Teaching an AI to Argue With You: What Vibe Rounds Gets Right About Clinical Reasoning

A field note on the Cognitive Analytics Series

There's a version of "AI in medicine" that everyone has already imagined: you feed it a case, it spits out a diagnosis, and you either trust it or you don't. Vibe Rounds, the clinical-reasoning framework at the center of this series, is deliberately not that. It's built on a much stranger and more useful premise — that the most valuable thing a language model can do with a patient case isn't answer it, but argue with the person answering it.

That premise turns out to have a lot of moving parts, and the series traces them honestly, in the order they actually got built: a working demo first, then the machinery to keep it honest, then the hard question of when that machinery is worth the trouble, and finally — the part most frameworks skip — a version small enough that someone might actually use it on a Tuesday.

Two modes, one job: slow the thinking down

The foundation is a dual-mode architecture with two jobs that pull in opposite directions on purpose.

Promption takes a messy clinical narrative — vitals, labs, a chief complaint, half-formed history — and scaffolds it into something structured: a clean problem representation, a differential built from the actual findings, an illness script instead of a vibe. It's the mode that organizes.

Provocation does the opposite. Once a hypothesis exists, it goes looking for reasons to doubt it — anchoring bias, premature closure, the one lab value nobody's explaining. It's the mode that argues.

The series makes its case with a single worked example: a 48-year-old woman in DKA, tachypneic, severely acidotic, with ground-glass opacities on CT. Promption mode builds the obvious story fast — infection triggering DKA. Provocation mode then does something a single confident answer wouldn't: it flags that her HbA1c is only 6.8%, which is an odd number to see attached to a pH of 6.77. Well-controlled diabetics don't usually crash that hard. That one flagged inconsistency opens the door to alternatives nobody had raised yet — metformin-associated lactic acidosis, ARDS as a consequence of the acidosis rather than its cause. Nothing here diagnoses the patient. It just makes sure the easy answer has to survive an argument before anyone acts on it.

The problem with fluent arguments: they can be fluently wrong

Here's the catch the series doesn't try to hide: both modes are, by construction, only as good as the model's internal coherence. A hallucinated anchor gets extracted just as smoothly as a real one. Fluency isn't the same as being right, and a stress-test built entirely out of fluency can manufacture doubt about a correct answer just as easily as it catches a wrong one.

So the second piece of the series does something unglamorous but necessary: it anchors the two cognitive modes to actual structure outside the model. A knowledge graph for chronic kidney disease turns a hedge like "consistent with CKD" into a checkable criterion — eGFR under 60 for three months or more, sitting there as a fact you can verify instead of a phrase you have to trust. An institutional pathway for acute pancreatitis does something a graph structurally can't: it encodes when, not just what — the 48-hour organ-failure threshold, fluid resuscitation titrated hour by hour. And a layer of appraised evidence answers a third question neither of the first two touches: how strong is the justification, really, and under what conditions does it hold?

Each layer earns its keep by catching a different failure. And each layer, honestly, introduces a new one. A correct fact from the graph can still get misapplied to a case that only looks like the textbook entry. A pathway-derived instruction carries the borrowed authority of "per protocol," which makes a wrong instruction more dangerous, not less, because people trust it faster. And evidence-appraisal language — GRADE tiers, effect sizes, confidence intervals — is arguably the most convincing hallucination surface in the entire stack, precisely because it sounds the most rigorous. Precision without real retrieval behind it is just fluency wearing a lab coat.

Not every case deserves the full stack

The natural next question — should every case get the graph, the pathway, and the evidence layer? — gets a refreshingly non-uniform answer.

For common, well-documented presentations, the base model is often already close to the truth, so stacking three independently-sourced layers on top mostly adds seams where they can quietly contradict each other, for not much accuracy gain. For the ultra-rare and genuinely novel, there's a different problem: there's no ground truth to anchor to in the first place, so a structure-first approach can force an unfamiliar case into the nearest available graph entry — a mistake that's harder to catch precisely because it looks verified.

The sweet spot sits in between — the rare-but-known presentation, where ground truth exists but the base model tends to over-fit the case to the nearest common script anyway. That's exactly where Provocation's anchor-hunting and the graph's differential field earn their cost. The series' sharpest line on this: the highest-leverage fix isn't adding or withholding structure uniformly, it's making "how much ground truth do I actually have for this case" a visible output in its own right. A system that fails loudly — an honest "no strong match" — beats one that fails quietly and confidently, every time.

There's also a category of uncertainty the series is refreshingly upfront about not solving: how sick the patient actually looked walking in the door, what was tried and failed at the bedside, whether the case narrative itself was a faithful capture of the real encounter. No graph, pathway, or evidence layer ever sees that gap. It's not a coverage problem a fourth layer could fix — it's a category the entire documented-representation approach can't reach, full stop.

The turn that makes it usable

Three pieces in, the series has been adding — a second cognitive mode, a graph, a pathway, an evidence layer, a rarity-weighting scheme. Every addition was justified on its own terms. And then the fourth piece does something most frameworks never bother to do: it turns around and asks whether anyone would actually use all of it.

The honest failure mode of a growing framework, as the series puts it, isn't that it's wrong — it's that it becomes something people admire from a distance and never open. So Part 4 proposes a ten-minute weekly loop: write the one-line problem representation from memory, name the single finding you're leaning on hardest, ask what would have to be true for that finding to point somewhere else, check exactly one fact against a source you trust, write one sentence on what you'd do differently. That's it. No graph, no pathway, no rarity tier — just Promption's first move, one Provocation question, and a single spot-check, small enough to survive an actual busy week.

The tell you've reached for too much tooling: you spent longer assembling ground truth than thinking about the patient. The tell you've reached for too little: you can't say, out loud, what would have changed your mind.

Why the shape of the series matters as much as the content

What makes this series worth reading in order isn't any single layer — it's the honesty of the arc. Tooling and claimed rigor climb for three parts straight, and then the fourth part deliberately climbs back down, not because the earlier layers were wrong, but because a framework only teaches you something if you actually open it. That's a rarer instinct than it should be. Most technical writeups about AI-augmented reasoning stop at "here's the architecture" and never get to "here's the version you'll actually run." Vibe Rounds does, and the ten-minute loop is arguably the most quietly ambitious idea in the whole series — not because it's clever, but because it's the one piece designed to survive contact with a real week.

Vibe Rounds CCOS Framework — for educational and clinical-reasoning practice only, not a clinical decision tool.