Preface: Healthcare Is Compressed Coordination

This curriculum is one idea, walked across eleven parts: healthcare is compressed coordination under uncertainty. No finite agent can carry the whole patient: the physiology, history, trajectory, risk, uncertainty, social context, and plan. Each clinician, consultant, nurse, AI system, patient, and family member carries a compressed representation of the patient, shaped by what they need to do next.

Clinical communication works when agents reconstruct sufficiently aligned belief states from minimal messages. Clinical safety fails when those compressed models silently diverge. Clinical AI should become the infrastructure that exposes, synchronizes, and corrects those belief states before irreversible action. The receivers do not need identical belief states; the nurse, the consultant, the patient, and the AI each need a task-shaped version. What they need is alignment sufficient for safe action, and the architecture that catches drift when alignment slips.

A sign-out that says "COPD exacerbation, improving, discharge tomorrow" is a compressed belief state, not a summary of the patient. It tells the receiver what disease model to load, what trajectory to expect, what risks are probably low, and what actions are likely next. If the sender omitted that the patient had rising CO₂, poor home support, or an unresolved PE concern, the receiver may reconstruct the wrong patient from a confident message. The compression worked at the surface and failed at the belief state.

Compression fails in three ways. The wrong thing gets dropped: an action-changing fact never reaches the receiver. The wrong thing gets filled in: the receiver reconstructs a confident picture from a confident message, and the picture is wrong. Or the gap is never checked: no one verifies whether the receiver's picture matches what the sender meant. The work of safe clinical communication is the architecture that catches all three failures, and the work of clinical AI is to run that architecture continuously across the team.

I came to this from medicine, where the failure modes are loud. A clinical sign-out is a wildly compressed message: a dozen words in place of a person, handed off in seconds at change of shift. The compression works when the receiver shares the codec; it fails, sometimes catastrophically, when they do not. The same compression happens at every scale of clinical communication. The same compression happens, with new failure modes, every time an AI joins the team.

The arc, in four phases

The eleven parts trace one arc across four phases.

Phase 1 (Parts 1-6): the compression mechanics. How clinical messages work, from the substrate up.

• 1. The Compression Substrate. Reality is too large; clinical agents communicate by compression.
• 2. The Function of the Message. Compression is purpose-relative, utility-based, and receiver-architecture-relative. The best message is the lowest-cost sufficient representation for the receiver's objective and architecture.
• 3. The Highest Accurate Abstraction. Disease, syndrome, complaint-pattern, raw observations: pick the highest accurate level the receiver can decompress.
• 4. The Decompression Order. Base model first, relevant nuance second, evidence and action third.
• 5. The Minimum Sufficient Message. What it looks like when a message has been tuned to a single point.
• 6. The Grounding Constraint. Anchor only what prevents dangerous ambiguity; include the evidence that would reverse the receiver's reconstruction if absent.

Phase 2 (Part 7): the cognitive object.

• 7. The Belief-State Object. What is actually being compressed is a structured belief state: a mixture over hypothesis-components, each carrying its own evidence dependencies, predicted trajectory, action threshold, and disposition implication.

Phase 3 (Part 8): the failure mode.

• 8. The Same Wall. Safety fails when private compressed models silently diverge. The same substrate-independent failure mode appears in intelligence analysis, aviation, and distributed systems.

Phase 4 (Parts 9-11): the clinical AI architecture.

• 9. The Defense Architecture. Clinical AI implements shared compression and error correction. Two halves, both required.
• 10. The Temporal Loop. Align before, synchronize during, update after. Three temporal layers compounding into a continuous loop.
  • 10.1 The Connection. The verification architecture: signal meeting explicit state across four channels.
  • 10.2 Quiet Verification. Friction discipline for the verification half of runtime synchronization.
  • 10.3 Quiet Acquisition. Friction discipline for the acquisition half.
• 11. The Irreversible-Action Check. The capstone. Before discharge, escalation, sign-off, or handoff: expose, synchronize, and correct belief states across the team.

Where this sits among existing frameworks

The FDA's Software as a Medical Device guidance, the Coalition for Health AI (CHAI) assurance standards, the FUTURE-AI consortium's guidelines, the TRIPOD-AI reporting checklist, the Duke deployment framework from Sendak and colleagues, and the ONC's HTI-1 transparency rule all touch parts of what this curriculum names. Most put their weight on uncertainty and reporting transparency, and on post-deployment monitoring and bias evaluation — the operational maintenance layer the curriculum returns to repeatedly. None, as far as I can find, names belief-state divergence between the model and the clinician at the moment of action as a first-class property a system either has or lacks. That is the gap this curriculum names.

How to read it

If you are a clinician, Phase 1 is about the work you already do, named for what it is. Phase 2 is what you implicitly hold in your head when you reason about a patient. Phase 3 is the failure mode that hurts patients when no one notices. Phase 4 is what clinical AI should be doing instead of writing notes.

If you build clinical AI, Phase 1 is the missing background most ML papers assume but rarely write down. Phase 2 is the data structure your system probably does not maintain explicitly. Phase 3 is the failure mode your evaluations probably do not catch. Phase 4 is the product category most current clinical AI sits adjacent to without occupying.

If you are neither, the whole sequence is a general theory of compressed coordination, clinically instantiated — one way of seeing how minds communicate across rooms, across professions, and across substrates, where compression is the medium and silent divergence is the danger. Medicine is where the stakes are visible in time-stamped, attestable, irreversible form, but the failure modes are substrate-independent.

Each part stands alone. The order above is the recommended order, but the parts are not chained. Pick whichever title looks most interesting and start there. If you have time for two, read Part 1 (the substrate) and Part 11 (the capstone): one tells you what the medium is, the other tells you what the architecture is for.

The capstone names what the rest builds toward: clinical AI as the belief-state coordination layer for healthcare, with the highest-value intervention at the moment before irreversible action.

Compressed Medicine · Preface · 1. The Compression Substrate · 2. The Function of the Message · 3. The Highest Accurate Abstraction · 4. The Decompression Order · 5. The Minimum Sufficient Message · 6. The Grounding Constraint · 7. The Belief-State Object · 8. The Same Wall · 9. The Defense Architecture · 10. The Temporal Loop · 10.1 The Connection · 10.2 Quiet Verification · 10.3 Quiet Acquisition · 11. The Irreversible-Action Check