May 23, 2026 · 18 min read

Compressed Coordination

Companion to Compressed Medicine · The general theory the clinical curriculum instantiates

By Sunny Harris, MD

A bounded agent — finite sensors, finite memory, finite attention, finite time — cannot hold the world. The constraint is simple. Let N be the cardinality of distinguishable world states the agent might have to represent; let M be the cardinality of internal states the agent's machinery actually distinguishes. M < N. Treat this as compact notation for the qualitative constraint, not as load-bearing formalism — the inequality names what follows from being a bounded agent in a richer world, and that's all it needs to do. Reality has, for practical purposes, inexhaustible detail; the agent has a few pounds of brain or a few terabytes of RAM. The math does not work; more hardware does not eliminate the need for compression.

Compression is what the inequality forces. The agent's representation must collapse multiple distinct world states onto a single internal state — keep what changes action, drop the rest. A single bounded agent acting alone compresses, acts, observes, updates, compresses again. The hard problem starts when two such agents have to act on a world they're each compressing separately.

The constraint

When two bounded agents must coordinate, neither can hand the other their full representation. There is no full representation; both ends are already operating under M < N. The handoff itself is constrained by the same inequality — the channel admits some number of distinguishable signals smaller than the sender's already-compressed internal state, which is in turn smaller than N. The signal that crosses is a compression of a compression.

The consequence is mechanical: the signal underdetermines the source state. If the sender has N possible source states and the channel admits M < N possible signals, then some source states share a signal. The receiver, looking at the signal, cannot uniquely identify which source state it came from. Multiple source states are consistent with any given signal; the receiver cannot recover the sender's full state from the signal alone.

This is not a flaw of any particular codec or any particular matching mechanism. It is a structural property of the inequality. The receiver always does inferential work the signal does not fix.

The frictionless-vacuum check. Imagine for a moment that M ≥ N along the resource axes — an agent with infinite memory, infinite bandwidth, infinite compute, infinite attention. The bounded-resources inequality dissolves. Compression along those axes is no longer mandatory; the agent can ship its complete representation; the receiver can hold the sender's full state without remainder from that source. Under M ≥ N, the bounded-pathway contributions to the three failure modes vanish — but the failure modes themselves can still arise from non-bound pathways: model heterogeneity from differing learning histories, channel reliability and SNR issues, incentive asymmetries, and multi-agent common-knowledge gaps. The check isolates boundedness as a causal axis, not the causal axis. Boundedness is one independent variable; the framework has others, and they don't dissolve when the inequality does.

Bounds have shape. The inequality is universal; its character varies. A memory-bounded system (the brain's working capacity, the model's context window) compresses differently from a bandwidth-bounded one (the wire between two services, the seconds available at sign-out). A temporally-bounded system (the resident with thirty seconds to deliver a consult) compresses differently from a vocabulary-bounded one (the receiver lacking the codec the message points at). The strategy follows the binding constraint. The same agent under different bounds produces different lossy signals about the same source state, and where the strongest bind lies determines which failure modes show up first.

One axis among several. Boundedness is the axis this essay develops most carefully, because it's where the M < N derivation lands cleanly and where most of the curriculum's failure modes connect. But coordination failure has at least five other axes that produce distinct failure pressure independently of bounded resources: learning history (different training data, different priors, different representational capacity — agents with identical runtime resources can diverge because they learned different partitions from different data); substrate heterogeneity (different sensor modalities, effector types, computational architectures — what an agent can represent at all); channel properties beyond raw bandwidth (reliability, SNR, latency, authentication); incentive structure (utility alignment, trust, fault model); and system-level structure in multi-agent settings (synchrony, common-knowledge availability, network topology). The companion essay What Determines Whether Coordination Survives develops the full multi-axis framework with falsifiability conditions for each axis. The rest of this essay treats the bounded-resources axis as the canonical case — the failure modes named here are produced through that axis, even when they also have pathways through the others.

The consequence

If the source state cannot be recovered uniquely, then communication is not the transfer of state. It is the receiver-side reconstruction of state, from the signal plus the receiver's own priors plus the surrounding context.

This sounds like a problem. It would be a problem if coordination required full-state recovery. It doesn't. Coordination requires that the receiver's reconstruction preserves the distinctions that would change the next action.

State the criterion sharply: the task partitions source states into equivalence classes. Two source states belong to the same class if the optimal next action under each is identical. Communication need not preserve every difference between source states; it need only preserve the differences that move source states across class boundaries. This is the rate-distortion move, recast for coordination — a compressed signal is task-sufficient if it preserves the distinctions that would change the receiver's policy. Underdetermined dimensions that fall inside an equivalence class are harmless. Underdetermined dimensions that span class boundaries are the dangerous ones; the receiver's reconstruction might land in the wrong class and drive the wrong action.

With one caveat that does most of the work. The task assumes a shared objective — both agents partitioning source states by the same criterion. When two agents coordinate on a single common goal, the partition is shared and "task-relevant distinction" is unambiguous. When two agents have overlapping but distinct utility functions — which is most real institutional communication — each agent partitions source states by their own utility, and the partitions disagree. A clinical AI optimizing for thirty-day readmission rate draws sharp class boundaries between 95% and 99% stable; a throughput-pressured administrator collapses both states into the same discharge class. Same signal, same source state, different partitions, both internally coherent under their own utility function. The "right" reconstruction is no longer a property of the message; it is a property of whose utility function defines the partition.

The framework does not collapse under this. It widens. There is not one task partition; there are partitions, one per agent, with their boundaries dictated by each agent's payoff structure. Coordination across agents with identical partitions is the easy case (rate-distortion under shared loss). Coordination across agents with different partitions is the hard case, and most of clinical practice, institutional communication, and multi-agent AI lives there.

This does not eliminate underdetermination. It defines when underdetermination is harmless (inside the relevant partition) and surfaces a new failure mode the next section names. The whole pragmatic content of compressed coordination sits in this one move.

This is not, and is not pretending to be, a philosophical settlement of the indeterminacy-of-meaning problem. It is a task-relative engineering criterion. The receiver's reconstruction is not "the correct meaning" of the signal; it is a representation sufficient for the next action, under a partition that depends on whose utility function is being optimized.

The mechanism

Reconstruction is not a single operation. It runs through several stages, each with its own failure surface.

Common ground. The signal only constrains a reconstruction in the receiver if the receiver has stored patterns the signal can resolve to. A handle like DKA only lands in a receiver who has the illness script installed. Sell at the bid only resolves into a market action in someone fluent in equities. git rebase only resolves into a meaningful operation in someone with a built model of distributed version control. The shared infrastructure — vocabulary, conventions, priors — is what every signal is implicitly a pointer into. Both ends must have invested in the same infrastructure for the pointer to resolve.

Matching. When the signal arrives, the receiver runs a matching operation: from the signal, against stored patterns, to a candidate reconstruction. The matching is fuzzy by construction. Multiple surface forms point at the same pattern; single surface forms point at multiple patterns; matching fires probabilistically with context as the disambiguator. Matching is also multi-level — a face matches simultaneously on visual features, identity continuity, behavioral history, emotional signature, and each level can succeed or fail independently.

Priors and context. Matching does not run on the signal in isolation. It runs on the signal weighted by everything the receiver already believes — the leading hypothesis, the surrounding signals, the recent action history. The reconstruction is the joint optimum of signal and priors, not the signal alone. When priors are strong, they can resolve underdetermination cleanly. When priors are wrong, they can drag matching toward a confident but incorrect reconstruction.

Verification. For coordination to actually close, the sender needs to know the receiver's reconstruction landed in the right equivalence class. If verification is implicit — silence treated as agreement — the sender assumes the landing was correct. If it's explicit (read-back, confirmation, observable downstream action), the loop closes. If it's neither, divergence accumulates silently.

Temporal stability. All of the above must remain stable across time. Common ground drifts (training distributions shift; conventions change). Priors update (the receiver yesterday is not the receiver tomorrow). Verification mechanisms degrade. The architecture is not a static contract; it's a maintained one.

The breakdowns

If reconstruction runs through five stages, it can break at any of them, producing four broad classes of failure. Three are accidents of bounded reconstruction. The fourth is design under misaligned utility, and it is the framework's most original move.

Failure to resolve. The signal arrives but cannot be matched. The receiver has no stored pattern that fits; or multiple stored patterns fit equally well and context doesn't disambiguate. The receiver knows they don't know. This is the safest failure because it surfaces openly: the receiver asks for clarification, the sender provides more signal, the loop continues.

Wrong resolution. The signal matches a stored pattern, but not the one the sender meant. This is the dangerous family. It happens when:

Resolution drift. Matching that worked at one time fails at another. Common ground has aged. Conventions shifted. Training distributions diverged from current cases. The same signal that landed correctly six months ago lands in a different equivalence class today.

Strategic underdetermination. A fourth class sits adjacent to wrong resolution, and only becomes visible once the consequence section's caveat is taken seriously. Strategic underdetermination is the case where the sender exploits the receiver's reconstruction process to land the receiver in an equivalence class the sender prefers — under the sender's utility function, not a shared task. A transferring physician compresses a sign-out ambiguously to push the receiving service into the "must-accept" class. A vendor's release notes describe a security-relevant change at a level of abstraction that lands customers in the "no action needed" class even when their own utility function would have required action. A regulator's filing answers the literal question without surfacing the discriminating signal that would have flipped the supervisor's decision. Strategic underdetermination is coordination success under the sender's utility (the receiver lands where the sender wanted them to land) and coordination failure under any utility function the receiver would have chosen for themselves. The other three breakdown classes treat wrong landings as accidents; strategic underdetermination treats wrong landings as design. The defenses that work against accidental wrong resolution — more shared codec, more verification — do not work against strategic underdetermination, because the sender controls the very signals the receiver would verify against.

The canonical example of wrong resolution at multiple feature levels is entity resolution. Two records share name, date of birth, demographics, and most surface features; they get merged because the surface match succeeds. The identity binding fails — they are not the same person. Category-level matching landed; identity-level matching didn't. The downstream action — one person's medication history applied to the other's care — is catastrophically wrong.

The same shape recurs across substrates. In distributed systems: a client resolves the correct service identifier to a stale leader after failover; name-level matching succeeds, temporal-state matching fails, the action executes against the wrong epoch. In retrieval-augmented AI: a model retrieves the semantically nearest document, which is the wrong patient, the wrong environment, or the wrong time slice; topic-level matching succeeds, referent-level matching fails. In human cognition: the Capgras phenomenon, where visual features match the loved one but identity continuity does not and the integration mechanism that normally settles the conflict has broken. Different substrates, same architecture.

A variable-by-variable mapping makes the structural analogy explicit. In clinical entity resolution: a surface handle (name + date of birth + demographics) activates the wrong referent (Patient A's record retrieved for Patient B), because the temporal/causal binding that would have distinguished them (actual care-history continuity, provenance per assertion) was absent or unindexed. In Knight Capital 2012: a surface handle (the Power Peg flag name) activated the wrong referent (the retired code path on one server), because the deploy-version binding that would have distinguished old code from new was inconsistent across the fleet. Same shape, substituted variable-for-variable: surface match succeeds at the level the channel observes; identity match fails at a level the channel doesn't carry; the missing variable is provenance — under what state was this binding made. The lesson transfers structurally: high-stakes systems that resolve handles to referents need an explicit provenance dimension or they will run this failure mode, regardless of substrate.

The three failure modes the Compressed Medicine curriculum named at the clinical level are now expressible within this framework rather than asserted from outside it:

All three are special cases of the same underlying breakdown: the receiver's reconstruction landed in the wrong action-relevant equivalence class, and the channel could not surface the divergence in time to act on it.

Strategic underdetermination sits outside this synthesis. The other three failure modes are versions of the same accident — reconstruction landing in the wrong class against the shared task. Strategic underdetermination is what happens when there is no shared task to begin with, and the sender's utility function is satisfied by landing the receiver in the wrong class for the receiver. The first three are accident; the fourth is design.

The countermeasures

If the failures are universal, the defenses are too. Three operations recur across every domain that has solved any version of this problem.

Shrink the harmful underdetermination space. Invest in common ground. Train both sides on the same patterns, the same conventions, the same priors. The more two agents share, the more the signal constrains reconstruction to a narrow set of action-relevant equivalence classes. Medicine has done this for a century with medical school, residency, illness scripts, and standardized order sets. Aviation does it with Crew Resource Management and ICAO phraseology. Distributed systems do it with shared schemas and ontologies. Science does it with notation conventions and citation infrastructure. The investment is heavy upfront; the return is that subsequent signals constrain reconstruction more tightly and land in the right class more reliably.

Detect and surface wrong landings. Even with strong common ground, reconstruction can land in the wrong class. The second defense is mechanisms that catch this — feedback from receiver to sender, divergence detection across replicas, predicted-versus-observed checks, action-precondition gates, structured read-back at high-stakes commits. The mechanisms differ across domains; the operation is the same: surface divergence before action makes it irreversible.

Align partitions, or surface their misalignment. A third defense is implied by the utility-function caveat: when each agent's partition is defined by their own utility, sharing a codec is necessary but not sufficient. Two agents can have perfect common ground at the vocabulary level and still partition source states differently because their payoffs differ. The defense is to either align the utility functions explicitly (rare in practice; most institutional incentives are set by structures the agents don't control) or to make the misalignment visible before action commits. The irreversible-action check from the clinical curriculum's capstone is not just error correction. It is a continuous alignment protocol: a forced moment where the team's distinct utility functions get reconciled before action commits. Without it, two agents can coordinate fluently inside their respective partitions and still drive opposite actions on the same source state, neither noticing, because each reconstruction was internally correct.

Every other defense — uncertainty markers, decompression order, grounding constraints, structured handoff protocols, cross-modal verification — is an implementation detail of these three. Shrink the space, catch what slips through, surface the partition mismatches. That is the whole architecture, at any substrate.

Why this generalizes, and why now

The architecture is substrate-independent because the constraints producing it are substrate-independent. Any system with bounded agents, lossy channels, and a task that partitions source states into action-relevant equivalence classes runs this architecture. The alternative is incoherent action.

Clinical medicine. The case worked out in detail in the eleven-part Compressed Medicine curriculum. Bounded clinicians compress patients into structured belief states; handoffs transmit a compression-of-a-compression; verification mechanisms at action-changing moments catch divergence before discharge, sign-off, or escalation. Medicine forces the architecture to be honest because the stakes are visible, the agents are heterogeneous, the codecs are documented, and irreversibility is tracked. It is one instantiation, not the boundary.

Multi-modal AI ingestion. A model that ingests video, audio, structured records, internet content, and outputs of other models is receiving multiple compressions of underlying reality, each with its own codec — a video encoder, an audio codec, a database schema, a training distribution, an ontology. The model must run matching against each, reconcile across modalities, hold a unified reconstruction, and act. The clinical AI thesis lifts directly: belief-state coordination across heterogeneous channels, verification at action-changing moments, common-ground investment across modalities. Same architecture, wider substrate.

AI-to-AI coordination. When both ends of a handoff are automated, no human is in the loop to catch silent reconstruction. The three failure modes recur faster and at scale. The "small active read-back at workflow seams" from the curriculum's Quiet Verification companion becomes a cross-AI handshake protocol that does not yet exist; the field has not built it because the human-mediated case has dominated attention.

Distributed systems. Lamport's 1978 logical-clocks paper named the structural problem: mutual knowledge of meaning is unreachable over an imperfect channel between two finite nodes. Knight Capital's 2012 Power Peg outage was wrong resolution at scale — eight trading servers, one missing the new code, all matching at the protocol level while diverging at the semantic level; $440 million lost in forty-five minutes. Vector clocks are the partial mitigation; same architecture as clinical AI's provenance-per-assertion discipline.

Intelligence analysis. Yom Kippur 1973. A standing doctrine — Egypt would not attack without air parity — biased matching toward an incumbent hypothesis; mobilization signals were resolved into the wrong equivalence class. Richards Heuer's Analysis of Competing Hypotheses (1999) is the procedural mitigation: forced scoring of evidence against every hypothesis on the table, not only the leader.

Aviation. Tenerife 1977 was wrong resolution at the matching stage: a clipped controller transmission, KLM's matching landed on "clearance," the underdetermined dimension was action-relevant, the integration mechanism (read-back) was absent. Crew Resource Management and mandatory read-back are the structural mitigations that close the loop.

Biology. Cells signal compressed states to other cells through receptor-ligand pairs — handles built from molecular codecs. The immune system's failure modes look like codec failures: autoimmunity (self codecs misclassified as foreign), immune escape (pathogens evolved to evade the codec), allergic responses (the codec is too generous). Related signaling problems, at radically different scale.

Science. A paper compresses years of work into a few thousand words. Peer review is constrained reconstruction with verification. Reproducibility crises are silent-reconstruction failures at the field scale, where receivers resolve papers against priors that don't match what the original work actually established.

Cross-cultural communication. Idioms are compressed cultural patterns. Translation fails not because words lack equivalents but because receivers do — the underdetermined dimensions of a phrase span equivalence classes the source language doesn't make explicit. The architecture from the Receiver Is the Codec companion, stated outside the clinical case.

Different stakes, different speeds, different agents, same architecture.

What sits outside the framework. Naming where the theory doesn't apply makes the claim more precise. Quantum entanglement has no bounded channel between entangled particles, no codec, no resolution operation — compressed coordination simply doesn't engage. Stateless synchronized microservices, where each request is independent and no state is carried between them, have no source state to underdetermine; the framework is empty in that limit. Deterministic single-agent computation has no second agent and no channel; nothing to coordinate. The bounded-resources axis this essay develops applies wherever three conditions hold simultaneously: (a) at least two agents are bounded (M < N), (b) some state must be carried across a lossy channel between them, and (c) a task partitions source states into equivalence classes — even if the partitions disagree across agents. Outside those three conditions, this axis has nothing to say (other axes may still bind — that is the broader framework's territory). Inside them, the architecture has the same things to say about every substrate.

Why this matters now. Three converging pressures make the framework load-bearing rather than a curiosity.

AI systems are becoming multi-modal. Each stream is a codec; reconciling them is compressed coordination at industrial scale. Building this without the framework means re-inventing the breakdowns domain by domain, painfully, under different names.

AI systems are becoming multi-agent. Agentic frameworks, multi-step tool use, AI-to-AI handoffs. The human verification clinical AI still depends on isn't available when both ends are automated.

Every system integration is a codec-translation operation. Cross-organizational data flows, schema migrations, API integrations, EHR-to-EHR transfers — all instantiate the architecture, with the same four breakdown classes and the same three defenses. The cost of misalignment scales with the irreversibility of the resulting action.

The framework is a useful lens. Without it, each domain rediscovers the same breakdowns and installs partial, incompatible mitigations. With it, the mitigations become more portable: aviation's read-back discipline can lift to clinical AI; clinical AI's structured belief states can lift to multi-modal model design; distributed systems' provenance can lift to AI-to-AI protocol.

The clinical curriculum was the first place this architecture got stress-tested in production at high stakes. The architecture lives one level up. The multi-axis framework develops it further; the substrate-specific instantiations beyond clinical are the next stress tests still ahead.

A standalone companion to the Compressed Medicine curriculum. Related: What Determines Whether Coordination Survives · The Receiver Is the Codec · The Algebra of Thought.