The Grounding Constraint

ED-to-floor transfer at 19:00. Mr. J, 62, came in with chest pain. The ED team has done the workup; he is admitted for overnight observation. The handoff to the floor hospitalist can be one of two messages.

Version one: "Mr. J, 62, chest pain, low-risk. Stable. Plan: obs overnight."

Version two: "Mr. J, 62, low-risk chest pain by HEART score 2, serial troponins flat, ECG nonischemic. Hemodynamically stable on room air, no current pain. Plan: observe overnight, repeat troponin at 0500, ambulation test in the morning, discharge if all stable."

Same patient. Same evidence. The first message uses "low-risk" and "stable" as compressions the receiver will act on, without grounding. The second message anchors each compression to the evidence the receiver would need to reconstruct what was meant. The first hands the receiver four ambiguous words. The second hands them a model they can act on.

For the fresh hospitalist coming on shift, version one is dangerous. "Low-risk" by what criterion: HEART, TIMI, gestalt? "Stable" how: vitals stable now, trajectory stable over the last hour, stable off pressors, stable on room air? "Obs overnight" with what criteria for discharge, what to call about, what to repeat? The receiver has to phone the ED for clarification. Version two prevents that call by grounding the words the receiver will act on.

The principle

Ground only what prevents dangerous ambiguity. The job of grounding is to attach enough evidence to each compressed claim that the receiver can reconstruct the same model the sender meant, with the same thresholds, the same trajectory expectations, and the same action implications. The chart contains everything; the message should contain what the reconstruction depends on.

The goal is controlled decompression rather than maximal brevity. The receiver should reconstruct the intended state, not merely hear the words.

Two failure modes

The principle has two failure modes that pull in opposite directions.

Under-grounding. The compression the receiver will act on is asserted without the evidence that would let them bind it. "Patient is stable" is the canonical instance. The sender knows what stable means in this case; the receiver does not. If the receiver decompresses "stable" the same way the sender meant it, no harm is done. If the receiver decompresses it differently (the way the verification series called silent reconstruction), the team is now operating on two different models with the same word in front of them. Under-grounding is how "stable" becomes a token doing more work than the evidence in the message supports.

Over-grounding. The opposite failure. The sender lists every vital sign, every lab, every symptom, every historical fact, every imaging result, every prior consultant's opinion. The receiver gets buried. The relevant evidence is in there somewhere, but it is sitting next to thirty pieces of irrelevant evidence and the receiver has to do the extraction themselves. Over-grounding is the chart note that runs five paragraphs because the writer cannot decide what to leave out. It is the consult request that includes the patient's history of present illness when only the current question matters. Over-grounding is not "thorough"; it is offloaded triage.

The grounding rule

The rule that resolves both failures is the same in both directions. Include the evidence that would cause the receiver to reconstruct a different model if absent. Omit the rest.

Applied to under-grounding: if the receiver might decompress "stable" or "low-risk" or "improving" differently from how the sender meant it, the sender owes the receiver the discriminating evidence. The HEART score, the troponin pattern, the hemodynamic specifics, the trajectory direction.

Applied to over-grounding: if a piece of evidence would not change the receiver's reconstruction whether present or absent, it does not belong in the message. The normal labs, the unremarkable history, the irrelevant prior visits, the default-correct details. They can live in the chart for later retrieval. They do not belong in the active message.

The criterion is reconstruction. The compressed claim plus the included grounding should let the receiver build the same model the sender has. Less, and the model diverges. More, and the receiver has to filter through irrelevant material to find the pieces that actually matter.

Variation by function and receiver

The grounding budget varies across both the function the message is performing and the receiver who will decompress it.

What needs grounding for one receiver does not need grounding for another. "Stable" said to a senior cardiologist decompresses into the same model the sender meant, because the receiver shares the codec; "stable" said to a primary care doctor returning a referral is a different message that needs different grounding to land safely. The same handoff sentence carries a different grounding requirement when it lands on a clinician trained on the same service versus one cross-covering from another. Receiver-architecture is one of the inputs to the grounding decision; what would be reconstructed wrongly without grounding is a property of the receiver as much as of the message.

The grounding budget also varies across the functions Part 2 named.

For handoff, ground the words the next clinician will be acting on. "Stable on pressors" instead of "stable" if pressors are running. "Trending up" instead of "stable" if the trajectory is the claim the receiver will act on. The grounding is small (one or two phrases) but specific to what could change the call.

For consult requests, ground the decision context. The consultant needs the question first and the evidence relevant to that question. Other evidence is available in the chart; it does not need to be in the message.

For chart notes, the grounding budget is larger because the future reader is unknown. A reviewer six months from now does not have the team's context. The note has to include enough evidence that the assessment and plan can be reconstructed without contemporaneous knowledge. Even here, the budget is not unlimited: the evidence must be relevant to the assessment, not to every claim that could conceivably be made about the patient.

For teaching, ground the rule. The discriminating features of the case need to be visible so the rule generalizes. Other features can be omitted or named only in passing.

The function determines how much evidence the grounded compression carries. The principle is invariant. Ground what prevents dangerous reconstruction. Omit the rest.

The clinical AI implication

A clinical AI system inherits the same constraint, and inherits both failure modes in their characteristic shapes.

The under-grounding failure for AI is the fluent summary. The output reads cleanly, asserts "low-risk" or "stable" or "anxiety," and does not anchor the claim to the evidence that would let the clinician verify or reject the reconstruction. The clinician reads "anxiety at rank one" and the system has not surfaced what features that ranking depended on. If the clinician's reconstruction of "anxiety" differs from the AI's, the divergence stays invisible.

The over-grounding failure for AI is the wall of context. The system surfaces every chart finding, every lab trend, every imaging result, every prior note, in a single dashboard. The clinician has to do the relevance triage the AI was supposed to do. The fact that the system "has access to the entire chart" becomes a liability rather than an asset when nothing is filtered.

The grounding rule for the system is the same as for the clinician. Anchor the claims that would be reconstructed wrongly without grounding. Omit the evidence that would not change the reconstruction. The system has to know which claims carry the current decision and which evidence would actually move the receiver's model.

At the bedside

The ED team that wrote version two of Mr. J's handoff knew that "low-risk" and "stable" were going to do too much work alone. They grounded each decision-driving word with the evidence that would let the floor hospitalist reconstruct the same picture. The hospitalist took the patient at 19:00 and did not call the ED. The next morning Mr. J ambulated, troponins stayed flat, and he was discharged from the floor by 10:00. The grounding did the binding work that "stable" alone could not have done.

Part 5 (already in the series as The Minimum Sufficient Message) handles what to omit when the receiver does not need it. Part 6 handles what to include when the receiver could be misled without it. Both are corollaries of the same principle: every word in the message has a reconstruction cost, and the message that costs the least without ambiguity is the one that did the job.

---

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 · 11. The Irreversible-Action Check