Lesson 9 ran the full engine, once, on one case. That was the last lesson about getting to a diagnosis. Everything from here forward assumes you can already do that competently, and asks a different question: what does clinical reasoning look like once the diagnosis is the least interesting part of the problem — because the case is a multi-step management plan, an ambiguous outcome, or a decision that has to be defended to someone who wasn't in the room?
The hinge for that shift is meta-cognition — reasoning about your own reasoning, not about the patient. Modules 18 and 20, which Lesson 3 introduced narrowly as a way to close out a differential, turn out to generalize into something much bigger: a habit of treating your own trace as data, worth auditing on its own terms, at any point in a case, not just at the end of a differential.
Most clinical training ends at "did you get the diagnosis right." This lesson is about the layer above that — did you know how confident to be, did you know which part of your reasoning to distrust, and could you say so before someone else pointed it out.
This is the shift from evaluating a model's output to evaluating its self-report about its own output — from "was the answer right" to "was the confidence score, and the stated reasoning for it, actually trustworthy."
Two Kinds of Being Wrong
A reasoner can fail in two different ways that get flattened into "wrong" if you only grade the final answer. The first is a wrong diagnosis reached through sound reasoning — bad luck, or a genuinely atypical case, and largely uninteresting from a meta-cognitive standpoint. The second is a correct diagnosis reached through reasoning that would have failed on a slightly different case — right for the wrong reasons, and the far more dangerous failure mode, because nothing about the outcome flags it. Module 18's self-critique pass exists specifically to catch the second kind, since the first kind is usually visible from the outcome alone.
"Right for the wrong reasons" is invisible in a chart review that only checks outcomes — it only shows up when someone asks you to defend the reasoning itself, which is exactly what this habit trains you to do to yourself, first.
This is the difference between a model that happens to pass an eval and one that would generalize to a shifted distribution — the eval score alone can't tell them apart, but interrogating the reasoning chain that produced the answer often can.
Confidence Isn't One Number: Module 20, Generalized
Lesson 3 used Module 20 narrowly, as a confidence rating on one weak inferential link. Beyond a single diagnosis, the same habit has to be applied across several independent axes at once, because a plan can be reasoned soundly on one axis and shakily on another without either failure showing up in a single blended score. Three axes recur across most non-diagnostic reasoning: confidence in the underlying facts, confidence in the plan's logic given those facts, and confidence that the plan is still right if the patient's circumstances (adherence, access, preference) diverge from the assumed default.
Module 56's cross-case pattern bank does real work here too, in a different direction than in Lesson 3 — instead of only storing near-miss diagnostic pairs, it now stores cases where a plan looked sound at commitment time and turned out, on later review, to have a specific meta-cognitive gap. That bank becomes the raw material for training the self-critique habit on patterns that recur across cases, not just within one.
"I'm confident in the diagnosis but not confident this patient will actually be able to follow the plan" is a completely different statement from a single blended confidence score — and it's the one that actually changes what you do next.
This is multi-dimensional calibration instead of a single scalar confidence score — closer to reporting separate error bars per subsystem than one aggregate accuracy number that hides which subsystem is actually unreliable.
Module 57 and Module 31, Rejoined
Lesson 1 previewed two modules without being able to use them yet: Module 31's single-number confidence habit, and Module 57 as "the orchestrator we haven't written yet." This lesson is where both come due. Module 31's habit — commit to a confidence number before the outcome is known — is what actually makes Module 20's multi-axis rating trustworthy rather than performative, because a rating made after the fact is worth almost nothing. Module 57, meanwhile, is the meta-cognitive layer applied to the whole course so far: given everything Lessons 1 through 9 built, which module should run, in what order, on a case that doesn't announce which lesson it belongs to — the same sequencing problem the Master Protocol solved for diagnosis, now extended to reasoning that continues past it.
That second half of the prompt is deliberate. The most common meta-cognitive failure isn't picking the wrong tool — it's reaching for the familiar tool out of habit when a different one is actually indicated, which is itself a bias worth naming before it's acted on.
Every clinician has a favorite move — a test they order too readily, a diagnosis they reach for too fast. Naming which module you'd default to "out of habit" is a direct way of surfacing that same tendency.
This is a routing decision made explicit and interrogated, rather than left to whichever handler happens to fire first — the meta-cognitive equivalent of asking a router not just what it chose, but what it almost chose instead.
Homework for Lesson 10
- Take a case you've reasoned through recently — ideally the one you've carried since Lesson 2. Run Module 18's generalized self-critique on your management plan, not just your diagnosis: name the single assumption the plan depends on most, and state whether the plan survives that assumption being false.
- Rate your confidence on all three Module 20 axes separately — facts, logic, and circumstance-robustness — rather than giving one blended number. Where do the three numbers diverge most?
- Run Module 57's routing question on the same case: which module from Lessons 1–9 would most reduce your uncertainty right now, and which one were you tempted to reach for instead, out of habit? Bring both answers forward — Lesson 11 will ask what changes when this same judgment has to be made across a team rather than by one reasoner alone.
This lesson draws on the generalized form of Module 18 — Meta-Cognitive Self-Critique and Module 20 — Confidence Interval Self-Rating, first introduced narrowly in Lesson 3, together with Module 31's calibration habit and Module 57's capstone-routing framing from Lesson 1, and Module 56's cross-case pattern bank, all from the VibeRounds Prompt Directory. If you're coming from the evidence side, the companion Evidence-Based Medicine for Techies course treats calibration from the statistical side; this lesson treats it from the reasoner's side. Neither course is a clinical decision tool; see the VibeRounds disclosure statement for full terms.