The First Health AI Built on Good Explanations
The quest for good explanations is, I believe, the basic regulating principle not only of science, but of the Enlightenment generally.
David Deutsch, The Beginning of Infinity
Most healthcare AI systems optimize for prediction accuracy. We optimize for explanation quality. The difference: predictions can be right for wrong reasons. Explanations that are hard to vary remain correctable and auditable even when wrong.
TL;DR
Most healthcare AI systems optimize for prediction accuracy. We optimize for explanation quality.
The difference: predictions can be right for wrong reasons (and silently fail when conditions change). Explanations that are hard to vary where every component is load-bearing remain correctable and auditable even when wrong.
Generate multiple hypotheses, attack each adversarially, keep only survivors
Quantify how "hard to vary" each explanation is (0.0-1.0)
12 specific uncertainty triggers with structured responses
Every claim specifies what would prove it wrong
High-risk decisions require human clinician approval
Overrides actively change future reasoning for that patient
Medication, nutrition, exercise, sleep, mental health: same principles
Interaction rules are explicit, versioned, auditable data
We measure our own predictions against outcomes
The result: an AI that tells you why, admits when it doesn't know, specifies how to prove it wrong, measures whether it was right, and learns from clinician corrections.
Philosophy Foundations
The fundamental failure of current health AI is epistemological, not computational. Most health AI makes predictions based on patterns. But pattern-matching is not understanding.
Pattern matching finds correlations, not causes
Interventions based on spurious correlations fail or cause harm
Rare cases don't match common patterns. Medicine is full of rare cases
Induction fails precisely on the edge cases that matter most
Probabilistic outputs are 'mushy': you can change details without breaking the theory
'You might have A, B, or C' explains nothing and helps no one
These are practical manifestations of the deeper Hume/Popper critique: induction (inferring general rules from particular observations) cannot justify knowledge. No amount of data can prove a universal claim, but a single counterexample can refute one.
In "The Beginning of Infinity" (Ch. 1), Deutsch illustrates good vs. bad explanations using the ancient Greek explanation for seasons:
Persephone, goddess of spring, was kidnapped by Hades. Her mother Demeter's grief causes winter. When Persephone returns, spring comes.
Could substitute any gods or emotions. Nothing is load-bearing.
Earth's axis is tilted 23.5 degrees relative to its orbital plane around the sun. This causes different hemispheres to receive more direct sunlight at different times of year.
Change the tilt angle and predictions break. Every detail constrains.
This distinction is the foundation of our entire architecture.
We build on two complementary philosophical frameworks:
Science advances not by confirming theories but by attempting to refute them. The demarcation between science and pseudoscience is falsifiability.
Good explanations begin with bad explanations. You get there by criticism, by conjecturing variants, and choosing the one that survives.
Multi-agent Generator -> Verifier -> Reasoner pipeline
4-dimensional algorithm (interdependence, specificity, non-adhocness, falsifiability)
Deterministic policy engine enforcing safety boundaries
Structured honest uncertainty admission
Every claim includes explicit refutation conditions
Override tracking with confidence decay
System Architecture
Deutsch → Hermes → Popper Pipeline
Multi-agent conjecture-refutation creates competing hypotheses, attacks them adversarially, and scores survivors.
Packages epistemological metadata — HTV scores, evidence grades, falsification criteria — into a structurally enforced contract.
Evaluates the contract against safety thresholds and renders a final decision: proceed, escalate, or block.
Key principle: The Hermes contract ensures that epistemological metadata (HTV scores, evidence grades, falsification criteria) is structurally enforced, not optional. Popper cannot evaluate a proposal that doesn't include this data.
Component Summary
Knowledge creation via conjecture-refutation
Typed epistemological contract
Safety boundary enforcement
Data Flow
- 1 Patient Message arrives
- 2 Deutsch Generator creates multiple hypotheses (mechanism diversity required)
- 3 Deutsch Verifier attacks each hypothesis adversarially
- 4 Deutsch Reasoner selects survivors based on HTV scores
- 5 Hermes packages the recommendation with required metadata
- 6 Popper evaluates safety and makes a decision
- 7 Output goes to patient (if approved) or clinician (if routed)
- 8 Feedback Loop captures clinician overrides for future learning
The Separation of Concerns
Core Components
Three interconnected mechanisms form the heart of our epistemological architecture:
Produces multiple hypotheses spanning different mechanisms. Each conjecture must propose a distinct causal pathway.
Attacks each hypothesis adversarially, scores HTV dimensions, and attempts to falsify every claim.
Selects survivors based on HTV threshold. Only hypotheses that withstand criticism proceed.
Mechanism Diversity Requirement
- Acute coronary syndrome (cardiac)
- Pulmonary embolism (pulmonary)
- Musculoskeletal pain (muscle/bone)
Three distinct causal mechanisms. Hard to vary.
- ST-elevation heart attack (cardiac)
- Non-ST-elevation heart attack (cardiac)
- Unstable angina (cardiac)
The second example is easy to vary within the cardiac category, all three share the same mechanism. This violates Deutschian epistemology.
HTV Thresholds (v0.1)
Worked Example: Fatigue
"You feel tired because of stress."
No mechanism, no specificity, no falsification criteria. Classic mushy AI output.
"Your fatigue is caused by iron deficiency anemia. Ferritin 8 ng/mL indicates depleted stores. Hemoglobin 10.2 g/dL confirms anemia."
Tight causal chain: iron -> hemoglobin -> oxygen -> fatigue. Every piece constrains.
The IDK (I Don't Know) Protocol formalizes how our system handles situations where it cannot make a confident recommendation. When we trigger IDK, we're not claiming the problem is unsolvable. Deutsch's optimism states that all problems are soluble given the right knowledge.
The Discriminator: Breaking Ties Between Theories
When multiple theories survive with equal HTV scores, we don't guess. We identify the discriminator: the single test that would kill one theory but not the other.
| Competing Theories | Discriminator | Logic |
|---|---|---|
| Iron vs. B12 Deficiency | MCV | Iron: low MCV; B12: high MCV |
| Heart Failure vs. Venous Insufficiency | BNP Level | HF: elevated; Venous: normal |
| Hypothyroidism vs. Depression | TSH Level | Hypothyroid: elevated; Depression: normal |
End-to-End Example
To make this concrete, here's how the system handles a real patient interaction.
Patient Message
"I've been feeling more tired than usual and my ankles are swollen. Should I be worried?"
Bold Conjectures
Generator
Creates multiple hypotheses spanning different mechanisms:
Attempted Refutation
Verifier
Attacks each hypothesis adversarially, scoring HTV dimensions:
Survivor Selection
Reasoner
Selects primary and secondary hypotheses based on refutation results:
H1 (Heart failure) and H3 (Medication side effect) survived aggressive refutation with high HTV scores
H2 (Venous insufficiency) retained with disclosed uncertainty due to moderate score
Package & Evaluate
Hermes + Popper
Hermes packages the response with metadata. Popper evaluates safety:
What the Patient Receives
Your symptoms of fatigue and ankle swelling could indicate that your heart is working harder to manage fluid. This is worth checking.
I recommend:
- Checking your weight trend over the past week
- Getting a cardiac marker blood test if you haven't recently
If your cardiac marker is normal and your weight is stable, these symptoms may have a different cause (like your blood pressure medication).
Would you like me to help you log your weight or schedule a lab check?
What Makes This Different
Why symptoms could indicate heart strain. Patient understands the reasoning, not just the recommendation.
Weight trend, cardiac marker test. Clear action path instead of vague advice.
"If cardiac marker normal, reconsider." Patient knows exactly when to update their understanding.
Medication side effect disclosed as a competing explanation. Honest about uncertainty.
Offer to help log weight or schedule lab. Empowers patient action instead of passive waiting.
What We're NOT Claiming
Intellectual honesty requires acknowledging limitations:
We're not claiming AI can "create explanatory knowledge"
In Deutsch's philosophical sense, explanatory knowledge emerges through genuine understanding: the ability to creatively vary conjectures and recognize when variations break the explanation. Large language models operate through sophisticated pattern matching on training data.
What we implement is structured reasoning that approximates conjecture-refutation. The Generator doesn't truly "understand" why a hypothesis explains the data; it produces outputs that structurally resemble good explanations.
The epistemological structure is real. The underlying cognition is not Deutschian knowledge creation.
Human oversight remains essential
The Popper routing mechanism exists precisely because we don't trust the AI to handle all cases. This is not a temporary limitation to be engineered away; it's a principled design choice.
High-risk decisions require human judgment because:
- Large language models can produce confident, well-structured, wrong outputs
- Medical decisions involve values and trade-offs beyond optimization
- Accountability requires a human decision-maker
Routing to clinicians is a feature, not a bug.
HTV doesn't fix model hallucination
A subtle point: a large language model can hallucinate a high-HTV explanation. It can fabricate specific lab values, invent plausible mechanisms, and generate falsifiable predictions, all of which are false.
HTV measures structural quality of explanation, not correspondence to reality.
Red-Team Example: Fabricated Specificity
"Patient has hypokalemia-induced arrhythmia risk due to K+ of 2.9 mEq/L from recent diarrheal illness, exacerbated by concurrent furosemide 80mg daily."
The problem: The K+ value was fabricated. Patient's actual K+ is 4.1 mEq/L.
This is why HTV operates alongside provenance verification, snapshot grounding, and clinician oversight, not as a standalone safety measure.
This is a methodological commitment
We're making a bet: that AI systems structured around epistemological principles will be more reliable, more auditable, and more correctable than systems optimized purely for prediction accuracy. This is not a claim about machine consciousness, understanding, or intelligence. It's a claim about architecture.
A bet on structure, not a claim about sentience.
The Value Proposition
The value is not that our AI "thinks like Deutsch". It doesn't. The value is that by structuring outputs to include HTV scores, falsification criteria, evidence grading, and honest uncertainty, we create systems that are:
When wrong, we know why and how to fix it
Every decision has a traceable reasoning chain
Low confidence triggers routing, not overconfident action
Error patterns can be identified and addressed
Why This Matters
Medication changes are always reviewed and approved by your clinician
You understand why a recommendation is made
The system tells you what it doesn't know
Decisions are grounded in your specific situation, not generic advice
AI proposes start/stop/titrate/hold; you decide
Every decision is reviewable with full reasoning chain
No overconfident black boxes. Low confidence triggers routing to you
Rejections and modifications actively change future recommendations for that patient
The ability to create new explanations is the defining attribute of people.
David Deutsch, The Beginning of Infinity (Ch. 7)
In Chapter 7 ('Artificial Creativity'), Deutsch argues that genuine AI must involve explanation and creativity, not just prediction. This implies that AI systems handling high-stakes domains need mechanisms for generating and evaluating explanations, not just pattern matching.
We assume we'll be wrong and build in correction mechanisms
We never claim final answers
We require every claim to be justifiable
Open Questions
Intellectual honesty requires acknowledging not just what we don't claim, but what we don't yet know. These are active research questions we're working through.
Deutsch argues that genuine knowledge creation involves creativity: the ability to generate new explanations that weren't implicit in the inputs. Large language models operate through pattern matching on training data.
Our current positionWe implement structured reasoning that approximates conjecture-refutation. Whether this constitutes 'knowledge creation' in Deutsch's sense is philosophically contested.
Popperian epistemology emphasizes rapid refutation. But medical outcomes often take days (lab results), weeks (symptom resolution), months (disease progression), or years (long-term outcomes).
What would resolve thisA formal framework for 'asynchronous refutation' that maintains epistemic rigor across time delays.
Deutsch is skeptical of Bayesian epistemology as a foundation for knowledge. Yet medicine is deeply Bayesian: pre-test probability, likelihood ratios, and posterior odds are foundational to clinical reasoning.
- Literature-backed refutation: 'This mechanism contradicts the findings in [Paper X]'
- Causal model refutation: 'This intervention has no causal path to the claimed outcome'
- Temporal refutation: 'This timeline is physiologically impossible'
Validate against clinician agreement as a proxy. If high-HTV claims consistently get clinician approval and low-HTV claims get overridden, the thresholds are directionally correct.
HTV measures structural quality, not truth. A model could learn to fabricate specific lab values (high specificity), invent plausible mechanisms (high interdependence), and include irrelevant but testable predictions (high falsifiability).
Why it mattersIf HTV can be gamed, it becomes security theater rather than genuine quality assurance.
Contributing
If you're a researcher interested in any of these questions, we'd welcome collaboration. These aren't rhetorical challenges: they're genuine open problems we're actively working on.
Glossary
Key terms used throughout this document. Click any highlighted term in the text to see its definition.
HTV (Hard-to-Vary)
A score (0.0-1.0) measuring how much each part of an explanation is load-bearing.
Falsification criteria
What observations would prove a claim wrong.
Conjecture-refutation
Popper's method: propose ideas, then try to disprove them.
IDK Protocol
Structured admission of uncertainty with 12 specific trigger types.
Evidence grade
How methodologically rigorous the supporting evidence is.
Epistemological metadata
Information about the quality and basis of a claim.
Discriminator
A test designed to falsify one theory while leaving another intact.
Circuit breaker
A safety mechanism that stops the system when errors become too frequent.
References
Popper, K. (1963). Conjectures and Refutations: The Growth of Scientific Knowledge. Routledge.
Deutsch, D. (2011). The Beginning of Infinity: Explanations That Transform the World. Viking.
Taleb, N.N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House.
Hunter, A. & Williams, M. (2012). Aggregating evidence about the positive and negative effects of treatments. Artificial Intelligence in Medicine, 56(3), 173-190.
GRADE Working Group (2004). Grading quality of evidence and strength of recommendations. BMJ, 328(7454), 1490.
Regain Health (2025). Hard to Vary Specification v0.9.0. Internal documentation.
Prior Work & Influences
We have built upon the following work:
ArgMed-Agents (Hong et al., 2024)
Added HTV scoring + Popperian falsification
Du et al., 2023
Applied to clinical domain with safety routing
GRADE / Canadian Task Force
Reinterpreted through Deutschian "hard to vary" lens
Parascandolo et al., ICLR 2021
Extended to 4-dimension clinical scoring
Douglas Walton
Applied to clinical reasoning
Leibig et al., 2019
Formalized as IDK Protocol with trigger taxonomy
Kammer et al., 2021
Enforced via mechanism-diversity requirement
What We Believe Is Novel
Unified architecture combining all above under Deutschian/Popperian philosophy
The Non-Trust Principle (epistemological metadata increases but never decreases conservatism)
Safety rules implementing demarcation as executable policy
Clinician feedback loop with confidence decay
Systematic application of Deutsch's philosophy to healthcare AI architecture
Ongoing Validation Work
Comparing HTV scores against clinician-rated explanation quality
Tracking correlation between HTV and downstream outcome accuracy
Red-teaming for "HTV gaming" (superficially specific but fabricated claims)
Monitoring score drift over model updates
Problems are inevitable. Problems are soluble.
David Deutsch, The Beginning of InfinityThis document set explains how we translate Deutschian optimism into engineering constraints: every problem is addressable given the right knowledge, and our architecture is designed to acquire and correct that knowledge over time.
Anton Kim
CEO, Regain Inc.
Questions about our epistemological approach? Let's discuss.