Essay · agentic systems · part II

Adoptable by Design

The first essay argued the constrained expert system is a good idea and a buildable one. But the systems it descends from — Arden, GLIF, Drools — didn't fail on the tech. They failed on the third question: will the real, regulated, risk-averse world actually let it in? Here's why this design is unusually well-placed to survive that — and where it still might not.

the third question: adoptable? the CDS exemption audit trail as a shield why the barriers are lower now
The gap the first essay left

Innovative and buildable ≠ adoptable

An idea can be genuinely new, and genuinely buildable, and still die — because it isn't legal to use, nobody will pay for it, or the liability terrifies everyone. That third gate is where the ancestors of this idea actually fell.

Arden Syntax and GLIF worked. Drools worked. They didn't collapse because the rules engine was broken; they collapsed because integrating them cost a fortune per site, keeping them current was nobody's job, and no clinician wanted to be the one who trusted the machine. So a serious case has to clear the gate the first essay skipped.

Innovative?✓ essay I Buildable?✓ essay I Adoptable?← the predecessors died here shipped the same idea has to pass all three gates
Two gates cleared, one to go. The tech was never the thing that buried Arden or Drools. This essay is about the red gate.
Regulation

Is it even allowed?

Software that helps a clinician make a decision can, in principle, count as a medical device — which means a slow, costly approval path. That alone has scared off plenty of good ideas.

But there's a well-known carve-out. In the US, following the 21st Century Cures Act, the FDA's Clinical Decision Support guidance treats certain CDS software as non-device — roughly, when it:

criteria (roughly) our design shows the basis for its advice the clinician can review & override not for a time-critical emergency every rule cites its guideline passage human-gated by construction (approve / edit) an advisory checker, not an autonomous actor
The exemption criteria read like our design spec. Transparent, overridable, advisory — the exact properties the "constrain more, keep the human in charge" discipline produces.

The same discipline that makes the system trustworthy is what makes it lighter to regulate. "Constrain more" pays off twice — once in engineering, once at the regulator's door.

◆ Not legal advice — and this is the honest part
The lines above are the shape of the argument, not a clearance. The real criteria have nuance, they've shifted over time, they differ by jurisdiction, and some uses (autonomous, high-risk, or that a clinician can't reasonably independently review) fall back under device rules. Any real product needs current regulatory guidance read by someone who does this for a living — not an essay. The claim here is narrow: this architecture sits on the friendly side of the line by design, not that it's automatically exempt.
Liability

Who's on the hook if it's wrong?

Clinicians dread "the computer told me to." It's the question that kills pilots. And here the architecture gives a clean answer that a hot-path LLM simply cannot.

hot-path LLM

A recommendation generated fresh at decision time: non-reproducible, no named author, no fixed basis. Under scrutiny, there's nothing to point at but "the model said so." A liability black hole.

constrained expert system

Every rule was reviewed and signed by a named clinician, with its guideline citation attached, evaluated deterministically, and logged. The audit trail is a shield: here is the rule, the evidence, who approved it, and when.

The audit trail we built for trust turns out to be the same artifact you need for defensibility. Keeping the LLM out of the live decision isn't just a reliability choice — it's a liability posture.

The market

Who buys it — and why they said no before

The predecessors didn't lack buyers because the value was unclear. They lacked buyers because the cost of getting value was brutal: plugging a rule system into each hospital's bespoke, messy data was a custom integration project every single time (Arden's "curly braces," again), and once installed, keeping the rules current was a chore nobody owned. The ROI never showed up.

BuyerTheir painWhat changed
Health systemsoutdated protocols, variation in care, audit pressureFHIR lowers integration cost; LLM keeps rules current
EHR vendorsCDS is a checkbox feature that's expensive to maintainauthoring becomes cheap and continuous
Specialty societiesthey publish the guideline; adoption is slow & inconsistentpublish a checkable protocol, not just a PDF
Payersenforcing evidence-based criteria consistentlydeterministic, auditable eligibility logic

The wedge isn't "sell to everyone." It's one buyer with acute pain and a clean-ish data model — most plausibly a specialty society that wants its guideline used, or a health system under audit pressure.

Timing

Why the barriers are lower now

The "why now" for adoption is different from the "why now" for the tech — and just as important. Two of the three things that buried the predecessors have quietly collapsed.

THEN — why it stalled integration = custom per sitehigh cost, every hospital different maintenance = manual, unownedrules rot as guidelines change NOW — why it's viable integration = FHIR, standardizedshared data model, mandated APIs maintenance = LLM-authoredupdates arrive as reviewable diffs the third barrier — clinician trust — is the one still standing (see next section)
Two of three barriers fell. The common data standard (FHIR) shrinks the integration tax that killed the predecessors, and LLM-authoring removes the maintenance chore. The human barrier remains.
The honest ledger

Where it still might fail

It would be its own kind of overselling to pretend the path is clear. The real risks:

clinician trust

A generation of clinicians was burned by noisy, low-value alert systems. "Another CDS tool" starts underwater. Earning trust is a product and change-management problem, not a technical one.

messy data

Even with FHIR, real records are dirty and incomplete. Mapping them to the rules' terms — curly braces, still — is real work, and it's where a deterministic system quietly depends on not-so-deterministic inputs.

glacial procurement

Healthcare buys slowly, through long pilots and committees. Runway and patience matter as much as the idea.

regulatory gray zones

The exemption has fuzzy edges. Push into higher-risk or more-autonomous territory and you're back under device rules — with the cost and timeline that implies.

◆ The honest reframe
None of these are reasons not to build it — they're the map of where the effort actually goes. Notice that three of the four are human and organizational, not technical. That's usually a sign the technical bet is the easy part.
The through-line

Adoptable because trustworthy

Here is the thing that makes me think this generation is different, and it's almost too neat: the properties that make the constrained expert system trustworthy are the very same properties that make it adoptable.

Design propertyBuys trustBuys adoption
Transparent (every rule cited)the clinician can check itfits the CDS exemption
Human-gated (approve / override)a person owns each rulefits the exemption + liability shield
Deterministic + auditedsame answer, every timedefensible under review
LLM at authoring, not decisionreproduciblemaintenance stops being a chore

Every prior gap-closer had to trade trust for capability, or capability for compliance. This one doesn't. The same discipline earns all three gates.

◆ The one-sentence version
It's innovative, it's buildable — and it may be adoptable precisely because the things that make it safe (transparent, human-gated, auditable) are exactly what regulators, hospitals, and courts are already asking for. The predecessors had to fight the gatekeepers. This one is shaped like what the gatekeepers want.

Read part I: The Constrained Expert System. The substrate: Context Graph architecture. This essay makes a business/regulatory argument, not a legal one — treat the regulatory sections as a starting hypothesis to verify, not a conclusion.

↑ Top