From Stable Addresses to Earned Dispositions

Anyone who has worked a prior authorization queue knows the moment.

The case is incomplete. Technically.

But you can see what is actually happening.

The evidence is there, three pages into a discharge summary. The denial reason has an exception path nobody pulled. The member is adherence-fragile in a way no rule will catch. The criterion is satisfied if you read the policy the way it was meant to operate, not the way it was flattened into a checkbox.

You see the disposition.

The system sees the description.

That gap shows up everywhere in managed care. A care manager reads a draft care plan and sees the intervention will not land. A medical director recognizes the appeal pattern before the file is complete. A fraud investigator sees which evidence will matter before the case is built. A pharmacist reads a refill history and feels the cost sensitivity before any score says it out loud.

Healthcare expertise lives in that recognition.

Healthcare infrastructure mostly does not.

This essay is about the missing substrate underneath that gap.

Not a dashboard. Not an agent pitch. Not a brand doctrine.

A substrate.

The claim is simple, and it has to be built in the right order:

First, healthcare work needs stable addresses for its recurring primitives.

Then, and only then, those primitives can accumulate dispositions through receipts and outcomes.

Do not reverse that order.

That was the mistake.

The honest split

There are two claims that want to fuse.

They should not.

The first is the canonicalization claim.

Healthcare workflows contain recurring primitives. Problems. Goals. Interventions. Coverage criteria. Evidence requirements. Denial reasons. Appeal pathways. Suspect conditions. Adherence barriers. Investigation rationales.

These primitives repeat, but they are locally named. They appear as row labels, note fragments, policy clauses, dropdown options, task reasons, denial explanations, appeal templates, internal shorthand. They carry operational knowledge, but they do not carry stable identity.

The canonicalization claim says: give the recurring primitive a stable address so it can be reused, compared, governed, audited, and improved.

That claim is demonstrable now.

The second is the dispositional claim.

A primitive does not merely exist. It may carry a tendency. This intervention may be engagement-warm for one population and weak for another. This denial reason may be appeal-reversible when a specific evidence object is added. This medication regimen may be adherence-fragile under cost pressure. This care-plan signature may be closure-ready after one outreach.

The dispositional claim says: once a primitive is addressable, receipts and outcomes can reveal what that primitive tends to do under specified conditions.

That claim is not fully proven by the primitive library itself.

It has to be earned.

Canonicalization first. Dispositions later. Receipts in between.

What a PGI library proves

A care management PGI library is useful because it does not prove too much.

That sounds like a weakness.

It is not.

It proves the first claim cleanly.

A care management library with tens of thousands of rows collapses to thousands of distinct problem-goal-intervention triples. That is structural redundancy. Inside it are hundreds of problems, goals and interventions. One intervention, “care manager will follow up” recurs across dozens of distinct problems.

That is not a philosophical argument.

That is trapped action potential.

The same intervention keeps reappearing under different local contexts. The organization is already reusing primitives. It just does not have enough addressability to know what it is reusing.

So the system cannot ask the questions it should be able to ask:

Where else did this intervention appear?

Which problems did it attach to?

Which goals did it support?

Which care managers edited it?

Which variants should be deprecated?

Which populations responded?

Which outcomes followed?

The work has memory, but not operational memory.

It has repetition but not compounding.

That is the canonicalization claim.

Proven enough to build on.

What the PGI library does not prove

The PGI library does not prove the dispositional claim.

Not yet.

It contains rows. It contains problems, goals, and interventions. It contains repeated structures. It does not contain conditions, triggers, receipts, or outcomes.

That means it cannot yet tell us whether “Care manager will follow up” is effective, for whom, under what conditions, through which channel, after which prior intervention, with what failure mode, and against which outcome.

It can tell us the primitive recurs.

It cannot yet tell us what the primitive tends to do.

This distinction matters because it keeps the framework honest.

A stable address is not an empirical claim.

A disposition is.

A canonical ID says: let us treat these as the same concept for operational purposes.

A disposition says: this concept tends toward X under condition Y.

The first can be governed as reference.

The second has to be tested against the world.

If you fuse them, you get ontology theater.

Clean names pretending to be knowledge.

Description is not action

Most healthcare AI systems still behave as descriptive engines.

They extract what exists in the record: diagnoses, medications, notes, claims, policies, encounter histories, codes, call transcripts, faxes.

Then they represent it in a cleaner form.

Useful.

Often very useful.

Still not enough.

A care manager looking at a complete summary still has to decide who to call, what to say, which barrier is real, which intervention belongs, which risk is urgent, which gap can close, which escalation is worth spending mental bandwidth on.

A prior authorization specialist can find the policy and still has to translate it into evidence requirements that match a mess chart note.

A pharmacist can identify an adherence issue and still has to decide whether the member is cost-sensitive, refill-fragile, switch-ready, outreach-responsive, confused, embarrassed, angry, or just done with the whole insurance thing.

The descriptive layer leaves the action layer underspecified.

A description says what is there.

An action-bearing truth says what is likely to matter next.

A member is 73 years old.

Descriptive.

A member is fall-prone under polypharmacy conditions.

Action-bearing.

A claim has a CPT code.

Descriptive.

A claim is audit-vulnerable under a known outlier pattern.

Action-bearing.

A care plan contains a diabetes problem, an A1C goal, and an adherence intervention.

Descriptive.

That problem-goal-intervention bundle is engagement-warm for SMS-reachable members but weak for members with transportation barriers and uncontrolled food insecurity.

Action-bearing.

Healthcare work does not run on status alone.

It runs on conditional action potential.

What is a disposition?

A disposition is a real property that manifests under certain conditions.

Fragility is a disposition. Glass is fragile because, when struck with sufficient force, it tends to break. The glass is still fragile when nobody strikes it.

Solubility is a disposition. Salt is soluble because, when placed in water, it tends to dissolve. Salt sitting dry on a shelf has not lost the property of solubility in water (under generally common conditions).

Healthcare is full of these properties.

A member may be readmission-prone under transition stress. It’s why we pay ToC premiums and incentivize risk managers to prevent re-hospitalizations.

A medication regimen may be adherence-fragile under cost pressure. It’s why GoodRx and direct to consumer has gained considerable mindshare

A care gap may be closure-ready after one clean outreach. It’s why ranked queues are value capture optimizers.

A denial may be appeal-reversible if the missing documentation is supplied. This is request for information 101. What if we had access without requesting?

A provider pattern may be suspicious under peer comparison. Malignment to best practice is identifiable when we have complete provider pictures.

A coverage criterion may be evidence-satisfied given the right lab value and diagnosis history.

A disposition is not the same thing as a prediction.

A prediction says: this will happen.

A disposition says: this is the kind of thing that tends to happen under these conditions.

That difference matters.

A member can be readmission-prone and not be readmitted because a transition-of-care intervention worked. The disposition was not falsified. It was managed.

That is why dispositional reasoning matters for healthcare operations. It asks a better set of questions:

What is this case inclined toward?

What would trigger the disposition?

Which action targets the disposition itself?

Which action targets the trigger?

Which receipt would prove whether the action mattered?

Prediction wants the answer.

Disposition wants the action surface.

The substrate

A dispositional ontology needs a substrate.

You cannot attach durable dispositions to loose free text.

“Weekly medication adherence outreach,” “call member weekly about meds,” “CM adherence follow-up every Monday,” and “weekly med compliance call” may be the same operational primitive.

Or not.

The difference matters.

If each instance stays trapped in local wording, the system cannot learn across them. If the system collapses them carelessly, it erases distinctions that may matter.

Canonical workflow intelligence is the disciplined middle.

It decomposes healthcare workflows into stable, addressable, governed primitives. Each primitive receives a canonical identity. Local instances bind to that identity with evidence and governance state. Receipts record what was proposed, committed, edited, approved, rejected, executed, or measured. Outcomes attach back over time.

The hash is not the product.

The hash is the address.

The product is what addressability unlocks: retrieval, reuse, comparison, governance, auditability, similar-case search, AI grounding, outcome learning, and operational memory.

This is the foundation of Emergent Portability. Here, the only point is the substrate: the primitive needs an address before it can carry history.

A care plan stops being a document.

It becomes a composition of addressable objects.

Problem.

Goal.

Intervention.

Binding.

Signature.

Receipt.

Identity and relationship stay separate.

That matters. “Weekly adherence outreach by care manager” may be the same intervention across several goals. Binding it to A1C improvement, medication possession ratio improvement, or hospitalization reduction creates different relationships. The relationship changes. The primitive identity should not disappear every time the relationship changes.

This is the move from local documentation to operational knowledge.

Reference truth and empirical truth

Ground truth is not one thing.

There is referential ground truth: the address. The canonical ID. The hash. The governed decision that says, for operational purposes, these things co-refer.

There is empirical ground truth: the receipt. What actually happened. Who proposed it, who approved it, what evidence supported it, what changed, what outcome followed.

The disposition sits between them.

Addressed by the ID.

Tested by the receipt.

Refined by the outcome.

This is why objective versus subjective is the wrong cut.

Canonicalization is not objective in that sense. It is stipulated, governed, nominal. It can be well-formed or malformed. It can be useful or violent. It can preserve meaning or erase it.

But a disposition is empirical.

If we claim that a denial reason is appeal-reversible when a lab value is supplied, the receipts can test that. If we claim an intervention is engagement-supporting for a population, outcomes can test that.

The identity decision gives the claim a place to live.

The receipt tells us whether the claim survived contact.

Propose and commit

This architecture does not require pretending AI owns truth.

That would be dangerous.

The safer pattern is simple:

LLMs propose.

Governed, human-led systems commit.

A model may extract a candidate intervention from a note. It may offer evidence, confidence, and rationale.

But the binding must carry governance state.

Was it auto bound under a validated rule?

Was it proposed by a model?

Was it approved by a clinician?

Was it overridden by a care manager?

Was it deprecated by a curator?

Was it rejected because the apparent match was too broad?

The system should know the difference between proposed truth and committed truth.

Without that distinction, canonicalization becomes another opaque AI assertion.

With it, canonicalization becomes a governed resolution pipeline:

extract candidate, normalize, search registry, propose match, attach evidence, score confidence, route to rule or human approval, commit binding, generate receipt, measure what happened, refine the model.

AI can participate in building operational truth without pretending to own it.

Receipts earn dispositions

Receipts are where the dispositional claim becomes honest.

A receipt is the structured record of what happened to a primitive, binding, signature, or action. It records who or what proposed it, what evidence supported it, whether it was accepted, modified, rejected, superseded, executed, or measured, and what happened next.

If we say a care-plan signature is engagement-warm, receipts tell us whether engagement occurred.

If we say a denial reason is appeal-reversible when a specific evidence object is added, receipts tell us how often that was true.

If we say a member is closure-ready after one hypertension outreach, receipts tell us whether the gap closed.

If we say an intervention is adherence-supporting under cost pressure, receipts tell us whether adherence improved, whether the intervention was completed, whether the member responded, and whether the care manager changed the plan.

Receipts turn dispositions into measurable hypotheses.

Without receipts, the system cannot learn what happened.

It can only generate more output.

With receipts, every action becomes a data point in the refinement of the operational ontology.

The Information Action Potential Series

Information also has a motion pattern: it enters a workflow, distributes through systems and people, gets metabolized into decisions and tasks, and either disappears or becomes memory.

The Information Action Potential Series names the gates where that motion becomes usable:

Structure: observed, parsed, canonicalized, bound, composed.

Memory: receipted, compared.

Action: retrieved, acted upon, measured.

Recursion: compounded.

Each stage adds action potential.

Each gate costs engineering work.

Production systems often have pieces of the sequence.

Observed-to-parsed is extraction.

Parsed-to-canonicalized is normalization.

Canonicalized-to-bound is governance.

Bound-to-composed is workflow intelligence.

Composed-to-receipted is operational memory.

Receipted-to-measured is empirical accountability.

Measured-to-compounded is learning.

Most healthcare AI investment has crowded the early gates:

extraction, summarization, drafting, search.

The trapped value is in the middle and the end:

bound, composed, receipted, measured, compounded.

That is where action potential accumulates.

The human role

This does not remove humans from healthcare work.

It changes where human judgment concentrates.

Humans should not spend endless hours re-finding the same policy source, re-checking the same criterion, retyping the same intervention, reclassifying the same denial reason, rediscovering the same care gap pathway, or reconstructing provenance after the fact when the information should already be there.

That is not where expertise lives.

That is where expertise gets spent down.

Humans should own accountability, empathy, ambiguity, ethics, exception handling, clinical nuance, and trust.

The medical director’s judgment on the edge case.

The care manager’s read about whether this member will pick up the phone.

The pharmacist’s instinct about which adherence barrier is real.

The investigator’s distinction between legitimate fraud and erring in good faith.

Canonical workflow intelligence compresses the mechanical substrate: extraction, matching, lookup, comparison, routing, templating, evidence assembly, version checking, audit trail generation.

Compression is not replacement.

It is labor reallocation.

The system proposes. The governed workflow commits. The human remains accountable. The receipt remembers.

Closing

Action-bearing truths are not predictions.

They are not mere classifications.

They are claims about what something is inclined to do or become under specified conditions, attached to stable operational primitives, tested through receipts, and refined through outcomes.

The first move is humble: give the recurring primitive an address.

The second is disciplined: bind local instances to that address without erasing meaningful difference.

The third is empirical: record what happened.

The fourth is dispositional: learn what tends to happen under conditions.

The fifth is recursive: let the next decision start from a richer substrate.

We have built many libraries of description.

We need registries of action-bearing truths.

Not because every action should be automated.

Because every action should be able to teach the system what happened.

Healthcare does not only need more intelligence.

It needs knowledge that can travel, reconnect, be challenged, be acted on, and remember its consequences.

It needs truths that bear action.