Building the First Healthcare AI on Sound Epistemological Foundations
Based on work of Karl Popper and David Deutsch
Regain Health
Patient Health OS
AI-powered mobile app delivering personalized health protocols, daily coaching, and continuous optimization based on lab results and wearable data.
Regain Medical
Provider OS
AI-native Medical Information System reducing documentation time by 75%, automating clinical workflows, and enabling data-driven decision making.
Healthcare is an
Epistemology Problem
A good explanation is hard to vary while still accounting for what it purports to explain.
David Deutsch, The Beginning of Infinity
The fundamental failure of current health AI is epistemological, not computational. We apply Popperian critical rationalism - conjecture and refutation - to build AI that creates genuine explanations, not probabilistic guesses.
The Inductive Fallacy in Healthcare AI
Most health AI makes predictions based on patterns. But pattern-matching is not understanding.
Correlation ≠ Causation
Most AI health tools confuse 'A usually follows B' with 'B causes A'
Pattern matching finds correlations, not causes.
The Black Swan Problem
Rare cases don't match common patterns. Medicine is full of rare cases.
Induction fails on edge cases that matter most.
Easy-to-Vary Explanations
Probabilistic outputs are 'mushy' - you can change details without breaking the theory
'You might have A, B, or C' explains nothing.
We don't want an AI that hallucinates; we want an AI that arrives at hard-to-vary explanations.
Critical Rationalism in Medicine
We do not look for confirmation; we look for good explanations.
Conjecture
Generate bold explanations for the patient's health state
"What biological mechanisms could explain this?"
Refutation
Systematically attempt to destroy each explanation
"What evidence would kill this theory? Does the data already contain it?"
Survivor Selection
Only explanations that withstand rigorous criticism survive
"Theories survive by being hard-to-vary, not by being likely"
Conjecture
Generate bold explanations for the patient's health state
"What biological mechanisms could explain this?"
Refutation
Systematically attempt to destroy each explanation
"What evidence would kill this theory? Does the data already contain it?"
Survivor Selection
Only explanations that withstand rigorous criticism survive
"Theories survive by being hard-to-vary, not by being likely"
What Makes an Explanation 'Good'?
A good explanation is one that is hard to vary while still accounting for the phenomenon it explains.
Interdependence
Change one part, the whole thing collapses
Components are logically connected
Specificity
Makes precise, testable predictions
Not vague statements but concrete claims
Parsimony
Explains without arbitrary assumptions
No ad-hoc additions to save the theory
Falsifiability
There exists a clear data point that would prove it wrong
The theory makes itself vulnerable
See the Difference
Bad Explanation
You feel tired because of a virus.
Good Explanation
Your fatigue is caused by Epstein-Barr reactivation. The virus infects B cells, triggering a cytokine response. Your elevated liver enzymes (ALT 67) confirm hepatic involvement.
The Debate Architecture
Simulating a World-Class Medical Team via Adversarial Debate
The Intake Agent
Build the Problem
Maps the patient's Health State into a structured timeline of unexplained phenomena
The Conjecturer
Generate Theories
Bold hypothesis generation: What biological mechanisms could explain this?
The Critic
Destroy Theories
Systematic refutation: What evidence would kill each theory?
The Synthesizer
Select Survivors
Picks the theory with fewest contradictions
Feedback Loop: If refuted, return to Conjecturer for new hypotheses
The Intake Agent
Build the Problem
Maps the patient's Health State into a structured timeline of unexplained phenomena
The Conjecturer
Generate Theories
Bold hypothesis generation: What biological mechanisms could explain this?
The Critic
Destroy Theories
Systematic refutation: What evidence would kill each theory?
The Synthesizer
Select Survivors
Picks the theory with fewest contradictions
Technical Implementation
The system implements the Generator-Verifier-Reasoner (ArgMed) pattern within a LangGraph state machine. Theories are validated against medical consistency schemas using formal verification.
In Popper's framework, a theory that survives rigorous attempts at refutation earns provisional acceptance - never certainty, but the highest epistemic status achievable.
The IDK Protocol
When We Say 'I Don't Know'
A critical failure of current AI is hallucination under pressure - generating confident-sounding answers when it doesn't know. We prioritize epistemic humility.
Three Triggers That Activate IDK
All Theories Refuted
Every candidate explanation has been falsified by available evidence. No surviving hypothesis remains.
Equally Easy-to-Vary
Surviving theories score identically on hard-to-vary criteria. No explanation stands out as more robust.
Missing Critical Data
The discriminating data point that would decide between theories is not available in the current dataset.
What Happens Instead of Guessing
"Current knowledge is insufficient to distinguish between [Theory A] and [Theory B]."
"The primary bottleneck is the lack of [Specific Data Point]."
"Action: Safest-case posture until [Metric] is updated."
We do not guess when we don't know.
The Role of Discriminators
When theories tie, we identify the discriminator - the single test that would kill one theory but not the other.
| Surviving Theories | Discriminator | Logic |
|---|---|---|
| Overtraining vs. Insulin Resistance | Morning Cortisol | Overtraining → elevated; IR → normal |
By identifying the discriminator, we guide the next data collection step rather than making unfounded claims.
Open Questions
Problems We're Still Working On
Knowledge Creation vs. Retrieval
"Is our multi-agent debate genuinely creating new knowledge, or is it sophisticated retrieval dressed up as knowledge-creation?"
An LLM Critic might just be retrieving counterexamples from training data, not generating genuine counter-arguments.
Falsifiability with Delayed Ground Truth
"In medicine, ground truth often takes weeks or months."
How do you think about falsifiability in domains where refutation is delayed?
Hard-to-Vary vs. Prior Probability
"Is there a coherent way to combine 'hard-to-vary' with 'prior probability'?"
Or are they fundamentally incompatible epistemologies?
Scaling Refutation
"Our Critic uses schema-driven refutation. What other refutation mechanisms should we consider?"
Empirical falsification, logical consistency checking, expert panel review, patient outcome feedback loops.
Vision & Product Docs
Explore the technical and philosophical foundation of the Popperian Assessment Engine.
Vision & Philosophy
Product & Architecture
Join Our Team
We're looking for AI researchers with a strong interest in epistemology to develop the Popper-Deutsch assessment agentic system and apply it to healthcare.
