Clinical Epistemology · working map

What all of this actually is

Eleven numbered folders, five reasoning frameworks, four synthetic corpora, and a governance overlay — but one argument, applied over and over, and now being formalized into something computable. This page is the map of that argument.
Built July 24, 2026 Instance MB-M1 Scope whole portfolio All patient material fully synthetic
The whole portfolio is one claim: a system must earn the right to make a recommendation — through explicit, sequential, inspectable gates — before it interprets anything. And “I cannot determine this” is a first-class output, not a failure.
Everything else is that claim wearing different clothes. The frameworks apply it to different clinical problems. The synthetic corpora exist so the claim can be scored instead of asserted. The governance overlay maps it onto CHAI. And folder 11 — the newest work — is the attempt to stop describing “earned” in prose and start computing it.
01

The spine — four beats

If someone gives you ninety seconds, this is the whole thing.
The claim

Refusal is a valid answer

Most clinical AI is built to always produce something. This portfolio starts from the opposite premise: the honest exit — “I cannot determine this, and here is what would change that” — is a complete, correct output.

The machinery

Gates that cap, not gates that stop

Nine sequential readiness checks. Failing one doesn’t halt the system — it caps how much authority the system has earned. A tool that can’t see the whole picture may still ask for data; it may not advise.

The proof

Ground truth you can grade against

18 synthetic patients, 6 hospitalization packets with machine-readable answer keys, 78 lab and imaging documents, and 40 deliberately implanted implausible values. The claim becomes a score, not a position.

The formalization

From slogan to computable object

A knowledge graph as substrate, and — new as of today — decision theory as the reasoner on top. “Earned” becomes maximum expected utility with an explicit abstain action, gated by value of information.

02

The three-layer stack

folder 11 · design only
This is the architecture the newest work describes: three layers of reasoning sitting on a substrate that — importantly — already exists. Read it top-down as “what the system does,” or bottom-up as “what has actually been built.”
3
Reason

Decision-theoretic reasoning layer

Bayesian and decision networks (chance nodes = patient state, decision nodes = next-best-action, utility nodes = benefit/harm), belief-state / POMDP reasoning for diagnoses that shift over time, and value of information to decide whether to act now, gather more, or decline. This is what turns the graph from something you traverse into something you compute over.

design note
added 2026-07-24
▲ computes over ▼
2
Judge

Epistemic overlay graph — the differentiator

Nodes are reasoning objects, not clinical facts: CRRF gates, BRIDGE stages, next-best-actions, confidence axes, CHAI modifiers, the evidence base, and an explicit first-class Unable-to-determine. Edges are epistemic — earned-through, grounded-in, operationalizes, drifted-to, silent-on. Layer 1 is commodity; this is the layer that isn’t — it’s what makes the portfolio more than a chart summarizer. (The novelty position itself was red-teamed in folder 07.)

design v0.1
added 2026-07-23
▲ joined at the recommendation node ▼
1
Know

Clinical world-model graph

The commodity layer: patients, encounters, labs, medications, findings — every edge carrying provenance back to a source document and a certainty value. Resolved against RxNorm / LOINC / SNOMED. The new insight from today’s work: this graph is not only a fact store, it is a candidate Bayesian network — so design the edges now to carry conditional-dependency meaning, not just provenance.

design v0.1
added 2026-07-23
▲ built from ▼
0
Substrate

What already exists — the fuel and the gold sets

18 canonical synthetic patients · 6 scored hospitalization packets with answer keys · 78 lab, imaging and pathology documents with 1,682 PHI tokens · 20 aberrant reports with 40 implanted errors · a 46-source verified evidence base · five working framework apps. No new labeling is needed to start evaluating.

built & live
folders 01–10
Read the honesty here: built  Layer 0 is real and running. designed  Layers 1–3 are written down, not implemented. That distinction is deliberate — it’s the same discipline applied to the project itself.
03

The eleven folders, grouped by what they’re for

The numbering is chronological, which hides the structure. Grouped by role, the portfolio reads as: five frameworks that reason, four corpora that test them, two layers that check them, and one folder describing where it all goes next.

A · The frameworks that reason

The argument applied to five different clinical problems. Each one is a working artifact, not a concept paper.
02
Hypertension Management Framework
The original proof of “earned recommendation” — checklist plus next-best-action. The simplest complete statement of the idea.
app · live v2.0
01
CKD Care Orchestration Framework
The flagship. Longitudinal, multi-module architecture across the CKD→ESKD continuum, with agentic module orchestration. Includes a timeline viewer.
app · live
04
Complex Multi-Disease Framework (CMDF)
17 module-agents plus the fluid/volume triage tool — DIAL and CKD pathways. Where the confidence axes (data completeness, signal strength, context adequacy) live.
app · live v2
07
Clinical Recommendation Readiness Framework (CRRF)
The generalized engine: nine gates, six tiers of earned authority, and the move that makes it matter — it applies to a patient or a whole population. White paper, leadership brief, red-teamed novelty position, and a 46-source verified evidence base.
docs + flow sheet v0.3
06
BRIDGE — Hospitalization Reconstruction
Reads a chart backward: Baseline · Run-up · In-hospital evolution · Definitive anchors · Gestalt at discharge · Evaluate. Produces a diagnostic trajectory, a 0–100 certainty, an independent coherence verdict, and the signature output — the ranked known-unknowns list: what a careful clinician would still ask, surfaced because the chart is silent. Six worked cases, each externally red-teamed; plus a scoring harness.
specs + 6 cases v0.1

B · The ground truth

What makes the claims scoreable rather than assertable. This is the quietly expensive part of the portfolio and the part that makes the rest defensible.
Patients/ — the canonical cohort
18 synthetic patients, source of truth for everything else: master JSON, timeline, labs, next-best-action snapshots, dated ADT events. App data mirrors this, never the reverse. Ships with an interactive explorer.
canonical data + explorer app
05
Synthetic hospitalization packets
6 packets, each deliberately built so the ED’s first impression drifts to a different discharge diagnosis — and each shipped with a machine-readable answer key. These are BRIDGE’s evaluation inputs.
data + answer keys
08
Synthetic lab & clinical-document corpus
54 scanned LabCorp-style reports (3 per patient), 24 imaging/pathology reports, 3 multi-patient fax combos, named ordering and interpreting physicians. Dual-purpose: OCR/de-identification ground truth (1,682 PHI tokens, with injected de-ID failure modes) and extraction ground truth.
data + generators
09
Aberrant / counterfactual lab corpus
20 reports with 40 deliberately implausible values and a ground-truth list of every one. Explicitly non-canonical — it exists to benchmark whether a system notices that a number cannot be true. This is the adversarial test set.
data + ground truth non-canonical

C · The checks

Two different auditors: does it work clinically, and does it survive external governance?
03
Clinical validation
Validation report(s) supporting the frameworks — the clinical-accuracy check.
docs
10
CHAI governance overlay
Alignment of the portfolio to the CHAI Responsible-AI T&E frameworks and the Risk Categorization Tool (v3) — including a worked risk categorization of the Care Orchestrator. Explicitly not a clinical framework; it’s the layer that translates the epistemics into language a health system’s governance committee already accepts.
docs + fillable template

D · Where it’s going

The only folder that is design rather than artifact — and the one that determines whether the portfolio becomes infrastructure or stays a set of demos.
11
Knowledge Graph Engineering
Two coupled graphs (clinical world-model + epistemic overlay), a domain-tailored schema, a four-stage build, evaluation against the gold sets that already exist, and a construction pattern — fan-out extraction → verifier-on-edge → resolution barrier → synthesize — which is the ingestion-layer sibling of the CRRF gates: a fact must earn the right to be asserted.
design note · v0.1
11+
Decision theory as the reasoning layer ★ new
Today’s addition. Kochenderfer, Wheeler & Wray’s Algorithms for Decision Making (MIT, 2022) supplies the third layer: the formal machinery the portfolio has been inventing in prose. See the Rosetta Stone below.
strategy note · 2026-07-24
04

The engine — nine gates, six tiers

CRRF v0.3
This is the mechanism everything else references. The critical design choice is in the right-hand column: failing a gate doesn’t stop the system, it lowers the ceiling on what the system is allowed to say. That’s what prevents the usual failure mode where a blocked system either crashes or confabulates.
1
Understand the question
What clinical decision is being asked — and for whom: one patient, or a defined population? Declared explicitly. No undirected retrieval.
Unclear question or unit → no authority
2
Know what data is needed
The question defines the data model — what a competent clinician would need, sorted into required, material, contextual, or irrelevant.
3
Check what we actually have
Provenance, completeness, currency and gaps versus what’s expected — of this record, or of the dataset’s representativeness and selection bias.
Data unfit for this question → authority capped
4
Interpret before using
Raw values become evidence only through interpretation. A potassium from a hemolyzed sample is not a potassium; a code is not a phenotype until validated at scale.
5
Judge sufficiency
Sufficient, conditionally sufficient, insufficient, or contradictory — judged against the question, not by model confidence.
Insufficient or contradictory → say so, out loud
6
See the whole picture
The data must cohere — into a coherent patient state, or a stable, cross-stratum-consistent cohort.
Incoherent picture → authority capped
7
Earn the authority level
Sufficiency plus coherence jointly set which of six tiers is earned. Authority is earned, never presumed.
8
Show the work
The answer ships as a package: support, counter-evidence, assumptions, provenance, confidence, blind spots, and the conditions that would reverse it.
9
Learn under governance
Observation, interpretation, action and outcome recorded separately; durable knowledge and versioned cohort definitions governed and auditable.

Six tiers of earned authority

T5 · ActionableFull authority to recommend or conclude — all nine gates passed cleanly.
T4 · AdvisoryMay recommend or report, with full package, for review.
T3 · ConsiderationMay offer possibilities with explicit uncertainty.
T2 · ObservationMay surface interpreted findings without judgment.
T1 · Data requestMay only identify what is missing and ask for it.
T0 · SilentStates plainly what it cannot determine — and what would change that.
The honest exit

At any gate, “I cannot determine this — and here is what would change that” is a complete, correct answer. For a population, its analog — “this dataset cannot answer this question” — is the guardrail against confidently-wrong analysis.

05

The Rosetta Stone

added 2026-07-24
This is the most consequential thing in the newest work. Every signature idea in the portfolio turns out to have a precise, named counterpart in decision theory. That does three things at once: it gives the project a rigorous vocabulary, a normative reference model to be measured against, and — for the first time — a real evaluation story.
Portfolio idea (prose today)Formal constructWhere
“Earn the right to make a recommendation” Maximum expected utility with an explicit abstain action, plus a value-of-information gate Ch. 6 — Simple Decisions
“I cannot determine this” as first-class output Decision-theoretic abstention — expected utility of acting falls below threshold, or VOI says gather more first Ch. 6.6 — Value of Information
Calibrated certainty (BRIDGE 0–100) Posterior probability — genuinely calibratable and scoreable (Brier, log loss) Ch. 2–4
The nine CRRF gates The structure of a decision network: chance nodes (state), decision nodes (NBA), utility nodes (benefit/harm) Ch. 6.5
Confidence axes — data completeness, signal strength, context adequacy State uncertainty vs. model uncertainty; expected information gain Parts III & IV
BRIDGE diagnostic trajectory (ED impression → discharge dx) Belief-state updating in a POMDP — the working diagnosis is a belief that updates as evidence arrives Ch. 19 — Beliefs
BRIDGE “silent-chart” known-unknowns The unobserved variables with the highest value of information / largest entropy reduction — so the list should be ranked, not just enumerated Ch. 6.6; Part IV
Fluid-framework NBAs — RN data-gathering vs. clinician review Information-gathering actions (observe) vs. interventions (act), ranked by VOI Ch. 6.6; Part II
CKD→ESKD longitudinal management A sequential decision problem — an MDP/POMDP over disease states Parts II & IV
Aberrant-labs red-teaming + multi-model QA Policy validation — robustness, rare-event simulation, adversarial analysis Ch. 14
The one-liner this buys you: “AI that knows when not to answer” is, precisely, value-of-information-gated maximum expected utility with an abstain action — a defensible, literature-grounded formulation for a governance or leadership audience.
06

Where this is genuinely hard, and what’s next

Keeping faith with the project’s own ethos means being explicit about what isn’t known and what could go wrong. These are the three real tensions, and the four open moves.

The honest tensions

Where not to over-fit the new decision-theory framing
Quantification
Full Bayes nets, decision networks and POMDPs need numbers — conditional probability tables, transition, observation and utility models. Medicine rarely hands these over, and 18 synthetic patients cannot learn them. Near-term use is conceptual scaffolding plus selective formalization, not a rebuild.
Optimization vs.
inspectability
Decision theory says “take the highest expected utility action.” The portfolio deliberately privileges transparency and abstention — a single utility number can be less auditable than a visible chain of gates. The resolution: decision theory is the reference model; the gates are the auditable runtime that approximates it.
Calibration
Once certainties are posterior probabilities, they can be scored. That’s a gift and an obligation — an uncalibrated “0–100 certainty” stops being a vibe and becomes a defect you can name.

The open moves

Carried forward from the instance log — nothing here is started yet
1
Formalize one decision, end to end. Pick a single decision — the fluid-overload NBA, or one BRIDGE known-unknown — and build it as a decision network with an actual VOI calculation and clearly-labeled probabilities. The reasoning-layer analogue of “prototype on one packet.”
2
Prototype the ingestion workflow on one hospital packet. Fan-out extraction → verifier-on-edge → resolution barrier → synthesize, run against a packet whose answer key already exists.
3
Rank the known-unknowns. BRIDGE’s silent-chart list currently enumerates. VOI says it should order — the question worth asking first is the one that most reduces decision-relevant uncertainty.
4
The missing folder. Lab Summary Agent — Sequential Axis Model (NephAssist v1.4) is described in the portfolio brief but has no folder yet.