A Receipt

I killed Policy Vault because the thing it proved was not the thing I thought I was building.

That sentence took longer to earn than the product did.

Policy Vault began as a practical irritation.

Drug coverage rules were hard to find. Not impossible. Worse. Technically available, functionally hidden.

Portals inside portals. PDFs behind inconsistent page shells. Dead links that stayed official. Criteria split across target-agent tables, clinical approval sections, renewals, exclusions, quantity limits, and benefit-design notes. Public rules arranged like private knowledge.

I did not start with a theory of canonicalization.

I started with a pharmacist’s frustration.

If a rule governs whether someone can access therapy, it should not take a pilgrimage to find it.

So I built the parallel lane.

What I claimed

The early claim was straightforward:

Policy coverage criteria could be made searchable, comparable, citable, diffable, and computable without waiting for the incumbents to make them pleasant.

Not by stealing anything.

Not by pretending public data was proprietary.

By taking rules that already existed and making them walkable.

The product shape followed the claim.

Ingest payer policy content from PDFs and HTML. Clean the text. Segment true prior authorization criteria. Preserve hashes, dates, and URLs. Publish the result through a Coverage Explorer for users and a computable rules layer for downstream systems.

The method had a spine:

crawl, clean, segment, extract, normalize, validate, publish.

Every artifact needed provenance. Source URL. Hash. First seen. Last seen. Effective date where available. Enough traceability that a user could walk back from a rendered criterion to the thing it came from.

That was the promise.

Coverage rules with receipts.

What the evidence was then

The evidence was contact.

Not market research contact. Real contact.

Extraction runs that failed at 2 a.m. Policy clauses that refused to parse cleanly. Pages that worked on ten documents and collapsed at a hundred. Criteria that looked identical until one threshold changed. Program summaries that bundled multiple drugs under one policy while another payer fractured the same class into separate documents. Cross-referential nightmares with no real traceability.

The frustration was not imagined.

The web is full of policy content, but practitioner workflows needed computable coverage rules with provenance, fewer clicks, and less scavenger hunting.

Policy Vault touched that surface.

There was a short spike of unusual traffic. Enough that I noticed. Enough that I let it change my confidence.

But I did not instrument it well enough to characterize it cleanly, or productize it.

That matters.

In earlier drafts I wanted the number because the number looked like proof. That is exactly the trap. If I cannot reconstruct the analytics, the honest receipt cannot lean on the figure.

So the claim has to stay weaker:

The spike did not prove product-market fit.

It proved contact.

Someone needed the pathway.

I do not know how much of that need was durable.

What would have changed my mind

A few things should have changed my mind earlier.

If users came for answers but did not return for workflow.

If buyers liked the directory but did not need the rules feed.

If criteria freshness became more expensive than the value of the interface.

If payer-specific variation made canonicalization too brittle without a governance layer.

If the front end got attention but the address layer did the real work.

That last one was the big one.

I thought the artifact was the product.

It was not.

The artifact was a surface where the addressability thesis became visible.

The real product was not the terminal, drug/indication/criteria triples page.

The real product was the ability to say: this criterion, this policy, this plan, this version, this source, this hash, this relationship, this difference.

The address was doing more work than the interface.

I did not see that clearly enough at first.

What Policy Vault proved

Policy Vault proved that public healthcare rules can be made more usable without permission from the systems that made them hard to use.

That matters.

It proved that coverage criteria are not just documents. They can be decomposed into smaller, bite-sized operational units.

It proved that provenance is not decoration. A criterion with a hash, URL, and timestamp becomes checkable.

It proved that policy content has recurring structure: drug, plan, indication, criterion, evidence requirement, approval pathway, renewal condition, exclusion, threshold, duration, source, references.

It proved that canonicalization is not an academic move. It is what lets scattered rules become comparable, benchmarkable and auditable.

And it proved that access friction untangled creates signal.

People did not care because Policy Vault was polished.

It was not.

They cared because the maze was real.

What it cost

The cost was maintenance.

Not in the vague founder way.

In the physical way.

Every policy source had its own personality. Every payer had a preferred mess. Portal chrome. Embedded PDFs. Broken tables. Drug aliases. Objects that have variable rendering. Program-summary sprawl. Renewal criteria buried in the wrong place. Step therapy language that shifted just enough to break an extractor.

Freshness was not a feature.

Freshness was the product tax.

If a rule changed and the system missed it, the interface became dangerous. Not just stale. Misleading.

That is the cruel part about healthcare knowledge products. The better the interface gets, the more users forget the underlying instability. Smoothness becomes liability if provenance and freshness do not keep up.

A careless system would collapse them.

A useful system has to preserve the difference.

That is where the theory got sharper.

“These are the same” is not a neutral statement.

It is a decision about which differences count.

Sometimes the decision helps.

Sometimes it does quiet violence.

Where I erased too much

Failure, intolerance, contraindication, and trial are not interchangeable. They are neighboring concepts with different consequences.

The act of making them comparable risked making them falsely identical.

That is the part I need to own.

Policy Vault did not just reveal the need for addresses.

It revealed the ethics of addressing.

The address is empty.

The act of addressing is where meaning gets baked in.

Who decided these criteria co-refer? Under what authority? Against what dissent? With what local context preserved? With what appeal path if the collapse was wrong? Where are the references?!

That is lineage entering at the canonicalization step.

Not as after-the-fact citation.

What did not convert

This is the part the clean version wants to skip.

Not every failure became a sharper thesis.

Some of it was just failure.

I did not prove that a public coverage directory could become a durable product. I proved that the maze was real. Those are not the same. Pain is not a buyer. Traffic is not a workflow. Appreciation is not willingness to pay.

I also did not prove that public rules infrastructure has an obvious owner.

That problem still bothers me.

The people harmed by opacity are not always the people with budget. The people with budget often benefit from local opacity, tolerate it, or have already built internal workarounds. The people who need the cleanest version of the truth may be members, clinicians, advocates, pharmacists, or small teams without purchasing power or operational bandwidth.

That does not resolve into addressability.

It is political economy.

It is distribution.

It is the old problem: public value, private maintenance burden, unclear buyer.

I do not have a clean answer for that.

I also built too much alone. That sounds romantic after the fact. Just a builder trying to hold source truth, parsing truth, and semantic truth in one head.

Useful for discovery.

Bad for stewardship.

Why I killed it

I killed Policy Vault because the standalone product thesis was no longer honest.

The directory was useful. The coverage explorer was useful. The computable rules concept was useful.

But the durable insight had moved underneath them.

Policy Vault was not mainly about prior authorization.

It was about addressability.

Once I saw that, keeping the product alive as a standalone surface risked trapping the larger idea inside the first domain where I happened to build it.

That would have been the wrong loyalty.

The correct loyalty was to what was born from contact.

The Clinical Node ID idea was born.

POID and COID as overlays were born as early forms of a broader enumeration instinct.

Provenance entered the fray.

Hashing was seen as valuable.

Diffing survived.

The idea that public rules need walkable interfaces survived.

The idea that healthcare primitives need stable addresses survived.

The idea that a criterion is valuable only as a webpage did not.

The idea that the artifact itself was the value did not.

The artifact was the training ground.

The addressability was the transferable thing.

The survival ledger

What survived:

The addressing instinct. Give the recurring primitive a stable identity.

The provenance reflex. Preserve source, version, hash, and timestamp because trust cannot be reconstructed from vibes.

The CNID family. Coverage criteria were only the first species. The same architecture applies to problems, goals, interventions, denial reasons, appeal pathways, suspect conditions, adherence barriers, evidence objects, and investigative rationales.

The diff instinct. Change is not noise. In regulated workflows, change is often the point.

The public-legibility ethic. If the rule governs real lives, a better interface is not a luxury.

The receipt thesis. A claim gets stronger when it records what happened to itself.

What did not survive:

The standalone directory thesis. Useful, but not enough.

The idea that normalized criteria are automatically better. They are better only when meaningful differences survive.

The assumption that computability is the end state. Computability without governance can accelerate the wrong collapse.

The belief that traffic was proof. It was signal. Not evidence enough.

The belief that the artifact was the product. The artifact was evidence. The address was the lesson.

What I would build differently

I would build less front end first.

More registry.

Less page.

More primitive.

Less “find the criterion.”

More “what is this criterion, what does it bind to, what changed, who accepted the collapse, and what receipt proves the downstream decision?”

The schema would start with governance, not just extraction.

CriterionCoreID.

PolicyObjectID.

CoverageObjectID.

EvidenceAnchorID.

BindingID.

DecisionReceiptID.

Every match would carry a status: proposed, committed, rejected, overridden, deprecated.

Every criterion would know whether it came from label, guideline, payer convention, commercial overlay, renewal logic, exclusion, or plan-specific rule.

Every decision would be traceable to the policy version used at the time.

Every change would be diffed.

Every collapse would be reversible.

That last word matters.

Reversible.

A canonicalization system needs an appeal path.

Not just for members.

For meaning.

What remains unresolved

The unresolved question is not whether addressability matters.

It does.

The unresolved question is whether the most socially valuable addressability work can sustain itself commercially without becoming captured by the institutions whose opacity made it necessary.

That is not a neat product lesson.

It is a governance problem.

A public rules substrate may need to be a utility, a coalition, a standards body, a payer-facing vendor, an advocacy infrastructure, or something stranger. I do not know. I only know the solo directory was not enough.

That is the loose end I do not want to sand down.

The addressability thesis survived.

The ownership model did not.

What the sunset means

The sunset is not failure.

It is calibration.

But calibration is not redemption.

Some things got sharper. Some things simply ended.

Before Policy Vault, I could say healthcare knowledge needed provenance.

After Policy Vault, I knew provenance was not enough.

Before Policy Vault, I could say coverage criteria needed structure.

After Policy Vault, I knew structure could erase meaning if the identity decision was not governed.

Before Policy Vault, I could say computable rules were valuable.

After Policy Vault, I knew computability was only safe when paired with lineage, receipts, and human review.

Before Policy Vault, I thought the system needed a better interface to rules.

After Policy Vault, I think healthcare work needs addressable primitives that can carry evidence, governance, dissent, action, and outcome.

That is a different claim.

A harder one.

Not complete.

More true.

The receipt

So here is the receipt.

I claimed that payer coverage rules could be made more searchable, computable, and trustworthy through structured extraction and provenance.

I built the thing far enough to prove the pain was real.

I did not prove the business model.

I did not instrument the traffic well enough to use the spike as hard evidence.

I did not build enough shared governance around the semantic decisions.

I learned that the interface was not the deepest value.

I learned that canonicalization is both necessary and dangerous.

I learned that the hash is the address, not the product.

I learned that meaning is made at the equivalence-class decision, not after it.

I learned that public value and commercial ownership do not automatically align.

I learned that a corpus preaching receipts has to produce one on itself.

Policy Vault is sunset.

The addressability thesis is not.

The project closed.

The primitive survived.

A question survived too.

Who should own the address layer when the people who need it most are least able to pay for it?

That is the part I cannot close.

That is the point of a receipt.

Not to preserve the old claim intact.

To show what reality did to it.