Guided Discovery, and Knowing Which of the 57 Prompts Modules to Trust

VibeRounds runs on a library of 57 modules — the full set lives at Prompts.html, everything from a Bayesian probability engine to a patient-advocate case log. Guided Discovery is the mode built around that library.

This post covers a governance pass over all 57 modules that asks, plainly, what each one can honestly claim to produce in a single AI pass, and where the honest answer is "not yet."

Read more here - https://avi33tbtt.github.io/articles/VibeRounds-Right-Sized-Rigor.html

Five tags, not a pass/fail

The instinct is to sort modules into "reliable" and "not reliable." That's too coarse. A module can fail in very different ways, and the fix depends entirely on which way it fails:

• A — Genuine AI Analysis. The task is analytical by nature. Single-pass is the correct way to run it — nothing is missing.
• B — AI Reasoning Substitute. The AI can derive a correct answer from the case data, but the module's actual purpose was for the learner to reason it out themselves. Run it single-pass and you get a fast answer with no learning rep behind it.
• C — AI Insight Approximation. A plausible, labeled guess about something that has a real, person-specific answer the AI cannot verify — a bias, a diagnostic anchor, a "why now" trigger. Useful as a prompt for reflection, never as a finding.
• D — Illusion of Mechanism. The module needs two genuinely distinct real signals to compare — a learner's actual reasoning trace against an AI-generated one, say. Single-pass has nothing real on either side. It produces text shaped like the exercise, without the exercise having happened.
• E — Real-Data Dependent. The module's whole output is hollow without real patient facts fed in step by step — a symptom, a medication, a day's observation. There's no substitute source for that data; it has to come from a real person, one true answer at a time.

None of these tags are a verdict on the idea behind a module. Modules 32, 35, and 54 — tagged D — aren't bad ideas; a System 1 vs. System 2 comparison or a confidence-calibration check is a genuinely good exercise. They're just structurally unable to happen in a single pass, because the "System 1 answer" and the "calibrated answer" both have to be AI-generated if nothing real is captured first, which means the comparison is empty on both sides.

The scorecard

Every module gets scored 1–5 on how much its output can be trusted as-is, and the numbers land better than you'd guess going in: 28 of the 57 modules are already sound. The other 29 split roughly three ways —

• A batch that's sound in its reasoning but fragile on specifics — guideline wording, cost figures, survival percentages that drift because they're coming from model memory rather than a source.
• A batch that's a plausible-inference tool by nature (Tag C) and always will be, no matter how much infrastructure gets built around it.
• A small, sharper-edged batch that structurally can't deliver anything real unmodified — Tag D modules with nothing to compare, Tag E modules with no real patient data behind them, and a couple of Tag A modules making claims (a queried registry, a live literature search) that a single AI pass simply cannot back up without fabricating the numbers underneath.

Fixing the right thing, not the expensive thing

The organizing principle behind the whole report is worth stating plainly: more architecture is not automatically better, and less architecture is not automatically safer. The fix has to match why a module is unreliable.

That plays out as four tiers, cheapest first:

1. Ship-first, no heavy infrastructure. The Tag D and Tag E modules don't need a knowledge graph — they need a real signal captured live, before generation. Get the learner's actual reasoning trace or stated confidence first, as a required input, and a D module becomes a legitimate B or A. Run the patient-advocate and medication-reconciliation modules genuinely Socratic, one real answer at a time. This is the cheapest fix on the whole list, and it's also the most important one, because it's a straightforward extension of the same mandatory-descriptor discipline that already exists for catching missed red flags elsewhere in the system.
2. Light grounding. For modules whose reasoning is already sound and only the specifics drift — guideline text, cost figures, prevalence numbers — RAG against the real source, or a knowledge graph, turns a hedge into a checkable claim.
3. Real infrastructure. A handful of modules imply capabilities — an actual patient registry, a live literature index, a real cross-case store — that no amount of clever prompting substitutes for. Worth building, but only because the payoff is a full tag upgrade, not because it's cheap.
4. Labeling discipline. For the Tag C modules — bias auditing, diagnostic-anchor extraction, the "why now" hunter — heavier architecture is the wrong instinct entirely. A knowledge graph can make a label checkable; it can never confirm the label is this learner's real bias. The correct fix here is cheap and consistent: label it "AI's best guess," every time, no exceptions.

The one gap nothing closes

There's a fifth category that doesn't get a tier, because no tier fixes it. Pathway grounding, knowledge graphs, RAG, GraphRAG, neuro-symbolic validation — every one of these 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.

That's not a backlog item — it's the reason VibeRounds stays scoped to self-audit and learning rather than decision support, no matter how much of the rest of this gets built out.

Why this belongs in Guided Discovery

Guided Discovery's whole premise is steering a learner toward the right module for the moment they're in. That steering is only honest if it also carries this map along with it — pointing someone toward Module 6's registry analytics or Module 54's System 1/System 2 comparison without flagging that both are currently Tag D/Tag A-overclaim would be worse than not building Guided Discovery at all. Usability is meant to be a tiebreaker for how to build well. It's never license to ship something that quietly manufactures confidence it hasn't earned.