For pharmaceutical manufacturers
Your AI will answer confidently. That is not the same as correctly.
A model has no suspicion. Hand it a number without its unit, its grain, its date or its source and it will fill those in for you — plausibly. Predicate RX ships the organisations' own filings with all of it attached, so what your systems produce can survive being questioned.
Ask it any way you like. The facts underneath don't move.
Talk to us See what's in the feed
Four agencies. Four numbers. No agreement.
Each agency sees the organization its own way and never checks with the others. None of those numbers is the organization — each is one agency's projection of it, carrying that agency's blind spots.
A billing entity. Cannot see a system parent, because CMS does not bill one.
A filer. Cannot distinguish two hospitals sitting under one filer.
A provider. Sees facilities and units, not corporate structure.
A registrant. Sees the owner, nothing about the care delivered.
Some of the organizations that matter most hold no identifier at all in the file you happen to be reading. System parents. Management companies. The intermediaries between them.
And they do not hold still. Systems merge. Hospitals are acquired and closed. Registrations terminate and reappear under new numbers. The organization you identified last year may not exist today under that name, that number, or that structure.
Three ways a number lies to you.
Each of these produces an answer that is fluent, formatted and confidently wrong. None of them fails loudly.
The value is right. The meaning never arrived.
A hospital reports zero charity care. Is that a zero, or did it never file? Those are opposite facts, and a dataset that does not distinguish them ships both as the same empty cell. Two adjacent fields overlap, and summing them double-counts — nothing in the file says so.
WHAT WE SHIPA Sourceprint™ on every value — unit, grain, the form and line it came from, what a blank means, and what not to conclude. Attached to the number, not filed in a PDF.
Two sources agreed. Both were right. The answer was wrong.
A pricing file resolved to a different hospital in a different town and was labelled corroborated — the highest confidence available. Both identifiers described a health system. The question was about a facility.
WHAT WE SHIPResolution at the level the question asks about, every link traced to a filing the organisation made itself, and agreement that counts only when the sources are independent and describing the same level.
By the time you look, it never happened.
An arrangement is listed on the 2nd and absent on the 3rd. The registry publishes only today, so by the 5th there is no record it ever existed — and no way to request one.
WHAT WE SHIPEvery capture held, hashed and locked. Any past date reconstructable, with the observation window stated rather than a change date we cannot support.
The analyst was the error-correction layer.
Someone who knew that field was a cost rather than a charge, and that the blank meant a non-filer, used to sit between the data and the conclusion. Automate that step and the correction leaves with the job.
What replaces it has no suspicion
A model has no memory of being burned by a field, no sense that a figure looks off, and no context it was not given. What it has is fluency. It will fill every gap you leave, plausibly.
So the requirement goes up, not down
Data quality is usually treated as an input you already own while the model is the thing being solved. It is the reverse. Models are commodity and improve monthly; the data layer sets how far any of it can be deployed.
Bad data does not fail loudly
A crash is a signal. A fluent, well-formatted, confidently wrong answer is not, and it survives review. In our own build, most of an initial cost-report mapping was wrong and every figure it produced looked entirely reasonable.
And it compounds
One output becomes another's input, then a decision, then a filing. Without provenance attached from the start, nobody can say afterwards which upstream fact was wrong.
What we sell is a fact that can survive being questioned. The value, what it means, which organisation it is about, when it was true, and where it came from — travelling together. Most of the market ships the number and leaves the other four to you.
Every field arrives with its Sourceprint.
Not a citation in a footnote. The Sourceprint™ is the unit, grain, source line, period, observation date, null semantics and do-not-conclude statement that travels with a value — wherever that value goes, into whatever system reads it.
- field
- charity_care_at_cost
- value
- 48,113,000 USD
- grain
- one value per CCN per fiscal year
- source
- CMS Form 2552-10, Worksheet S-10, Line 23, Column 1
- verified against
- CMS Pub. 15-2, chapter 40
- period
- FY2023, settled version
- observed
- 2026-08-14, capture 0f3a91c2
- if blank
- reported-zero and not-filed are distinct values, never merged
- do not conclude
- Not uncompensated care. Excludes bad debt and unreimbursed Medicaid. Do not sum with uncomp_care_total — they overlap.
That metadata is what makes it background reference data rather than another dataset. Load it beneath your CRM, your accounting, your territory models, your reporting. Every internal system then speaks the same organization language — and every number your organization produces from it traces back to a line on a filing somebody signed.
A conclusion you cannot explain is a conclusion you cannot put in a submission, a negotiation, or a deposition. Most healthcare data fails at exactly that step: the number is right and the meaning is missing.
The part that cannot be bought later.
Ask what a hospital's arrangements looked like on June 1st and nobody can tell you — not the agencies, not your current vendor. The federal registries publish current state and overwrite it. There is no archive and no way to request a prior day. A team starting tomorrow starts tomorrow.
First retained day of the registry record. Every capture is hashed, locked against modification for ten years, and pinned to the exact code that produced it. Gaps are recorded and published, never filled.
Because the source overwrites, we observe states rather than events. A value differing between two captures changed somewhere in the interval — not necessarily once, and not necessarily on the later date. Every change we publish carries that window rather than a date we cannot support.
That is why the cadence matters. A daily capture bounds a change to a day. A weekly one bounds it to a week, and loses entirely any state that appeared and disappeared inside it.
What you license is a feed, not a report.
Every component below arrives on its own clock, resolved onto one persistent organization record. You do not assemble anything.
| Component | Cadence | What it carries |
|---|---|---|
| Covered entity registry | Daily | Status, entity type, control type, certification period, termination, child sites |
| Contract pharmacy arrangements | Daily | Entity–pharmacy pairs, locations, arrival and departure |
| Provider enrollment and NPI | Weekly | Facility and organizational identifiers, addresses, taxonomy |
| Medicare cost reports | On release | Every version — as-submitted, reopened, settled — held separately |
| Form 990 and Schedule H | On release | Community benefit by category, parent and affiliate structure |
| Hospital pricing files | As published | Standard charges, encoded identifiers, conformance findings |
| Corporate ownership and filings | Event | Transactions, REIT exposure, management company relationships |
| State 340B reporting programs | Annual | Acquisition cost, payments received, contract pharmacy and administrator payments, by payer type — where a state requires it. Minnesota and Rhode Island ingested; Ohio scoped for October. |
| Government contracts | As obtained | The agreements gating eligibility, sourced by records request |
Delivered as a change feed, a bulk extract, a query endpoint, or a share into your warehouse. One-way, outward. No client data is ingested, so nothing you hold is at risk in it and nothing you ask reveals what you are working on.
One record, under the whole organisation.
Right now each of these teams answers its own questions from its own sources, with its own idea of what an account is. They disagree, and nobody can reconcile them, because none of the underlying identifiers match.
Market access
Exposure by corporate parent, answered from site-level sources.
Legal
What was true on the date, and whether the pattern is systematic.
Compliance
Was this arrangement listed when the transaction happened.
Government affairs
Who is affected in this state, evidenced to a filing line.
Commercial analytics
The rollup under the CRM, the forecast and the territory model.
Finance
Channel exposure and the population behind an accrual assumption.
Medical affairs
The legal entity behind a site, for engagement and reporting.
Corporate development
What consolidated, when the record reflected it, what changed after.
Eight desks, eight versions of the truth, and a number in a board deck that nobody can trace back to a document. Put one resolved record underneath all of them and every derived figure — in a model, a deck, a submission, a deposition — runs back to a line on a filing somebody signed.
Licensing reflects that: a flat annual fee, unlimited internal use. Metering seats would penalise the adoption that makes it worth having.
The question at the end of every review is what is this based on. The honest answer has usually been an analyst's spreadsheet. Internal numbers were never the weak half — accounting is audited, owned, and defended by people who can be asked. It is the external half that arrives without a source, and a conclusion inherits the weakest input it rests on.
Close that half and the question stops being a problem to reconstruct days later. The answer travels with the value and exists before anyone asks for it.
We make your numbers auditable. Not your reasoning.
Someone can still divide by the wrong denominator using perfectly sourced inputs. What disappears is an entire class of doubt — so what remains to argue about is method, not where the figures came from.
This is not our analysis of the market. It is the market's own account of itself — compiled from what these organisations asserted to federal and state regulators, in their words and their numbers, with the form and line each one came from still attached.
Your AI. Your analysts. Your questions.
We do not sell you an answer. We sell you the ground to stand on when you go looking for your own. This is the crossing into analysis, not the analysis.
Connect it directly
Grounded, not approximated
Always the current fact
Your analysis stays yours
Reproducible on demand
Some of it is not online at all.
The government contracts that gate eligibility for a large share of participating hospitals are not published anywhere. They exist as records held by federal agencies and by the counties, cities and health departments that signed them.
We obtain them by records request. A federal request is filed and acknowledged. State public records requests run in parallel, jurisdiction by jurisdiction, because every such contract has a government counterparty whose copy is a public record under state law.
Filed in our name, for research, and dated. Requests made by a party to a dispute create a record of what that party was looking for and when. Ours predate anyone's matter and are not made on anyone's behalf.
Documents arrive into the same feed, resolved to the same organization record, carrying the same provenance as everything else — including a finding where an agency holds no such contract at all.
Building it yourself is the expensive option, and the exposed one.
Every input is public, so nothing stops you. What stops most teams is what it costs after the first year — and what it puts on the record in your name.
Five people for a year, then people forever
You cannot buy the months you did not capture
Your network, hitting thousands of hospital sites, daily
A records request in your name states what you were looking for
What you assemble is a document you created
And there is very little here to attack. We do not characterise anyone. The content is filings the organizations made themselves — to federal agencies, to state agencies, on their own forms, under their own names. What we add is the resolution between them and the record over time, published, versioned and checkable line by line.
Which leaves a dispute in an awkward place. Challenge a figure and you are challenging the filing it came from.
What we publish that nobody else does.
Unresolved and contested are normal outcomes, stated plainly rather than papered over with a plausible guess.
What we could not resolve
Entities no source asserts, and matches we declined to make. The unresolved set shrinks as sources accumulate. It does not empty.
Where sources disagree
Two filings asserting different facts is a finding about the filer, not a defect to be collapsed into a winner.
Days we do not hold
A capture failing validation is recorded as a gap and never written. A recorded gap is a limitation. An unrecorded one is a failure.
What is not observable
Acquisition price paid by a specific entity. Per-prescription dispensing. Rebate flow through a specific contract. Not in the public record, at any price.
The corners are public. The crossing isn't.
Annual subscription to the feed, with onboarding into your environment. Records-request coverage is scoped separately.