evidence sources, heuristics, confidence, failure modes, escalation, validation, monitoring.
2
Load it as a subgraph. Instantiate that same specification in the folder-11 epistemic overlay — gates, evidence, NBAs, confidence axes, known-unknowns — and check whether the schema can actually hold everything the specification asked for. Where it can't is the schema's real to-do list.
3
Formalize one decision inside it. Take a single fluid-overload next-best-action and build it as a decision network with an explicit value-of-information calculation and clearly-labeled probabilities — which forces the missing utility question into the open on one concrete case instead of in the abstract.
Why this and not three separate pilots: each layer is currently justified by the layer above it and validated by nothing. One slice through all three — specification, subgraph, decision network — on one framework you already know cold is the cheapest way to find out which layers are load-bearing and which are elegant description. It also produces the first genuine specimen of an intelligence asset, which is the artifact the strategy argument needs and currently doesn't have.