Quiet Acquisition

Late morning. An elderly patient with two months of weight loss is seen for follow-up. The chart records 14 visits over the past year, three labs, one chest X-ray, an outpatient cardiology consult. The AI copilot summarizes: chronic weight loss, likely multifactorial, depression on the differential, anorexia of aging plausible, malignancy possible but no specific signal on workup. The clinician reviews, agrees, schedules another follow-up. The signals in the chart are consistent and the AI's reading of them is faithful. The diagnosis, made six weeks later from an emergency visit, is metastatic pancreatic cancer. No one ever asked about early satiety, postprandial pain, or new-onset back pain. The questions were not in any prior note; the AI cannot ask questions that were never recorded.

This is the upstream half of the verification problem. Quiet Verification covered what happens after signal arrives: the architecture for surfacing mismatches between what the system sees and what the system thinks, with passive default and small active closure. There is a symmetric problem on the other side. The inferences the system holds are only as good as the signals that fed them, and the signals that did not arrive are invisible by definition. A closed-loop architecture also needs an acquisition layer that detects what has not been asked, has not been measured, has not been observed.

Two layers

The architecture sees the world in two layers, and the acquisition problem only makes sense once they are named.

Signals are facts. Observable, attestable, time-stamped, recorded. A sensor result, a clinician's typed note, an order signed at 10:14, a medication administration logged at 11:02. Signal-fact verification (does the record match what was actually sensed, said, or done) is largely automatable; the architecture can check these without spending clinician attention.

Inferences are computed states. The differential is a mixture distribution over hypothesis-components, each carrying its own evidence dependencies, predicted trajectory, and action threshold. The trajectory is a derived sequence, computed from successive measurement signals, with an eventual ground-truth signal much later. The plan is an inference about what the clinician intends, derived from their action stream. The risk estimate is an inference parameterized by the diagnostic mixture. None of these is directly observable.

Verification (the previous essay) lives at the inference layer. The hard cases are not whether the signal arrived correctly but whether the inferences drawn from those signals match across agents and against reality. Acquisition lives one layer earlier: deciding which signals to request when the mixture cannot discriminate cleanly between its components on the evidence in hand.

The three forms of acquisition failure

The signal types each have their own acquisition failure mode.

Un-asked attested signal. The history element that was never elicited. The early-satiety question never put to the weight-loss patient. The OCP question never asked of a younger woman with pleuritic chest pain. The suicidal-ideation screen never run with the postpartum mood case. The chart records what was discussed. The gap is what was not.

Un-ordered sensor signal. The lab, image, or measurement that would have moved the posterior between hypothesis-components but was never requested. The d-dimer in a low-Wells case where PE sits below the leader on probability but above on action threshold. The troponin in an atypical presentation where ACS is action-dominant despite a lower rank. The structured exam maneuver that would have separated cardiac from pulmonary causes.

Un-followed trajectory. The measurement at hour four that would have confirmed or refuted the current inference but was never set as a follow-up. The recheck of vitals after intervention. The phone call the next morning. The outpatient lab two weeks out. Each working hypothesis-component predicts a trajectory; absence of the successor signal means the system never closes the inference loop on that hypothesis.

Each of these is a gap that the verification mechanisms in the previous essay cannot catch. There is no signal to verify against. The mismatch sits in the absence.

Why friction discipline applies here too

A naive solution to acquisition is to surface every missing question, every plausible test, every reasonable follow-up. The same evidence cited in Quiet Verification (Van der Sijs 2006 on override rates, Strom 2010 on hard-stop workarounds, Wong 2021 on the Epic Sepsis Model's per-clinician alert burden) tells us this fails. A recommended-tests panel firing on every patient is alert fatigue in another costume. A required-history checklist at intake re-installs the click-through reflex. The acquisition layer also has to be passive by default.

The passive substrate

The same three channels that carried verification in the previous essay carry acquisition with different semantics.

Ambient display surfaces the under-determination in the mixture. A peripheral indicator on each action-dominant hypothesis-component shows how confident the system is and which discriminating signal is currently missing. Wallpaper for the common case; brighten when an action-dominant component has high uncertainty and a discriminating signal is acquirable. The clinician's foveal attention is not claimed; the indicator sits in peripheral vision until something changes.

Behavioral inference reads the clinician's action stream for what they chose not to do. Non-orders, non-questions, non-follow-ups are partly inferable from the actions they did take. If the clinician orders a CT pulmonary angiogram, the system registers that PE has crossed the action threshold and is being worked up. If the clinician schedules no follow-up after an atypical presentation, the system registers the trajectory as un-tracked for the hypothesis-components that would have demanded one. The action stream's negative space is a signal about which acquisition gaps the clinician has implicitly chosen to leave open.

Action-precondition checks at the workflow commit. When the discharge button is pressed, the sign-off is finalized, the disposition is set: the system gates against a missing discriminating signal if the gap would change the call. Not as a standing alert during the encounter. Not as a modal in the middle of the workup. At the commit point, where the cost asymmetry favors a brief pause and the gate has a specific predicate to check.

These three channels carry acquisition without claiming attention until a gap crosses the action-relevance threshold.

What the floor cannot cover

Each passive channel has a documented failure mode, and the dangerous shared failure is symmetric with the inaction divergence Quiet Verification named.

Ambient display fails when the peripheral indicator is for a hypothesis-component the clinician has stopped considering. The cue is in the visual field; the attentional set is not.

Behavioral inference fails when the missing signal is for a hypothesis no one has entertained. There is no behavioral footprint of an unconsidered diagnosis. The action stream registers agreement with the leader precisely because no alternative is being worked up.

Action-precondition checks fail when the missing signal is for a discriminator no one anticipated. The gate map covers the predicates the designers thought of. The dangerous miss involves a predicate not in the map.

The deeper structural failure: the system cannot ask about a hypothesis it does not have. The mixture only contains the components someone put in. The dangerous miss is the diagnosis missing from the mixture entirely, which is the symmetric counterpart of inaction divergence. In verification, the clinician does not see, does not act, does not trigger anything. In acquisition, the system does not consider, does not surface, does not ask.

The closure

The closure mechanism mirrors Quiet Verification's structured read-back. Narrow, AI-targeted, at workflow seams. Pre-disposition or pre-sign-off, the system surfaces one or two missing discriminating signals (questions, tests, follow-ups) that would change the call between action-dominant hypothesis-components.

The features that distinguished the WHO Surgical Safety Checklist from the failed modal alerts apply here as well. Bolted to existing workflow boundaries the team has to pause at anyway. Distributed across the team where possible. Small enough to fit in working memory. AI-targeted because the slots that warrant friction are not enumerable in advance; only the divergence trace across the mixture's components knows which gap matters for this patient on this shift.

The active closure has one job: catch the residual that the three passive channels cannot. The friction budget is reserved for that and nothing else.

The architecture, restated

Verification and acquisition are two operations on the same signal substrate, governed by the same friction discipline, running in opposite temporal directions.

Verification: signal arrives, compare to inference, flag the mismatch or update the belief.

Acquisition: inference under-determined, identify the discriminating signal, request it at the workflow seam or wait for it to arrive passively.

Together they close the loop. Signals feed the mixture. The mixture is verified against new signal and against the clinician's parallel mixture. Gaps in the mixture surface the next signals to acquire. The next signals refine the mixture. Both halves run mostly invisibly. Both halves spend friction only when the friction is the only instrument available.

At the bedside

The elderly patient with weight loss is not coming back in time to ask the question that was never put. The clinician at 3 a.m. is not going to ask twenty more questions, click through twenty more recommended-tests panels, or set twenty more follow-up timers. The acquisition layer that survives this is the one whose default state is silent. The gaps surface only where a different next question would have changed the disposition, only at the moment the disposition is being made, and only for the hypothesis-components the system can name.

The other half of the friction budget is for that closure. The rest of the acquisition has to be quiet.

---

Compressed Medicine · 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

Companions to Part 10 · Quiet Verification · Quiet Acquisition