Why assembly of the public record has ceased to differentiate, and what an explicit account of the unanswerable does for provider decision-making
Provider-side deployments of artificial intelligence for strategic and financial decision support are typically framed as problems of data access and model capability. We argue that both constraints have substantially relaxed: the public healthcare record — quality measures, cost reports, payment files, market-share files, machine-readable price files and tax filings — has been assembled and released at low or zero cost by multiple parties within approximately eighteen months, and general-purpose reasoning models are adequate to the analytical tasks involved.
The binding constraint is instead negative knowledge: an explicit, dated, sourced account of what a given file cannot support. In a working corpus of 46 verified determinations about United States public healthcare data, 22 (48%) contradict an assumption that a competent, well-resourced analytical team would otherwise have acted upon. Because no schema field encodes "this file resembles a price file and is not one," systems built on assembly alone reproduce these errors with high apparent confidence.
We propose the determinations ledger — a versioned, citable store of negative knowledge with supersession semantics — as a first-class architectural component of operational AI in provider settings, describe its schema and refusal behaviour, and discuss the error-cost asymmetry that makes it economically rational for organizations in which a confidently wrong answer is more expensive than a delayed one.
Health systems have spent the better part of a decade acquiring analytical capability, most of it organized around the premise that better decisions follow from broader data. That premise was defensible when the relevant public files were dispersed, poorly documented and expensive to assemble. It is no longer descriptive of the market. Multiple vendors now publish assembled national corpora of hospital prices, quality measures and financial filings at nominal or zero cost, in at least one case explicitly to eliminate the market for reselling public information.
What has not changed is the failure rate of the analyses these corpora produce. The characteristic error in provider decision support is not an absent answer; it is a fluent, well-formatted, wrong one. A rate comparison returned for a hospital in a state that sets rates administratively will show parity across all competitors — a true arithmetic result and a meaningless strategic one. A market-share calculation drawn from a federal file whose inclusion criteria changed will understate share in exactly the segment growing fastest. A compliance memorandum will cite a deadline withdrawn by the issuing agency in 2014.
Each of these errors is invisible to schema validation, passes every completeness check, and is produced with the same presentation as a correct answer. We take this as the central problem of operational AI in healthcare and argue that its solution is architectural rather than model-theoretic.
Between late 2024 and mid-2026, the cost of obtaining a linked national view of the United States public hospital record fell to approximately zero. Machine-readable price files mandated by federal transparency rules were crawled and republished at national scale. Cost-report, quality and market-share data have long been available for a low per-seat subscription from at least one directory service holding filings back to fiscal year 1996. Tax filings describing community benefit spending, and the needs assessments they respond to, are retrievable from public indices. Individual practitioners have demonstrated non-trivial joins of these sources using general-purpose tooling.
The strategic implication is narrow but severe: any product whose value proposition is the assembly of public data is competing with free. The implication for research is different. If assembly no longer differentiates, then observed differences in analytical quality between systems must arise elsewhere — and the location of that difference is an empirical question.
We define negative knowledge, in this setting, as a verified, dated proposition about what a specified data source does not support, together with the reasoning and evidence establishing it. Three properties distinguish it from ordinary metadata.
It is not representable in the source. A file cannot carry the field "this resembles a price file and is not one." Negative knowledge is produced by inspection, comparison and domain judgement, then attached externally.
It is perishable and dated. Federal files change inclusion criteria, guidance is superseded, and vendors alter parse behaviour without notice. A determination made in 2026 does not describe the file in 2028; it describes the file in 2026, which is precisely why it must be dated and why a later entrant inherits none of it.
It is asymmetrically valuable. A positive finding improves an answer at the margin. A negative finding prevents a category of confident error. Where the cost of the two is asymmetric — as it is in board-level, regulatory and payer-facing work — the negative finding dominates.
In a working corpus of 46 determinations produced during structured examination of United States public healthcare data sources, 22 (48%) were corrective: they contradicted an assumption that a competent, well-resourced analytical team would plausibly have acted upon without independent verification. Table 1 summarizes four representative cases and the decision each protects.
| Domain | Determination | Error prevented |
|---|---|---|
| Price | In an administratively rate-set state, machine-readable files are analytically empty for competitive comparison. | A parity chart presented to a governing board as competitive intelligence. |
| Market share | The federal service-area file began including Medicare Advantage encounters without corresponding documentation change. | Systematic understatement of share in the fastest-growing Medicare segment. |
| Compliance | A widely cited implementation-strategy deadline was superseded by subsequent guidance in 2014. | Counsel citing a withdrawn requirement to a state regulator. |
| Community benefit | An index of community health needs assessments widely believed not to exist is retrievable as a structured field in federal filings. | Procurement of a manually assembled corpus already available in the public record. |
We propose treating negative knowledge as a first-class, versioned store rather than as tacit expertise held by analysts. Each entry comprises: a stable identifier; a verification date; the source under determination, by URL and version; the proposition itself, stated in the negative; a confidence level; a pointer to any prior entry it supersedes; and a "prevents" field naming the decision at risk.
Two behaviours follow. First, citation: every generated answer resolves against the ledger and surfaces the entries that shaped it, so a reader can audit not merely the number but the judgement behind its framing. Second, refusal: where an applicable entry establishes that the available sources cannot support the question as posed, the system returns the reason and the entry rather than a number. Refusal is not a degraded response; it is the operationally correct one, and it is the behaviour absent from systems built on assembly alone.
The ledger has an additional property of interest to those studying competitive dynamics in health information: it cannot be acquired retrospectively. Entries are dated observations of files that no longer exist in the observed state. A system beginning in 2028 inherits none of 2026's determinations, and no level of funding recovers them.
A determination about public data acquires operational value only when applied to a decision, and provider decisions are made against internal figures — cost accounting, executed payer contracts, board materials — that organizations decline to export. We note, importantly, that this is not primarily a regulatory constraint: the public corpus itself contains no protected health information, and major cloud providers execute business associate agreements as a matter of course. Sector guidance issued in April 2026 declines to mandate on-premise deployment.
The constraint is instead practical and contractual. Residency — deployment within the organization's own environment — is what permits a single query to span the public corpus and the internal ledger. It is properly characterized as a capability argument rather than a compliance one, and it is a delivery requirement rather than a defensible asset in itself.
Provider strategy exhibits a pronounced asymmetry between the cost of delay and the cost of error. A capital allocation deferred by a week has a bounded cost. A capital allocation, rate posture or regulatory filing made on a plausible but incorrect figure carries costs that are institutional and, for the executive presenting it, personal.
Under such asymmetry, the utility-maximizing system is not the one with the widest coverage but the one whose stated confidence tracks its actual reliability. This inverts a familiar evaluation criterion. A system that answers eight of ten questions and correctly refuses two is superior, in expectation, to one that answers ten with two undetected errors — even though the second scores higher on coverage benchmarks, which is what procurement processes typically measure.
The determination corpus described here is a working set produced in the course of building an operational system, not a randomly sampled or externally adjudicated one; the 48% corrective rate should be read as an observation about this corpus rather than an estimate of a population parameter. Determinations reflect the analyst's prior about what a competent team would assume, which is not directly measurable. Two classes of question — physician-level referral flow and all-payer volume — remain outside the reach of any public-record system and require licensed claims; nothing in this argument addresses them.
The useful question for provider organizations evaluating operational AI is no longer which datasets a system holds. It is whether the system can state, with a date and a source, what it will not tell you and why. Assembly is finished as a differentiator. Judgement about the limits of the assembled record is not, and unlike data it accumulates in one direction only.
Organizations already hold negative knowledge; it sits in analysts' recollection and in email. Converting it into an asset requires only that each finding receive an identifier, a date, a source, a confidence level, a supersession pointer and a statement of what it prevents — and that every generated answer cite from it. Absent that step, the knowledge departs with the analyst.