A Released File Is Not Necessarily a New Fact: Reading the JFK Collection’s Provenance

The National Archives page labeled “2025 Documents Release” includes a January 30, 2026 addition: 11,022 pages in 140 PDFs. The page label does not date every item. Nor does a posting date establish a record’s creation date, first public availability or novelty. National Archives release page.

A researcher can discover a document for the first time without discovering a previously unknown fact. A newly posted scan can contain an old public record. A newly readable passage can be important even when most of its surrounding file was available before. And a newly transferred group of records can require a different finding aid from the one used for older material. These possibilities are not interchangeable, but headlines often treat them as if they were.

The practical solution is provenance: a documented account of where a record came from, how it is identified, which version is being examined and what changed between versions. Provenance does not decide whether a historical claim is true. It supplies the context needed to test the claim responsibly.

This article examines the collection’s actual access tools and then works through an original, entirely fictional comparison exercise. It does not advance an allegation about a person, identify a new suspect or claim to resolve the assassination. Its question is more basic: before calling a passage a revelation, can another reader locate the same record, distinguish it from related versions and reproduce the comparison?

1. The collection is larger than any one release page

NARA describes the JFK collection as containing more than six million pages and explains that material can be publicly available in person without yet being digitized. Its collection page describes an ongoing effort to digitize the holdings and post records as they become available. A web download set is therefore not automatically a complete representation of the physical collection. NARA’s collection overview.

For a researcher, that distinction changes the meaning of a failed search. “I found no result in this online subset” is a defensible account of a search. “The archive contains no such record” is a much larger statement. Moving from the first to the second requires evidence about the coverage and indexing of the search system, not simply confidence in the search terms.

Start by writing down the universe being searched. Is it one dated release, the legacy reference-system export, a catalog series, an agency’s records, or every digitized item returned by a broad query? A result count is meaningful only relative to that universe. If the universe changes, comparisons of counts may stop being comparable.

A useful research log might say: this search examined the named release table as displayed on a specified date; it did not examine undigitized material, analog media or all separately cataloged series. That is not an excuse for incomplete work. It is the information another researcher needs to understand what the search established and what it did not.

The same discipline prevents a large download from becoming an impressive but misleading claim of completeness. A folder containing thousands of PDFs can still omit entire categories of records, contain repeat versions or mix several release events. Before calculating percentages or declaring coverage, identify the units and the source lists against which the folder was checked.

A complete investigation does not necessarily require downloading the whole collection. It requires matching the research question to a defensible scope. A narrow question can be answered thoroughly with a small set of well-identified records; a sweeping claim can remain weak even when accompanied by a very large file count.

2. A page reviewed in 2026 can describe a dataset frozen in 2021

The JFK reference-system page says its search system is down for maintenance and supplies six spreadsheet exports current as of May 17, 2021. The page itself was reviewed in February 2026. It lists entry counts of 53,329; 53,173; 52,923; 53,200; 53,358; and 53,123, and explains that filenames identify RIF-number ranges. The displayed counts sum to 319,106 entries. That sum is our arithmetic on the published table, not a new count of the entire collection or of unique historical documents. JFK Assassination Collection Reference System.

This is a concrete example of why “updated” needs an object. A webpage can be maintained while the dataset it links to retains an earlier cutoff. A researcher who records only the page’s review date could accidentally represent an older inventory as a current one.

Keep separate fields for the date of the underlying record, the date of the dataset snapshot, the date of the web page’s review and the date the researcher accessed it. Those fields answer different questions. None should silently substitute for another when a chronology is assembled.

The entry total also needs care. An entry is a unit in a reference system. A page is a unit in a document or scan. A PDF is a digital container. A historical record may have multiple pages and multiple digital versions. Adding quantities with different units can produce a precise number with no coherent meaning.

For example, it would be invalid to add the 319,106 displayed reference entries to the collection overview’s page count and describe the result as a total number of documents. It would also be invalid to treat the six spreadsheets’ combined entry count as proof that every later transfer is represented. The source explicitly states the snapshot date.

A practical data-cleaning step is to preserve the inventory’s original field names alongside any normalized names used for analysis. If a column is renamed from “entry” to “document” for convenience, that change can introduce an assumption before any substantive research begins. A data dictionary should explain such choices rather than hide them in a spreadsheet header.

3. An identifier points to a record; it does not summarize its evidentiary value

The JFK Records Act, reproduced by NARA, defines an identification aid as a written description prepared for a record and directs attention to the collection’s physical integrity and original provenance. Identification and provenance are built into the recordkeeping framework; they are not optional decorations added by a later researcher. NARA’s text of the JFK Records Act.

In a research notebook, preserve the identifier exactly as supplied before creating a simplified version for matching. Keep the agency file number, catalog identifier and local filename in separate fields. They may refer to related objects, but they are not necessarily interchangeable names for the same level of description.

Imagine an original fictional example with a historical file reference called “Office-X / Folder-7 / Item-12.” A scan repository assigns it digital identifier D-400, while a researcher downloads it as “memo-final.pdf.” Those three strings describe different parts of the record’s handling. Renaming the local file does not change the historical reference, and a historical reference alone does not identify which downloaded version the researcher examined.

A well-formed citation needs enough of this context to make retrieval reliable. The title helps a human recognize the document. The archival reference helps locate it within the collection. A specific file link or version reference helps identify the digital representation. A page reference points to the passage being discussed. The research log’s access date and integrity record help preserve what was actually used.

An identifier also does not certify the truth of everything written in a record. A memo can contain a report of a claim, a question, a mistaken assertion or a secondhand account. The identifier establishes how to find the memo, not how much weight to give its contents. That assessment depends on authorship, purpose, timing, sources and corroboration.

The distinction is especially important when a sentence begins “government records show.” That phrase can mean a record documents an event, or merely that a record contains someone’s statement about an event. A careful writer names which one is supported. The difference should survive into the headline, not appear only in a qualification several paragraphs later.

4. Understand the catalog hierarchy before treating search results as documents

NARA’s catalog guidance distinguishes record groups or collections, series, file units and items. A series groups records associated with an activity and connects them to an archival creator; file units and items sit below that level, and descriptions can contain digital objects. Searching descriptions and downloading digital objects are therefore related but distinct operations. Using the National Archives Catalog.

A fictional search can illustrate the consequence. Suppose a result called “Correspondence, 1962–1964” is a series description. Beneath it are twelve file-unit descriptions, and one file unit contains several digitized items. Counting the series result as one document, the file units as twelve more documents and every image as another document would mix levels of description. The resulting total could not be interpreted without a clear counting rule.

Before exporting a result set, identify what each row represents. Does the row describe an archival grouping, a discrete item or a digital file? Does the search include nested levels? Are repeated results alternate descriptions, alternate versions or genuinely different records? Those questions should be answered before a chart is made.

The hierarchy also supplies context that a standalone PDF can lose. A document’s placement may help explain the activity in which it was created or retained. That does not make every neighboring item relevant to the same historical claim, but it can reveal why a particular record appears in a particular file.

NARA’s electronic-records access guidance notes that a series can indicate that not all material is online even when some attached records are available at lower descriptive levels. A label on a higher-level description should therefore not be read as a definitive statement about every item beneath it. Accessing electronic records through the catalog.

For a reader following a citation, the practical task is to move from the broad description to the specific digital object and back again. The object provides the passage; the hierarchy helps explain its archival setting. A reliable account should preserve both rather than forcing the reader to infer context from a filename alone.

5. Not every later transfer fits the older identifier system

NARA’s June 2025 update says additional FBI records transferred that year were digitized and added to the catalog under series identifier 495982978. It explicitly states that those newly provided records do not have RIF numbers. A search strategy confined to the legacy RIF exports therefore cannot be assumed to cover that separately described transfer. NARA’s update on the 2025 FBI records.

This is not a reason to discard the older finding aid. It is a reason to record its boundary. An inventory can be useful and incomplete for a later research question at the same time. The proper response is to connect finding aids through documented relationships, not to force every record into a field that some records legitimately lack.

In an original research schema, the identifier column would not demand a RIF number for every row. It would contain an identifier type and a corresponding value, with additional fields for any known cross-references. A missing RIF value would be labeled “not assigned in this source,” when supported, rather than “missing record” or “unknown identity.”

That wording matters. “Not assigned” describes a feature of the record system. “Missing” can imply that an expected identifier or document was lost. “Unknown” describes the researcher’s knowledge. The terms may look similar in a spreadsheet, but they support different conclusions.

The schema should also preserve the evidence for a crosswalk. If two identifiers are believed to refer to related material because a source explicitly links them, record that link. If the relationship is inferred from matching titles, dates and contents, record the inference and its limits. A useful join is not the same as a proven identity.

This approach reduces the temptation to treat every unmatched row as a dramatic discovery. Some unmatched rows may be genuinely new to a dataset. Others may reflect formatting, different descriptive levels or separate identifier systems. The work is to determine which explanation applies to each case, not to attach a single interpretation to the entire unmatched set.

6. A larger PDF count can reflect repackaging rather than new records

NARA’s release notes explain that splitting large PDFs into parts changed the displayed file counts. A change in the number of digital containers should not automatically be read as additional historical content. Release notes accompanying the documents.

Consider a fictional package containing one hundred scanned pages in one PDF. A later distribution divides it into four files of twenty-five pages each, with no new page images. The number of files rises from one to four, while the represented page content remains the same. A statement that “three new files were posted” can be literally correct and still be misleading if presented as “three newly discovered records.”

The opposite can happen too. One file may be replaced by a new file containing additional pages. The file count remains one, while the represented material changes. Counting containers alone misses the change. A comparison needs to examine the relationship between the files and their contents.

A useful manifest distinguishes container count, page-image count and the intended archival unit. It can also record whether a file is a complete representation, a numbered part, an attachment or a derivative made for search. The categories should be defined before a release is compared with another release.

For historical research, the decisive question is not simply whether a filename is new. It is whether the representation contains new material, newly readable material, a different arrangement, a corrected scan or only a different digital packaging. Each can matter, but they matter in different ways.

This is why release notes belong in the evidence packet. They can explain a change that would otherwise be misread as expansion or loss. Ignoring them while treating a directory listing as a complete history of the archive creates avoidable errors before the documents themselves are even read.

7. Work a fictional three-version comparison at the page level

The following example is entirely invented. The filenames, identifiers, dates, page contents and changes do not describe a JFK record or any real person. They illustrate how to distinguish a digital version from a historical document without claiming that a particular change occurred in the collection.

A researcher has three files associated with a fictional memo called Record A. The first, A-old.pdf, contains five scanned pages: a cover sheet and four memo pages. The second, A-new.pdf, contains six: a different cover sheet, the same four memo pages, and a one-page attachment. The third, A-part1.pdf, contains only the first three pages of A-new.pdf and is explicitly labeled as the first part of a split distribution.

A filename-only comparison finds three different files. A page-count comparison finds five, six and three pages. Neither comparison tells the whole story. The researcher needs a page correspondence that identifies the role of each page and shows which page in one version matches which page in another.

The correspondence would begin with the four memo pages. If their visible content is the same, the report should say so, while separately noting any differences in scan quality or markings. The new attachment is a separate addition to this package. The changed cover sheet may contain processing information rather than new historical testimony. The split file is a partial representation of A-new.pdf, not automatically another memo.

Now suppose one memo page in A-old.pdf contains an unreadable rectangle and the corresponding page in A-new.pdf contains legible text in that region. The researcher can identify newly visible text in the comparison. Before calling it newly public, however, the researcher would need to check whether another publicly available version already contained that text. A difference between two selected files does not establish the history of every other copy.

This is the central discipline of version comparison: state the comparison set. “Newly visible relative to A-old.pdf” is a bounded claim. “Never previously available to the public” is a historical claim about a much larger universe. The first can be established by the two files; the second requires additional research.

8. Turn the comparison into a record another reader can check

Here is a page correspondence for the invented package. The page numbers refer to PDF positions, not to page numbers printed on the fictional memo. The distinction matters because adding a cover changes the former without changing the latter.

The comparison contains two different kinds of change: an addition to the package and a change in visibility within a shared memo page. Treating both as “new pages” would misstate the result. There is one additional page position in the new complete package; the newly legible region is on an existing memo page.

There is also a replaced cover. Whether that cover matters to the historical question depends on its contents. A processing date might establish when a representation was prepared. It would not automatically establish when the memo was written. A title supplied by a later processor should not silently become the memo’s original title.

The partial file introduces a third issue: absence from a part is not absence from the whole. The last memo pages and attachment are not in A-part1.pdf because that file contains only the first three pages. If a reader has not obtained the remaining part, the correct status is incomplete package, not proof that the archive removed those pages.

For this exercise, a reproducible finding would name the two complete files, identify the newly legible region as memo page three, describe the added attachment and state that the part file contributes no additional page content. It would also say whether the comparison was visual, textual or both. A text-only comparison cannot establish that a faint stamp, marginal note or diagram is unchanged.

The table is an analytical product, not a replacement for the files. It must remain linked to them. If a reviewer challenges the correspondence, the researcher should be able to show the paired pages and explain the match using layout, pagination and text rather than insisting that an automated similarity score is decisive.

9. A searchable text layer is a finding tool, not the page itself

NARA’s Citizen Archivist guidance explains that optical character recognition can struggle with such features as bleed-through, stamps and mixtures of handwriting and typed text. It describes transcription as a way to make difficult records more accessible and searchable. Search results depend partly on these representations, not solely on what a person can read in the underlying image. Citizen Archivist frequently asked questions.

Add a fourth object to the fictional package: a text extraction from A-new.pdf. On memo page three, the invented image says, “The shipment was not approved.” The extracted text omits the word “not.” This is a hypothetical error, not a quotation from a JFK record. It demonstrates why the image must be checked before a sentence is reported as evidence.

The extracted sentence and the visible sentence point in opposite directions. A researcher who compares only text layers could describe a substantive reversal that never occurred in the page images. The right correction is not to invent a reason for the discrepancy. It is to identify the extraction error, preserve the original extraction separately and attach a verified transcription to the specific page.

A transcription should distinguish readable text from editorial reconstruction. If a word remains unclear, mark the uncertainty rather than completing the sentence from expectation. A second reader may be able to resolve it; otherwise, the uncertainty belongs in the report. Confidence in the surrounding narrative is not a substitute for legibility.

Search coverage can fail in the other direction too. A name visible in an image may be absent from its text layer, misspelled by extraction or divided across lines. A no-result query therefore establishes no matching indexed text under those search conditions. It does not prove that every relevant image lacks the name.

For the page comparison, preserve three linked but distinct objects: the downloaded file, the extracted text used for searching and the checked transcription used for quotation. Record which one produced each finding. This limited separation is more useful than a blanket instruction to distrust all OCR. The text layer can efficiently locate candidates; the page image and context determine whether a candidate supports the claim.

10. What a file hash can establish—and what it cannot

The Library of Congress’s June 2017 PREMIS preservation-events vocabulary separates calculating a message digest from checking fixity. The latter concerns whether an object has changed over a period, using a previous value for comparison. It also distinguishes changes such as renaming, migration and metadata modification. Those distinctions help researchers describe digital handling precisely. PREMIS preservation-events vocabulary.

NIST’s August 2015 Secure Hash Standard describes message digests as a means of detecting changes since a digest was generated. NIST FIPS 180-4.

For our original example, suppose a researcher downloads A-new.pdf and computes a cryptographic hash. That result is an integrity reference for the bytes received. If the saved file later produces a different value under the same algorithm, it no longer matches that reference. The change needs investigation; the hash alone does not explain it.

A different hash need not mean a historical sentence changed. Adding an OCR layer, modifying embedded metadata or recompressing a PDF can alter the file’s bytes while leaving its visible pages apparently unchanged. Conversely, two visually similar files should not be called byte-identical merely because a quick inspection found no difference. Visual equivalence and byte identity are different tests.

The researcher should not manufacture a convincing-looking hexadecimal value for a file that has not been downloaded and measured. In this fictional exercise, no actual PDF exists, so the sample record uses the explicit field value “not calculated; hypothetical object.” In a real project, the value would be generated from the saved file and recorded with the algorithm, time and file size.

A useful integrity record would identify the object as the original download rather than a cleaned-up derivative. If a researcher adds searchable text, the new derivative receives its own identifier and hash, with a recorded relationship to the original. Replacing the original silently would destroy the ability to show what was downloaded before processing.

The hash does not authenticate the historical author. It does not establish that a statement in the document is true. Nor does a hash calculated today prove that the same bytes were online years ago. It supports a narrower question: whether the file being examined matches the reference that was actually recorded. The provenance of that reference still matters.

There is a useful two-stage test for a later download from the same address. First compare the file-level hashes. If they match, the integrity comparison supports byte identity within the limits of the method. If they differ, compare the page structure and relevant content rather than jumping directly to a claim of substantive alteration. Record whether the difference was packaging, metadata, image quality, page addition, removal or changed visible text; leave the category unresolved when the evidence does not decide it.

The PREMIS data model separately represents objects, events, rights and agents. A small research project need not implement the entire standard to benefit from that separation: which object, which operation, which permission and which person or system are different questions. PREMIS Data Dictionary version 3.0.

Applied to A-new.pdf, the object is the downloaded representation; the event might be extraction of searchable text; the agent is the tool or researcher that performed it; and the rights record concerns permitted use. None of those fields should be filled with assumptions about the memo’s historical content.

11. The identification system has its own history

The Assassination Records Review Board’s 1998 final report describes the development of a document-based tracking process. Chapter four explains the historical thirteen-digit identification number in three groups: agency, the disk used to create the identification aid, and the record on that disk. The report also contrasts the document-level approach with conventional archival description by larger groupings. This is an account of how the review system was developed, not a statement that every later accession must use that same system. Review Board final report, chapter four.

That history helps explain why treating an identifier as a simple chronological sequence can be misleading. Parts of the number encode administrative organization. Without a source defining the scheme, sorting identifiers numerically should not be presented as sorting documents by creation date, release date or importance.

It also explains why a digitized collection can contain several overlapping organizational layers. A record has a history within the creating agency, another history within a review process and another within digital publication. Each layer may supply its own identifiers and descriptive fields. The research task is to preserve their relationships, not to declare one layer the only real identity.

A transcription error in an identifier can make a useful citation fail. Preserve leading zeros, hyphens and the original text form unless the system explicitly treats them as insignificant. If a spreadsheet converts an identifier to a number, the normalized value should remain a separate field from the source string. This is an original data-handling recommendation: it follows from the need to reproduce a reference, not from an assertion that a particular NARA export contains such an error.

For the fictional Record A, a crosswalk might have one archival reference, two complete digital representations and a split derivative. The relationship is one-to-many at the digital-file level. Flattening it into one row per filename can be useful for downloads, but it should not silently become one row per historical record in a claim about discoveries.

12. Provenance makes a claim testable; corroboration gives it weight

NARA’s JFK FAQ describes an intentionally broad category of assassination-related records, encompassing material that reports on, analyzes or interprets relevant activities and events. Inclusion in that category is not a certification that every assertion in every document is correct. JFK collection frequently asked questions.

Consider three more fictional documents, unrelated to any actual case. Document B is a clerk’s contemporaneous log recording that a parcel arrived. Document C is a later memo saying that someone remembered the parcel arriving. Document D copies C’s description. All three can be authentic records. They are not three independent observations of the arrival.

An evidence map should distinguish the event being described from the chain of reporting. B may be evidence of an administrative entry made at the time. C may be evidence of a later recollection. D may establish that the recollection circulated. Whether the parcel actually arrived as described still depends on the records’ reliability, context and any independent corroboration.

The copying relationship matters when counting sources. If D repeats C, treating them as two independent confirmations exaggerates the evidence. A larger number of files can reflect the spread of one account rather than a larger number of witnesses or observations. Provenance can expose that dependency without requiring the researcher to dismiss either record.

Dates need the same separation. The date on C might be the date the memo was written, not the date of the remembered event. The date on D might describe copying or filing. A timeline that puts all three dates in a single “event date” column could invent a sequence that none of the documents supports.

A better claim record would say: “C records a later recollection of the parcel’s arrival; D repeats that account; B is a separate contemporaneous entry whose relationship to the same parcel needs checking.” That sentence may be less dramatic than “three documents confirm the delivery,” but it gives the reader specific propositions to test.

The method does not require reflexive skepticism toward every official record. It requires matching the conclusion to the record’s function. A routing slip can strongly establish a routing action while saying little about the truth of an enclosed allegation. A draft can illuminate a decision process without proving that its proposed action occurred. A question in an investigative note can show what investigators considered, not that the suspected explanation was established.

13. Build a small evidence packet, not an untraceable screenshot collection

For the fictional comparison, the final research packet would contain the complete representations that can appropriately be retained, a manifest, the page correspondence, a checked transcription of the disputed region and a short account of the search scope. The manifest would record the source page and direct file address separately, because the source page may provide release context that the file lacks.

One filled manifest entry could read in prose: local object A-new; represented archival unit Record A; complete six-page package; downloaded from the documented source address at the recorded time; integrity value not calculated because this is a hypothetical exercise; derivative A-text created for searching; comparison counterpart A-old; partial representation A-part1; unresolved question whether an earlier public copy already showed the now-legible region. The explicit unknown is part of the result, not a defect to conceal.

A screenshot alone would not preserve these relationships. It can illustrate the specific region under discussion, but it may omit the preceding sentence, page number, source identifier or marks showing that the passage is a quotation inside another document. Retain enough context for a reader to understand the image’s role without redistributing irrelevant sensitive material.

NARA’s October 2023 digitization-quality guide identifies risks including omitted records, insufficient image detail and inadequate metadata, and treats testing and validation as parts of a digitization process. A researcher inspecting a few downloaded pages is not thereby certifying an archive’s entire digitization program. The useful lesson here is narrower: legibility, completeness and contextual metadata are separate things to check. NARA Digitization Quality Management Guide.

In the exercise, the packet can pass its page-correspondence check while failing a quotation check if the disputed word is still unreadable. It can pass a file-integrity check while remaining incomplete because only the first split part was obtained. The final report should preserve those separate outcomes rather than assign a single vague label such as “verified.”

14. Public availability does not erase publication judgment

NARA warns that the release contains personal identifying information and describes mitigation work with the Social Security Administration. That warning is a reason to avoid reproducing unnecessary personal identifiers, not an invitation to collect them. NARA’s release notice.

A provenance article can explain a record’s location, version and evidentiary limits without republishing a private person’s identifying number or unrelated contact details. If an illustrative crop contains such material, an editorially redacted derivative can be used with a clear note that the publisher made the redaction. The derivative should not be misrepresented as the archive’s original version.

Availability and reuse rights are also different questions. NARA’s collection FAQ cautions that some photographs and images may be privately owned and that users must determine whether permission is needed. An online archival location should not be treated as a universal reuse license. NARA’s guidance on reproduction and copyright.

For the fictional Record A, the defensible published finding is specific: the compared complete package gained an attachment and a different cover; one region on a shared memo page became legible relative to the selected older file; the split part adds no further content; the exercise has not established when that region first became public anywhere. That finding answers the comparison question without inventing a historical revelation.

The same standard applies to real research. A newly available record can matter enormously, but its importance should emerge from a reproducible account of what it contains, what changed and how it relates to other evidence. A page count, a dramatic filename or an official-looking screenshot cannot do that work alone.

The JFK collection offers a large and evolving body of material for public study. Its scale makes careful description more important, not less. Distinguishing the record from its scan, the scan from its text layer, the release from the first public disclosure, and an assertion from an independently supported fact gives readers something more durable than a revelation headline: a path they can follow and check for themselves.