The Function of the Message

Tuesday afternoon. A 78-year-old woman on the cardiology floor, day three of admission for decompensated heart failure. Five clinicians will say something about her in the next forty-five minutes, each compressing the patient differently for a different audience.

The hospitalist on the receiving end of sign-out, around six, gets the handoff: "Mrs. M, decompensated HF on diuresis, net negative 2L today, K is 3.4 and she just got 40 mEq, recheck mag in the morning, pending echo. Watch for sudden room-air desaturation overnight; she had a brief episode at noon, resolved on her own. Call if it recurs."

The cardiology fellow gets a consult question from the medicine team: "Should we resume her metoprolol now that she's volume-down, or wait until tomorrow?"

The PCP, two days from now, will read the discharge summary: "Admitted with NYHA III decompensation in setting of medication non-adherence after spouse's hospitalization. Improved with IV diuresis. Discharged on oral regimen including resumed metoprolol at half-dose. Follow-up echo in two weeks."

The chart note for today reads differently still: "Day 3 HFrEF exacerbation, improving. Net -2L today, -6.5L cumulative. K 3.4 repleted. Brief episode of self-resolving desat at noon, no clear precipitant; will reconsider PE workup if recurs. Plan to titrate metoprolol back tomorrow. Discharge anticipated 1-2 days."

And on rounds, the attending tells the resident: "Notice she desaturated at noon. Volume is coming down but the right heart has not had time to remodel. That is why you do not restart beta-blockade until the diuresis is complete and the vitals have stabilized."

Five compressions of the same patient in the same afternoon. Each one is the right message for its job. Each one would be the wrong message in any of the other four contexts.

Compression is utility-relative

The principle the rest of this series rests on, stated plainly: compression is utility-relative. There is no universally optimal compression of a patient. Every clinical message is judged against three dimensions simultaneously.

Function. What job the message must perform: what decision it is supporting, what risk it is warning about, what evidence it is justifying, what concept it is teaching, what handoff of ownership it is performing. Different functions demand different content.

Receiver. Who or what must decompress the message: a trained clinician, a junior trainee, a layperson, a consultant from another service, a database, a language model, a retrieval system. The receiver determines what shapes are decodable and what codec the message can assume.

Cost. The total cost the message imposes on whoever must read it: transmission, decoding, attention, computation, time, and risk of error. The cheapest sufficient representation wins, where cheapness is measured for the receiver who actually has to do the decompression.

The three are not separable. The right compression is the one that performs the function, lands on a receiver who can decompress it, at the lowest total cost. Same patient, different functions, different receivers, different costs, different compressions. The "good" compression is judged against the triad, not against completeness, not against brevity in the abstract, not against any other message that did a different job for a different receiver. This is the first move that has to land before the rest of the curriculum can mean anything. Parts 3 through 6 will operationalize the principle (at what abstraction level to start, how to order the elements, what to leave out, what to keep when in doubt). Each of those is downstream of knowing the function the message is doing, the receiver who must decompress it, and the cost budget that frames both.

A fact has no intrinsic value to a clinical message. It has value only insofar as it changes the receiver's ability to act, monitor, decide, coordinate, teach, document, or reconstruct the patient state, given what the receiver is trying to do. A true fact can be high-utility, low-utility, or even negative-utility depending on context. Negative-utility facts distract, overload, mislead, and bury the active decision; including them costs the receiver attention without paying back in better action. The unit of value in a clinical message is not the fact. It is the fact's expected contribution to the receiver's objective, weighted by the cost the receiver pays to extract it.

Use the codec, or train it

Every clinical message does one of two things to the receiver. It updates their model of the patient at hand, or it updates the codec they carry from one patient to the next.

The model is the receiver's picture of this patient — a compressed generative representation that supports inference, prediction, explanation, and action, not a flat fact table.

Disease models, patient models, mechanistic models, predictive models, and policy models are not separate categories. They are the same underlying representation at different scopes or queried through different projections. A patient model is a disease (or world) model conditioned on this particular patient. A policy is the action-facing projection of that model — sometimes compiled into a trigger-action rule (allergy, DNR, "watch for sudden desat") that the receiver can use without fully decompressing the mechanism behind it. Part 7 (The Belief-State Object) develops the rigorous form: a mixture distribution over hypothesis-components, eight fields each.

The codec (Part 1) is the reusable machinery the receiver has built through training — the disease handles, causal structures, risk patterns, policy rules, and language conventions that build and interpret models. The codec produces a model when applied to a case. The codec is slow-changing; the model is fast-changing and specific to one encounter.

A message can therefore land in three places. It can update the patient model. It can transmit a projection of it — a policy handle, a risk handle, a sensor-meaning handle, an action-facing answer. Or it can update the codec itself, for future, different patients.

Use the codec. Apply the receiver's existing codec to update their model of this patient — either the full patient model (handoff, consult question) or a specific projection of it (a policy from a sign-out warning, an action-facing answer to a bounded question). A handoff that lets the next clinician make the next decisions. A consult question that asks for a bounded recommendation. A sign-out warning that pre-arms a response to a specific deterioration. A chart note that lets a future reader make a downstream decision. The codec is consulted but not changed. What changes is the receiver's model or its projection for this one patient, and the decisions that fall out of it.

Train the codec. Update the codec itself so the receiver will compress and decompress future patients differently. A teaching point on rounds that hangs a rule on one case. An M&M conference that updates the team's safety priors. A case report that revises a reader's mental DDx. Training data that updates an AI's weights. The current patient is the vehicle; the target is the codec the receiver will bring to the next, different patient.

These are different operations and the compression that fits one rarely fits the other. One utterance can do both across receivers — a rounds presentation uses the codec for the discussing attending and trains the codec for the resident.

Sender intent does not constrain receiver behavior. A handoff designed for use can be unintentionally trained on — the receiver notices a pattern and the codec quietly updates. A teaching point designed for training can be operationally used for the case it names. An AI at inference time uses; the same AI in a training pipeline trains on the same input. The writer optimizes for the intended operation. The receiver's options exceed the intent.

Five shapes

Within those two operations, five common compression shapes recur. Four are sub-shapes of using the codec, varying by how soon the decision happens and how bounded it is. One is the named shape within training — single-case rule extraction; other training shapes exist (population research, AI weight updates) but the rest of this curriculum focuses on the rule-extraction case. The list isn't exhaustive and the boundaries aren't sharp; the shapes are common enough to be worth naming so the rest of this curriculum can refer to them.

Manage (use). The receiver will be making an open-ended set of decisions about this patient. The compression demands trajectory, outstanding work, anticipated decision points, and what to call about. It does not need the full admission history, the imaging archive, or every lab value since arrival. The next clinician will be making decisions in the next twelve hours, not writing a paper. Typically a handoff or transfer of service, but multiple receivers can be in this shape at once — cross-cover, primary team, and bedside nurse may all be managing the same patient simultaneously, and the same message can serve any of them. The capability has landed when the receiver can write the next safe order without re-reading the chart.

Recommend (use). The receiver is being asked something specific and must return a bounded answer. The compression must lead with the question. The consultant's time is finite, the question is concrete, and the evidence relevant to it is a small subset of the chart. "Should we resume her metoprolol now that she is volume-down" is the right shape: bounded question, decision context implicit, the relevant variables (volume status, ejection fraction, medication tolerance) implied. The full HF history is available if the consultant wants it; the question does not require its prefiguration. Typically a consult, curbside, or focused page. Manage and Recommend share a soft border at consult-questions-that-extend-to-co-management; the difference is the scope of decisions the receiver must be ready for, and a single artifact can sit between them. The capability has landed when the receiver has a defensible recommendation with explicit criteria for when it would change.

Anticipate (use). The receiver may need to act on a deterioration that has not happened yet, and must be primed to catch the trigger when it occurs. Compression here is dominated by what could go wrong and what to do. "Watch for sudden room-air desaturation; she had a brief episode at noon, resolved on her own" tells the next clinician exactly what to look for and signals what to do without the receiver having to reconstruct the reasoning. Typically a sign-out warning or escalation precaution, though often embedded inside a handoff or a chart note rather than standing alone. The capability has landed when the receiver would recognize the trigger if it occurred and knows the response.

Reconstruct (use). A future reader, possibly unknown, will rebuild what happened, out of context, to make a downstream decision or to evaluate the case. Compression must preserve enough fact base for someone six months from now, possibly without team context, to reconstruct what happened and why. This is the most fact-dense and slowest-reading of the five; future readers can take their time. Numbers, dates, doses, and reasoning chains belong here in ways they do not belong in the other shapes. Typically a chart note, discharge summary, or audit-readable record. The capability has landed when a stranger six months later can rebuild the case and the reasoning behind it from the artifact alone.

Generalize (train). The receiver extracts a pattern they can apply to the next, different patient. The current case is the vehicle, not the target. What matters here is what discriminates: which findings would change the diagnosis, which actions would change the trajectory, which reasoning would transfer. The attending who says "you do not restart beta-blockade until the diuresis is complete" is teaching a rule by hanging it on one case, not summarizing the patient. Typically rounds, attending feedback, or post-call wrap-up. The capability has landed when the receiver can state the rule and apply it to a different patient tomorrow.

These shapes stack inside one utterance. The Mrs. M handoff above performs Manage and Anticipate simultaneously — both inside "use the codec" — because the receiver is both taking on open-ended decisions and being pre-armed for a specific deterioration. The attending's rounds line performs Generalize for the resident (training their codec) while also feeding the attending's own Manage the next time a similar patient lands (using a codec they are refining as they talk). Same utterance, different operations on different receivers.

Receiver architecture

The function-matching rule is necessary but not sufficient. A message is not small in the abstract. It is small for a receiver. The same representation can be effortless for one architecture and burdensome for another, and what counts as "the smallest message" depends on which architecture is doing the decompressing.

Consider four representations of the same patient state. A chest X-ray image is information-dense; for a trained radiologist it decompresses in under a second, and for a language-only AI fluent only in text it is far more expensive to interpret. A handoff sentence like "EF 25 percent, cold and wet, rising creatinine, failed outpatient diuresis" is compact for a clinician but points at nothing for a layperson. A structured table of vitals trends is awkward for a clinician to scan mid-shift, but trivial for a database to query and trivial for a dashboard to render. A vector embedding of the patient is information-rich for a retrieval system and useless to any human reader. Same patient state, four representations, four wildly different costs depending on who is doing the work.

The total cost of a message is the sum of its transmission cost, its decoding cost, the attention it claims, the computation it requires, and the risk of error it carries. The "smallest" message is not the one with the fewest tokens or the most compact form. It is the one that minimizes total cost for a particular receiver while still carrying the information the function requires.

This matters specifically for clinical AI. The representation that is most efficient for the AI to reason over is rarely the representation the clinician can act on. The representation the clinician naturally produces is rarely the one the AI can index efficiently. Communication across that gap requires translation, and translation is itself a compression operation that has to be tuned to its own receiver. The system that helps does this translation work invisibly, hiding the cost on each side from the other. The companion essay The Receiver Is the Codec develops the underlying claim — that no message has value independent of an accompanying codec — in non-clinical terms.

What this rules out

The most common failure of clinical communication is shape-mismatch — the artifact ships in the shape of one capability when the receiver's job requires another, or it does training work when the receiver needs to use, or use-work when the receiver needs to learn. A handoff that reads like a discharge summary: a Reconstruct artifact delivered to a clinician who needs to Manage a live patient in the next sixty seconds. A consult request that reads like a chart note: pages of evidence wrapped around a question the consultant only needs to Recommend on. A progress note that does anticipation work the future chart-reader cannot act on. A teaching point compressed into a one-liner so dense the trainee cannot Generalize from it. An AI inference that reads like a journal article when the clinician needs a bounded recommendation — train-shaped output handed to a receiver doing use.

A second class of failure is receiver-mismatch. The chart note written in radiology shorthand handed to a primary care reader. The vitals dashboard exported to a clinical AI that has no native handling for tabular streams. The clinical-AI output emitted in dense natural language to a clinician who is mid-procedure and cannot read prose. Each is a representation that would have been fine for a different receiver and is wrong for this one.

The dashboard a clinical AI emits at the bedside is another instance of both failures stacked. A general-purpose differential copilot that returns the same shape of output whether the clinician is signing out, ordering, asking a question, or charting is doing the same thing as a clinician who hands every audience the same paragraph. The output may be true. It will not be the right message for the job the clinician was doing when they opened the screen, and it will not be in a form the clinician can decompress at the cost they have available right then.

The operational rule

The best message is the lowest-cost sufficient representation for the receiver's objective and architecture.

Cost means total cost: transmission, storage, decoding, attention, computation, and risk of error. Sufficient means the message carries the information the receiver needs to perform the function, judged by each fact's expected contribution to that objective rather than by completeness. Receiver's architecture means the substrate that will be doing the decompression: a trained clinician, a layperson, a database, a language model, a retrieval system, a junior trainee, a senior attending. The same information can be sent in many shapes; the right shape is the one whose total cost is lowest for the receiver who will actually do the work.

This sentence is the principle the rest of the curriculum depends on. Part 3 (The Highest Accurate Abstraction) operationalizes how to pick the abstraction level once the function and receiver are known. Part 4 (The Decompression Order) operationalizes how to order the elements within that level. Part 5 (The Minimum Sufficient Message) operationalizes what to leave out for a receiver with the relevant codec. Part 6 (The Grounding Constraint) operationalizes what to keep when in doubt about the receiver's decompression. Each of those is downstream of this one: there is no universal smallest message, only the smallest one that does the job for the receiver in front of the writer.

At the bedside

The five messages about Mrs. M on Tuesday afternoon are a sign of clinical communication working as it should, not of disorganization: each compression fit to its function, each one wrong everywhere else, each one the right thing in the place it actually goes. A clinical AI that helps with any of them has to know which function it is helping with. Otherwise it will return the right shape of message for the wrong job, and the right shape for the wrong job is the same thing as the wrong message.

---

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