06-reference/research

mcd prototype phdata piia clearance

2026-10-05·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)·! medium
hclsphdatasnowflake-marketplacelicensingmedicare-coverage-database

Code-stripping solves the AMA license problem and does nothing for the inventions-assignment problem

The question

"Would a code-stripped Medicare Coverage Database (MCD) prototype survive phData's 5-clause PIIA check, and what specifically would need to be stripped or reframed to clear it?"

Context: [[2026-09-30-mercor-lessons-and-snowflake-dataset-monetization]] §5 names the five-clause read as the blocking first step before any HCLS dataset build, and §4 flags that the agreement is unread. This brief pins the prototype shape and runs the test. Two scope statements up front, both binding.

This is not legal advice. It is research supporting a founder decision. Every clause verdict below is conditional on text this brief has not read, and anything load-bearing should go to an employment lawyer in the founder's state before money changes hands.

The agreement is not on local disk. A sweep of ~/rdco-vault, ~/Documents, ~/Downloads, ~/Desktop, ~/Projects, iCloud Documents, and Mail attachments found no signed employment agreement, offer letter, PIIA, or handbook. A qmd sweep found no vault doc carrying its terms, which independently corroborates the same finding already recorded in [[2026-10-02-phdata-employee-vendor-community-role-precedent]]. So the clause test below is written as a keyed conditional template: for each of the five clause numbers from [[2026-07-31-fde-equity-kicker-retainer-pricing-shape]], it states what the clause governs, and what the verdict is under each of the two readings that clause commonly takes. No clause content is invented, quoted, or paraphrased from any phData document, because no phData document was available to read.

The prototype shape, pinned

The queue row rated this spec=4 because "a code-stripped CKE prototype over 20 LCDs" is not testable as written. Pinning it, in four axes, plus a phase split that turns out to carry the whole answer.

Data source. The MCD weekly download bundle — Current Local Coverage set plus the National Coverage set — restricted to the narrative and metadata tables only: document ID, title, authoring contractor and jurisdiction, original effective date, revision history, coverage indications and limitations narrative, summary and analysis of evidence, bibliography. The companion Billing and Coding Article code tables are not ingested at all, which is where the licensed code sets are concentrated.

Transformation. Chunk the narrative by document section. Attach a founder-authored tagging layer: service category, specialty, jurisdiction, effective-date window, coverage posture (covered / covered-with-conditions / not covered / investigational), and a source URL plus document version stamped on every chunk. The tagging layer is the asset, and it is also the only thing in the build that an inventions-assignment clause can attach to. The CMS narrative is not an invention; the schema and the judgment rules are.

Artifact — phase 1. A local, non-Snowflake prototype: 20-40 documents in a local store with a retrieval and answer harness, scored against 10 pre-written coverage questions with a written answer key. No Snowflake account involved.

Artifact — phase 2. Only if phase 1 passes its own test and the clause read clears: a Cortex Knowledge Extension / Cortex Search Service in the founder's own personal Snowflake account.

Distribution surface. Phase 1: nobody. Not shared, not listed, not published, no post. Phase 2a: private listing to named accounts. Phase 2b: public or paid Marketplace listing, with the founder or an RDCO entity as Seller of Record.

The phase split is the finding. The five-clause test does not return one verdict. Phase 1 plausibly clears all five. Phase 2b plausibly clears none of the three that matter.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The verdict: code-stripping clears the license problem and does nothing for the clause problem, and the clause that most likely blocks this is the inventions-assignment clause, for a reason that no amount of stripping can reach. The MCD narrative is not an invention. The tagging layer — the schema, the coverage-posture taxonomy, the judgment rules that make the product worth paying for — is. If clause 3 assigns inventions "relating to the Company's business," then a Snowflake-delivered healthcare data product built by a Snowflake-partner consultant is about as related as a thing can be, and relatedness is satisfied regardless of whose laptop it was built on. The own-time/own-resources carve-out that normally rescues side projects has a third prong in most statutory versions — unrelated to the employer's business — and this project fails that prong on its face. phData publishing "How to Monetize Healthcare Data in Snowflake" is not a coincidence to be managed; it is the fact that makes the relatedness test easy to satisfy against the founder.

So the honest ranking of danger is not the one the question implies. Clause 3 (PIIA) is first because it reaches the asset itself and because stripping is irrelevant to it. Clause 5 (corporate opportunity) is second and is the clause the framing ignores entirely — monetizing healthcare data on Snowflake is a stated phData service line, and if the opportunity was recognized through the role, a present-first clause reaches it. Clause 2 (conflict of interest) is third, and it fires for a specific reason worth naming: Snowflake is phData's Elite partner, a paid listing requires a Snowflake business-development contact and Stripe payouts, and the buyer universe is RCM vendors, payers and health systems — phData's prospect universe. That makes the founder a revenue counterparty of his employer's most important partner, selling to his employer's prospects. Clauses 1 and 4 are real but cheap to manage: phase 1 is unpaid and undistributed, and a Marketplace listing is pull-only by construction, which is also why the 09-30 brief's inbound kill criterion is fortunate rather than merely convenient.

The single reframe that does the most work is price, not scope. Going from a paid listing to a free listing removes the Stripe relationship, removes the Seller-of-Record indemnity, removes the business-development contact, and demotes clause 2 from a financial-interest question to an outside-activity question. It does not touch clause 3. The only thing that touches clause 3 is a written, project-named carve-out from phData — the standard mechanism, usually an excluded-inventions or prior-inventions schedule. That is a conversation, not a build task, and it is the actual gate. Everything else on the list is hygiene: personal hardware, personal accounts, personal time, zero reuse of phData-authored methodology (the asset-creator-framework material is the exact category to keep out), a timestamped build log, and a provenance note written now recording that the idea came from public reading about Mercor via a founder iMessage on 2026-09-30 rather than from phData work. That provenance note is cheap today and unreproducible in six months, and it is the single highest-value-per-minute item in this brief.

Two scope changes fall out of the license fetch and should be made before any build. Exclude dental coverage entirely rather than code-stripping it, because the CDT AI clause explicitly reaches outputs that rely on embedded CDT even when the codes are not visible. And anchor the prototype on NCDs rather than LCDs, because NCDs are CMS-authored and are clean government works, while LCD narrative is MAC-authored and carries an unresolved contractor-copyright question that no one has tested. That reverses the 09-30 brief's "20 LCDs" starting point. It also shrinks the corpus, which is a cost — there are far fewer NCDs than LCDs — so the honest framing is that NCD-first is the clean-title option and LCD-inclusive is the useful-corpus option, and that trade is a real one rather than a free improvement. Note also that the restriction frame should be license-tiered rather than blanket: CPT, CDT and UB-04 are out, but HCPCS Level II and ICD-10-CM are CMS- and CDC-published and are not subject to the three MCD licenses, so they can be retained as identifiers (medium confidence on ICD-10-CM, which is a US modification of a WHO classification; the CDC publication posture is permissive but the WHO base rights are not researched here).

Clause-by-clause verdict table

Verdicts are stated per phase because they diverge. Clause headings and numbering are from [[2026-07-31-fde-equity-kicker-retainer-pricing-shape]]; the characterization of what each governs is neutral and generic, not drawn from any phData document.

# Clause What it governs (one neutral line) Phase 1 (local, unpaid, undistributed) Phase 2b (paid public listing) The specific strip-or-reframe
1 Outside Activities / Outside Employment Whether activity outside the role requires prior written approval, or only non-interference with duties Clear under a non-interference reading; needs change if approval is required for any outside activity in the company's field regardless of pay Needs change (disclose) under non-interference; blocked pending written approval under an approval-required reading Keep phase 1 unpaid, unshared and unpublished so no approvable activity exists. Seek written approval before any listing, free or paid — the listing, not the money, is the trigger under an approval-required reading
2 Conflict of Interest Whether a financial or ownership interest in a customer, vendor, partner or competitor is prohibited or merely disclosable Clear — no financial interest exists Needs change under a disclose reading; blocked under a prohibit reading Drop the price. A free listing removes the Stripe payout, the Seller-of-Record indemnity and the required Snowflake business-development contact, which is what creates the counterparty relationship with phData's Elite partner. If paid is non-negotiable, this clause must be read before anything else
3 Proprietary Information and Inventions Assignment Ownership of inventions and works, and the scope test (relating to company business / when conceived / whose time and resources) Needs change — the tagging layer is assignable the moment it exists, regardless of distribution Blocked without a written carve-out, under a "relating to the Company's business" reading Code-stripping does nothing here. Required: a project-named written carve-out or excluded-inventions schedule from phData. Supporting hygiene: personal hardware, personal accounts, personal time, no phData-authored methodology reuse, timestamped build log. Note the own-time/own-resources carve-out likely fails its relatedness prong for a Snowflake data product
4 Non-solicitation of customers / clients Whether approaching phData customers or prospects is restricted Clear Needs change only if go-to-market involves outbound Pull-only distribution. No outbound, no named-account targeting, and no use of phData CRM or prospect knowledge to choose buyers. The 09-30 kill criterion (3 inbound access requests in 60 days) is already inbound-only — keep it that way rather than shortcutting it
5 Corporate opportunity Whether opportunities arising through the role must be presented to the employer first Clear if provenance is documentably personal Needs change → blocked depending on whether it is a present-first clause Write the provenance note today: the idea came from public reading about Mercor via founder iMessage 2026-09-30, not from phData work. If the clause is present-first, present it and obtain a written declination. This is the clause the question's framing omits and it is the second-most dangerous

Net. Phase 1 as pinned above is clear or cheaply-clearable on four of five clauses, with clause 3 the standing exception. Phase 2b is blocked on clause 3 and conditionally blocked on clauses 2 and 5. The decision this brief supports is therefore: build phase 1 now under the hygiene list, and treat the clause read plus a carve-out conversation as the gate in front of phase 2, not in front of phase 1. That inverts the 09-30 brief's §5 sequencing, and the inversion is the recommendation.

Why this is in the vault

It resolves the blocking first step that [[2026-09-30-mercor-lessons-and-snowflake-dataset-monetization]] §5 placed in front of every HCLS dataset idea, and it resolves it by splitting the gate: phase 1 of idea #1 can start now, while the clause read moves to the phase-2 boundary. It also corrects two things in that brief's §4 and §5 — that "code-stripped" addresses the clause risk (it does not), and that LCDs are the right starting corpus (NCDs have cleaner title). The Nov-30 credibility targets in [[2026-09-29-nov30-credibility-targets]] depend on something shippable in HCLS, and this is what makes a start permissible without a document the founder has not read.

Open follow-ups

Related

Sources

Vault

Web (all fetched 2026-10-05 unless noted)

Not consulted