VibeRounds · Module Governance Report

Right-Sized Rigor: Where the Stack Earns Its Keep, and Where It Doesn't

A complete view of module maturity, capability-fit, the heatmap behind it, and the scope for improvement it implies. The guiding idea throughout: build each module simply enough to be genuinely usable for learning today, while keeping a clear, honest scope for advancing the ones that are still low-maturity — never dressing up a weak module as finished, and letting the ones that are already decent for learning keep improving in accuracy and reliability as the underlying technology advances.

In plain words

What this page is: VibeRounds has 57 learning modules. Not all of them can be trusted to run in a single AI pass — some need a real student answer first, some need real patient facts, and some quietly make up numbers if you let them run unsupervised. This report sorts every module into what it can honestly claim to do, how mature it is, and what it would actually take to fix the weak ones.

How to read it:

Section 1 explains the five categories a module can fall into — real analysis, a shortcut that skips learning, a guess that needs a human check, a fake exercise, and one that needs real facts fed in step by step.
Section 2 sorts all 57 modules into those five categories.
Section 3 zooms into the "real analysis" group and flags which of those are still fragile — sound reasoning, but specific facts or numbers that can't be trusted without a real source.
Section 4 is a single scorecard: every module, rated 1–5 for how much you can trust its output as-is.
Section 5 lists what it would take to fix each weak spot, cheapest fixes first.
Section 6 makes the closing case: keep things simple wherever possible, but never at the cost of quietly making things up.

Bottom line: most modules (28 of 57) are already sound as-is. The rest either need a small process fix (capture a real answer before generating the comparison), a data-grounding fix (pull from real sources instead of memory), or, for a handful, real infrastructure that doesn't exist yet. One category — genuinely noticing something unusual in a patient — can never be fully handed to AI, no matter how much is built.

On this page

The five-tag use-classification framework
Full module classification, all five tags
Capability-fit within Tag A
Module maturity heatmap, full table
Scope for improvement, grouped by build cost
Why usability still wins the tie

The organizing principle

More architecture is not automatically better, and less architecture is not automatically safer. The right fix is sized to why a module is unreliable — a module that fabricates registry statistics needs a real data source, full stop; a module that's merely slow to word a guideline recommendation exactly right doesn't need a neuro-symbolic reasoner, it needs the guideline text in context. Matching the fix to the failure is the whole discipline. Skipping a needed fix in the name of simplicity reintroduces exactly the silent-fabrication risk this whole framework exists to catch.

1 · The five-tag use-classification framework

Whether a module is safe to run single-pass isn't a yes/no question — different modules fail in very different ways. Some give a correct answer but the learner's brain never engages. Some produce output that looks like the exercise ran but the module's core mechanism structurally requires a real person and can't function without one. Some need real facts supplied incrementally, step by step, or the whole downstream output is hollow. Five tags describe what claim the output is allowed to make about itself.

Overall disclaimer, applies to every module, every mode: all VibeRounds modules — Socratic, single-pass, or mixed — are learning and reflection tools only. No output from any module is a clinical decision, a diagnosis, a care recommendation, or a substitute for a real patient's or advocate's actual input. Nothing produced should enter a real patient record, guide real treatment, or replace a supervising clinician, without independent human verification at every step — even outputs that look complete, confident, or clinically fluent.

Tag	Meaning	What you get	What you don't get
A — Genuine AI Analysis	The task itself is analytical/synthetic by nature; single-pass is the correct way to do it	Real analytic value — this is what the module is for	Nothing missing — this is the intended mode
B — AI Reasoning Substitute	AI can derive a correct answer from case data, but the module's purpose was for the learner to reason it out	A correct answer, fast	The learner's own reasoning rep — no skill built, no error caught, no habit formed
C — AI Insight Approximation	AI produces a plausible, labeled inference about something with a real, person-specific true answer AI cannot verify	A reasonable starting hypothesis	Any guarantee it matches the real learner's mind, patient's situation, or team's knowledge — must be a prompt for reflection, not a finding
D — Illusion of Mechanism	The module needs two genuinely distinct real signals to compare; single-pass has nothing real on either side	Text with the shape/formatting of the exercise's output	The exercise itself — worse than B, a non-functional simulation of having done it
E — Sequential Real-Input Dependency	Real facts must be supplied incrementally because they don't exist anywhere else, or carry real care consequences if wrong	Nothing usable — any output is built on invented data	The entire task — fabrication of content that could be mistaken for a real clinical record

2 · Full module classification

Tag A — Genuine AI Analysis (28 modules; single-pass is correct)

#	Module
5	Real-Time Case Review & Data Audit
6	Registry-Level Analytics
7	Longitudinal & Cross-Case Learning
8	Socratic-Mode Design Specification
10	Journal & Article Reading
11	Patient Education Query Intelligence
19	Community & Social Medicine Insights
21	Evidence Frontier Search
22	Nested Analysis
23	Counterfactual Analysis (label output as hypothetical)
24	Heuristic Analysis
25	Thematic Analysis
27	Time-Series & Velocity Analyzer
29	The Iatrogenic Domino Effect
34	High-Value Care (HVC) Auditor
38	Poly-Crisis & Cascading Failure Simulator (label output as hypothetical)
39	"Global Knowledge Network" Diagnostic Matrix
40	Operational & Throughput Strategist
41	Clinical Workflow Implementation Science
43	Health Economics & Value-Based Care Alignment
45	Shadow Module 44 — Genetics Adversarial Counterpart
46	Evidence-Based Medicine Insights
47	Shadow Module — EBM Adversarial Counterpart
48	Treatment Comparative Analysis & Prognosis Trajectory
51	Systems-Based Clinical Analysis
52	Clinical Pearls Distillation
53	Clinical Guideline Intelligence Navigator
2.4–2.7	SOAP note, completeness audit, sign-off, handover brief — once real facts are already captured

Tag B — AI Reasoning Substitute (17 modules; correct answer, learner's brain bypassed)

#	Module
1	Socratic Clinical Reasoning
4	Peer-Level Ward Round Preparation
12	Differential Diagnosis Deepdive
14	Resource-Constrained Clinical Reasoning
15	Illness Script Acquisition
16	Basic Science ↔ Clinical Integration
17	Semantic Qualifiers & Problem Representation
18	Causal vs. Probabilistic (Network) Reasoning
20	Naturalistic Decision Making
28	Diagnostic Time-Out
31	First-Principles Pathophysiology Mapping
37	Red Herring / Signal-to-Noise Drill
44	Clinical Genetics Reasoning
50	Diagnostic Reasoning Map
56	Hypothetico-Deductive Reasoning
57	Clinical Cognition Deep Dive
36	Bayesian Probability / Likelihood Ratio Engine — arithmetic step only, once a probability exists

Tag C — AI Insight Approximation (6; plausible but unverifiable — flag as hypothesis)

#	Module	What's being approximated
26	Bias Auditing	A bias pattern inferred from written reasoning — may not match the learner's real bias
30	"Diagnostic Anchor" Extractor	Same limitation — inferred, not confirmed
33	"Why Now?" (Precipitant) Hunter	A plausible precipitant pattern-matched to the timeline — may not be the true trigger
36	Bayesian Probability Engine — prior-probability step	Substitutes population prevalence for the learner's own contextual judgment
42	Clinical Pre-Mortem	Generic failure scenarios — real personal/team blind spots may be missed
49	FMEA Analysis	Literature-derived failure modes — institution-specific ones may be missing

Tag D — Illusion of Mechanism (3; output looks complete, exercise didn't happen)

#	Module	Why the mechanism can't function single-pass
32	Clinical Cognition Loop	Reflects on a real reasoning process that, in single-pass, never occurred
35	Epistemic Certainty Mapping & Calibration	Calibration requires a real stated confidence vs. real accuracy; AI has no real confidence to report
54	System 1 & System 2 Thinking Question Generator	Requires two genuinely different real cognitive responses; AI generating both means the comparison is empty

Tag E — Sequential Real-Input Dependency (5; single-pass fabricates content with no real basis)

#	Module	Step(s) affected
2	Patient-Advocate Case Documentation	Steps 2.1, 2.2, 2.3, 2.8 — symptom, exam, medication capture
3	Extended Patient-Advocate Monitoring	Data-capture / monitoring-entry steps
9	N-of-1 Case Research Protocol	Real patient data-entry steps
13	Medication Reconciliation & Polypharmacy Audit	Medication-list capture step
55	Patient Needs Assessment	Full module — real-world care-decision stakes even though some inference is technically possible

3 · Capability-fit within Tag A

Being Tag A means the task itself is analytical — but that's a claim about design, not about whether a plain LLM can deliver it without fabricating specifics. A second axis matters: is the module reasoning over knowledge/text it genuinely has, or structurally implying access to data it was never given? Tag A splits into three tiers.

Tier 1 — Solid: real text/knowledge reasoning, low fabrication risk

Modules 8, 11, 19, 22, 23, 24, 25, 29, 34, 38, 41, 45, 47, 51, 52, and 2 (steps 2.4–2.7). The output is reasoning/prose applied to information already on the page — a case narrative, a set of facts. Nothing to fabricate, because there's no implied external dataset being consulted.

Tier 2 — Fragile: the reasoning shape is legitimate, precise specifics are where it breaks

#	Module	Where it's fragile
10	Journal & Article Reading	Fine if pasted in; if "recalled" from training, expect plausible-but-unverifiable details and stale coverage past the training cutoff
46	Evidence-Based Medicine Insights	Concepts usually right; exact recommendation wording is where LLMs drift subtly wrong
53	Clinical Guideline Intelligence Navigator	Guideline version and precise thresholds easy to get confidently wrong or out of date
43	Health Economics & Value-Based Care Alignment	Tradeoff reasoning fine; specific cost figures are illustrative, not sourced
48	Treatment Comparative Analysis & Prognosis Trajectory	Comparative reasoning fine; survival percentages not to be trusted as sourced facts
39	"Global Knowledge Network" Diagnostic Matrix	Name overclaims — it's a differential from trained knowledge, not a live network query
40	Operational & Throughput Strategist	Strategic reasoning fine; specific throughput/efficiency numbers invented unless supplied

Tier 3 — High-risk overclaim: the premise requires a capability a plain LLM structurally doesn't have

#	Module	The overclaim
6	Registry-Level Analytics	"Analytics" implies aggregation across a real case registry. Without an actual dataset in context or a real query tool, any numbers, prevalences, or trends are fabricated statistics dressed up as findings. The clearest example on the list.
7	Longitudinal & Cross-Case Learning	"Cross-case" implies real multiple cases were compared. Without them supplied, it's a single generic narrative wearing a "learned-across-cases" label.
21	Evidence Frontier Search	"Frontier" implies current literature awareness. Without a real search tool invoked, this is either stale or invented — dangerous because it reads exactly like a genuine literature review.
27	Time-Series & Velocity Analyzer	Different failure mode: not knowledge currency, but arithmetic reliability. LLMs are error-prone at precise rate/slope calculations token-by-token. Fine for a qualitative "trending up/down" read; not fine for exact rates without a calculator or code step.

Practical rule: treat Tier 2 and Tier 3 as needing one of — (a) real data or source text supplied in-context, (b) an actual tool/search call, or (c) explicit "illustrative, not sourced" labeling. Without one of these, a module in those tiers is Tag A in name only and behaves like an unlabeled Tag C: a plausible-sounding output standing in for a real finding.

4 · Module maturity heatmap, full table

Both axes above combine into one score per module — how much an unmodified single-pass run can be trusted.

5 — High maturity: reason over given/known material, trustworthy as-is

4 — Reliable output, but bypasses the learner's own reasoning

3 — Reasoning sound, precise specifics need verification

2 — Plausible inference only, must be labeled a hypothesis

1 — Structurally can't deliver unmodified

#	Module	Use-Tag	Maturity	Why
Level 5 — High Maturity
5	Real-Time Case Review & Data Audit	A	5	No external dataset implied
8	Socratic-Mode Design Specification	A	5	Design reasoning, no data claim
11	Patient Education Query Intelligence	A	5	Applied-knowledge prose
19	Community & Social Medicine Insights	A	5	Advisory synthesis, no data claim
22	Nested Analysis	A	5	Reasoning over given case
23	Counterfactual Analysis	A	5	Label as hypothetical, otherwise sound
24	Heuristic Analysis	A	5	Reasoning over given case
25	Thematic Analysis	A	5	Reasoning over given text
29	The Iatrogenic Domino Effect	A	5	Reasoning over given case
2.4–2.7	SOAP note / audit / sign-off / handover	A	5	Pure formatting/synthesis once facts captured
34	High-Value Care (HVC) Auditor	A	5	Applies known criteria to given text
38	Poly-Crisis & Cascading Failure Simulator	A	5	Label as hypothetical, otherwise sound
41	Clinical Workflow Implementation Science	A	5	Advisory prose, no data claim
45	Genetics Adversarial Counterpart	A	5	Reasoning over given case
47	EBM Adversarial Counterpart	A	5	Reasoning over given case
51	Systems-Based Clinical Analysis	A	5	Reasoning over given case
52	Clinical Pearls Distillation	A	5	Synthesis of given material
Level 4 — Reliable Output, Bypassed Learning
1	Socratic Clinical Reasoning	B	4	Correct answer, skips the reasoning rep
4	Peer-Level Ward Round Preparation	B	4	Same pattern
12	Differential Diagnosis Deepdive	B	4	Same pattern
14	Resource-Constrained Clinical Reasoning	B	4	Same pattern
15	Illness Script Acquisition	B	4	Same pattern
16	Basic Science ↔ Clinical Integration	B	4	Same pattern
17	Semantic Qualifiers & Problem Representation	B	4	Same pattern
18	Causal vs. Probabilistic (Network) Reasoning	B	4	Same pattern
20	Naturalistic Decision Making	B	4	Same pattern
28	Diagnostic Time-Out	B	4	Same pattern
31	First-Principles Pathophysiology Mapping	B	4	Same pattern
36 (LR math step)	Bayesian Probability / Likelihood Ratio Engine	B	4	Arithmetic once a probability exists
37	Red Herring / Signal-to-Noise Drill	B	4	Same pattern
44	Clinical Genetics Reasoning	B	4	Same pattern
50	Diagnostic Reasoning Map	B	4	Same pattern
56	Hypothetico-Deductive Reasoning	B	4	Same pattern
57	Clinical Cognition Deep Dive	B	4	Same pattern
Level 3 — Sound Reasoning, Fragile Specifics
10	Journal & Article Reading	A / Fragile	3	Fine if pasted in; risky if "recalled"
39	"Global Knowledge Network" Diagnostic Matrix	A / Fragile	3	Name overclaims — trained-knowledge differential
40	Operational & Throughput Strategist	A / Fragile	3	Strategy sound, specific numbers invented
43	Health Economics & Value-Based Care Alignment	A / Fragile	3	Cost figures illustrative, not sourced
46	Evidence-Based Medicine Insights	A / Fragile	3	Exact recommendation wording drifts
48	Treatment Comparative Analysis & Prognosis Trajectory	A / Fragile	3	Survival percentages not to be trusted
53	Clinical Guideline Intelligence Navigator	A / Fragile	3	Guideline version/thresholds drift
Level 2 — Plausible Inference Only
26	Bias Auditing	C	2	Inferred bias, may not match learner's real bias
30	"Diagnostic Anchor" Extractor	C	2	Inferred, not confirmed
33	"Why Now?" (Precipitant) Hunter	C	2	Pattern-matched, may not be the true trigger
36 (prior-prob step)	Bayesian Probability Engine	C	2	Population prevalence ≠ contextual judgment
42	Clinical Pre-Mortem	C	2	Generic scenarios, real blind spots may be missed
49	FMEA Analysis	C	2	Literature-derived, institution-specific gaps missed
Level 1 — Structurally Can't Deliver Unmodified
6	Registry-Level Analytics	A / High-risk overclaim	1	No real registry queried — numbers fabricated
7	Longitudinal & Cross-Case Learning	A / High-risk overclaim	1	No real multi-case comparison happened
21	Evidence Frontier Search	A / High-risk overclaim	1	No real search — stale or invented, reads like a lit review
27	Time-Series & Velocity Analyzer	A / High-risk overclaim	1	Precise rate math unreliable without a calculator/code step
32	Clinical Cognition Loop	D	1	Nothing real to reflect on in single-pass
35	Epistemic Certainty Mapping & Calibration	D	1	No real confidence to compare against
54	System 1 & System 2 Question Generator	D	1	Both "sides" AI-generated — comparison is empty
2.1–2.3, 2.8	Patient-Advocate Case Documentation (data steps)	E	1	Real symptom/exam/med facts have no substitute source
3	Extended Patient-Advocate Monitoring	E	1	Each entry is a real day's real observation
9	N-of-1 Case Research Protocol	E	1	Real patient data-entry steps
13 (med-list step)	Medication Reconciliation & Polypharmacy Audit	E	1	Real medication list has no substitute source
55	Patient Needs Assessment	E / special case	1	Plausible guess risks being mistaken for real patient wishes

5 · Scope for improvement, grouped by build cost

Grouped by what it actually costs to build — because the usability argument only holds if the cheap fixes get shipped first and the expensive ones are reserved for where they're truly needed.

SHIP FIRST — NO HEAVY INFRA Elicitation-first fixes

Not an architecture-stack problem at all — these need a real signal captured live, before generation, the same mandatory-descriptor mechanism proposed for a missed red-flag descriptor case. Cheapest, highest-integrity fix on the whole list.

Modules 32, 35, 54 (Tag D) — capture the learner's real reasoning trace, stated confidence, or fast System-1 answer first, as a required input, before the AI generates the comparison side. This converts D into a legitimate B/A once the real signal exists.
Modules 2 (data steps), 3, 9, 13 (med-list), 55 (Tag E) — run genuinely Socratic, one real answer at a time; no single-pass alternative should exist for these steps. If only the output shape is wanted, offer a clearly-labeled placeholder-data template demo instead — never presented as real.

LIGHT GROUNDING Pathway / knowledge-graph / RAG on already-sound modules

Reasoning shape is already sound — this is pure specifics-drift, and it's exactly what pathway grounding, a knowledge graph, or RAG is built to fix: converting a hedge into a checkable, sourced claim.

Modules 10, 46, 53 — RAG against the real article/guideline text instead of model memory.
Modules 43, 48 — knowledge-graph entries for specific cost/prognosis figures, or explicit "illustrative" labeling where no real source exists.
Module 39 — rescope the naming honestly, or back the "network" claim with a real knowledge graph.
Module 36 (prior-probability step) — a prevalence knowledge graph makes the number sourced, not just plausible.
Module 40 — needs real operational data supplied in-context; closer to the next tier's pattern than a pure grounding-layer fix.

REAL INFRASTRUCTURE Where the claim genuinely requires a data source

These modules imply capabilities — a real registry, real cross-case history, live literature search — that no amount of clever prompting substitutes for. Worth building because the payoff is a full tag upgrade, not because it's the cheapest option.

Module 6 — an actual structured registry, queried and cited, with the coverage flag surfaced rather than silently smoothed over.
Module 7 — a real case store with GraphRAG-style provenance on every cross-case claim.
Module 21 — RAG against a live literature index, not recall.
Module 27 — not a knowledge-stack fix — a code-execution step for the actual arithmetic, cleanest fix on the list.
Module 13, downstream of its med-list elicitation — a real RxNorm-anchored interaction check, exactly the bounded, well-structured use case neuro-symbolic validation is best suited for.

LABELING DISCIPLINE — DON'T OVER-BUILD Modules 26, 30, 33, 42, 49

Part of Tag C hits a hard ceiling directly: a bias, a diagnostic anchor, or a "why now" trigger is learner- or patient-specific information no reference material holds. A knowledge graph can make the label checkable; it can never confirm the label is this learner's real bias. Building heavier architecture here mistakes a category boundary for a reliability gap — don't. The correct, cheap fix is consistent "AI's best guess" labeling, every time, no exceptions. (42 and 49's institution-specific pieces are a partial exception — a real incident-report or failure-mode database is genuinely buildable there.)

PERMANENTLY OUT OF SCOPE The noticing gap

"Knowledge management and reasoning tools can give you knowledge and reasoning — they cannot tell you what knowledge or reasoning you need."

No layer above — pathway, graph, RAG, GraphRAG, neuro-symbolic, or a well-built elicitation checklist — closes this. Every one of these tools operates on reasoning, text, patterns, or logical consistency; none of them operate on the patient. A checklist only catches the red flags someone thought to encode in advance; a presentation nobody anticipated still gets no question asked about it. This isn't a backlog item. It's the reason VibeRounds stays scoped to self-audit and learning, not decision support, regardless of how much of the above eventually gets built.

6 · Why usability still wins the tie

Something small enough to actually use beats something comprehensive that nobody opens — a ten-minute loop wins over a heavier one people abandon. Nothing above contradicts that. The "ship first" and "labeling discipline" tiers are both usability-first moves: they cost little, ship fast, and are exactly where extra architecture would have been wasted effort anyway, since each additional layer's marginal value shrinks as the base model itself keeps improving. The one place usability doesn't get the final word is where a module is currently producing content that looks real but was fabricated — a registry statistic that was never queried, a cross-case pattern that never happened. There, the simple fix and the honest fix are the same fix: don't run it unmodified, full stop, until the real data source exists. Usability is a tiebreaker for how to build well. It's never a justification for shipping something that quietly manufactures false confidence.