1. A missing document leaves more than one unanswered question
A report refers to an attachment, but the attachment is absent. A payment appears in an account, but the project report cannot be found. A witness remembers a meeting whose minutes are unavailable. These situations invite a familiar shortcut: supply the missing detail with whichever story already seems most plausible. The shortcut is especially tempting when the surrounding history includes secrecy, institutional misconduct or demonstrated destruction. But knowing that an archive is damaged is not the same as knowing what every lost page contained.
The difficulty is not solved by assuming the opposite. Missing material cannot automatically clear an institution, either. A claim may remain unresolved precisely because the records needed to test it are gone. There is a difference between reporting that a specific allegation is not established and reporting that nothing objectionable happened. Both accusation and exoneration can exceed the surviving evidence. The useful question is what a particular record, absence or cross-reference permits a reader to conclude, and what additional material could change that conclusion.
The August 3, 1977 Senate hearing on MKULTRA records a concrete archival discontinuity. CIA testimony described seven newly discovered boxes, principally financial material, located in retired budget holdings rather than the expected project files. According to that account, the filing arrangement allowed material to escape destruction in 1973 and discovery during the 1975 investigations. The testimony also cautioned that the recovered material did not supply a complete picture of the activities. Those are attributed statements in a historical hearing, not a claim that today's public archive is complete. Senate joint hearing, August 3, 1977.
This article examines how to reason when evidence is uneven. It is not a new reconstruction of individual MKULTRA experiments, a catalogue of alleged perpetrators or a verdict about a living person. The historical material establishes why the problem matters. The worked examples below are deliberately invented and identified as such. Their purpose is to make the consequences of each bookkeeping choice visible without silently filling real archival gaps with fictional people, events or dialogue.
A good account can be firm about a verified limitation. It can say which attachment is absent from a particular copy, what a destruction notice covers, and which records were actually inspected. It can also identify uncertainty precisely enough to be useful. “We do not know” becomes more informative when followed by the question that remains open and the evidence that would help answer it. That is more demanding than choosing a dramatic ending, but it gives readers something they can test.
2. Separate the fate of the object from the truth of the story
Consider a fictional records project called the Alder archive. Alder is not an intelligence agency, a disguised historical program or a description of a real institution. Its register lists twenty-four document identifiers, D01 through D24. For this exercise, the register establishes that these twenty-four objects once existed. It does not establish that they contain every record the organization ever created. That distinction fixes the denominator for the calculations without turning a convenient list into an imaginary complete universe.
The first twelve documents survive as complete digital copies. Three more, D13 through D15, have partial digital copies. Three, D16 through D18, have been physically located but are not digitized. Two, D19 and D20, are covered by item-specific destruction documentation in the invented packet. D21 is identified as existing but withheld from this research release. The locations and fates of D22 through D24 remain unresolved. These categories are mutually exclusive descriptions of each identifier's current state in our exercise, not universal archival terminology.
The six groups sum to twenty-four. Nineteen documents are confirmed to survive in some form: twelve complete digital, three partial digital, three located physical and one existing withheld document. Two are documented as destroyed, and three have unresolved status. Saying “nine records are missing” because only fifteen have digital content available would collapse fundamentally different conditions. Saying “only two are missing” because two were destroyed would erase the three unresolved objects. Neither sentence communicates what the register actually shows.
NARA's explanation of its electronic-records catalog provides a relevant real-world caution: not every file in a series is available online, and access or use restrictions can affect availability. Its description also explains that online access may expose a selected version rather than every version in custody. A download interface is therefore not a complete inventory of everything held. This is a limited statement about the catalog, not an explanation for any specific unavailable intelligence record. National Archives catalog explanation.
For Alder, the next useful action depends on the category. D16 calls for access to a known physical object. D21 calls for understanding the stated restriction and any appropriate review route. D22 calls for further location work. D19 calls for examining the evidence and scope of destruction. Sending all four down the same research path wastes effort and creates misleading status reports. An unanswered location question should not quietly become a destruction finding just because the project has reached its publication deadline.
The register should also distinguish an object's survival from its contents. A surviving document can contain a mistaken statement. A destroyed document can have been accurately quoted elsewhere, inaccurately quoted elsewhere, or never quoted at all. “Survives” is not a reliability grade. “Destroyed” is not a content summary. Keeping those fields separate prevents the physical condition of the archive from deciding the historical argument before the evidence has been examined.
3. Count documents, copies and pages separately
The Alder download folder contains seventeen files, yet it represents only fifteen document identifiers. D01 and D05 each appear twice. Nothing new about the historical activity follows from the extra copies. They might still be useful for checking legibility or annotations, but they do not increase the number of distinct represented documents. A headline announcing seventeen newly available records would overstate this particular inventory by two. A headline announcing fifteen complete records would make a different mistake, because three are partial.
The partial documents illustrate why page counts need their own denominator. D13 is expected to contain four pages; pages one, two and four are present. D14 is expected to contain three pages, but only page one is available. D15 is expected to contain three pages, and pages two and three are available. Across those three documents, ten pages are expected and six are present. Four pages are unavailable. That is not equivalent to four absent documents, four suppressed claims or four confirmed instances of destruction.
In this fictional packet, the expected lengths come from reliable contemporaneous page inventories supplied as part of the example. Without that premise, a numeral on a scanned page would need further interpretation. A packet's page “4” might be its fourth leaf, an exhibit number, or part of an earlier numbering sequence. Our calculation is valid because the expected object boundaries were specified before counting. A real researcher needs to explain the corresponding basis rather than borrowing certainty from the clean structure of a teaching example.
The local calculation accompanying this article actually reconciled these invented rows. It found twenty-four register entries, seventeen downloaded files, fifteen represented identifiers, and four missing pages across the three partial documents. It also ran deliberate failure cases: a duplicate register identifier, a download pointing to nonexistent D99, and a fifth page added to a four-page document. Each was rejected. These are checks of a small synthetic inventory, not an audit of a government archive or a demonstration that real historical evidence has been recovered.
The distinction changes the questions an editor can ask. If page three of D13 is unavailable, a reference to “the conclusion on the following page” creates a specific lead. It does not tell us what the conclusion said. If D14's cover identifies three enclosures, the cover can establish that enclosures were listed, while the actual inclusion or later separation of those enclosures may require more evidence. The sentence published should identify which proposition the surviving page supports rather than presenting an absent attachment as though it had been read.
Copy comparison can also expose disagreement without resolving it. Suppose D05's first scan shows a handwritten marginal note and the second does not because its edge is cropped. That would explain a difference in visible content without requiring alteration of the original. If both scans show the complete margin but different writing, another explanation is needed. In either case, retain the two observations before choosing among explanations. A tidy merged transcript should not conceal a contradiction that materially affects the claim.
4. Financial traces answer financial questions first
A September 9, 1977 Justice Department memorandum addressed obstacles to potential criminal convictions relating to MKULTRA. It identified the passage of time, recordkeeping deficiencies and destruction as evidentiary problems, and regarded newly described financial documents as offering limited help on specific intent. This was a historical legal assessment, not a court judgment and not a statement of today's law. Its relevance here is the mismatch between finding additional records and finding the particular evidence needed for a particular question. Justice Department memorandum, September 9, 1977.
Return to Alder, where three authorizations allocate $5,000, $4,000 and $3,000 to a project. Three payment entries record $5,000, $2,500 and $1,000. The authorizations total $12,000 and the recorded payments total $8,500. The difference is $3,500. That calculation establishes a difference between two lists. It does not establish that $3,500 was stolen, returned, never drawn, paid from another account or spent without documentation. Those are competing possibilities requiring different additional records.
Now suppose the $2,500 payment has a narrative field saying “Phase B support.” A researcher might reasonably seek the Phase B agreement, deliverable acceptance or correspondence. The payment does not itself establish that a specified experiment occurred, that a participant consented, or that the work produced the result later claimed. The reason is concrete: none of those propositions is a field in this invented payment entry. Adding a dramatic interpretation to an ambiguous label does not make the missing fields appear.
Even the money needs disciplined treatment. An authorization is not necessarily a payment, a payment is not necessarily a final expense, and a refund changes a net amount without erasing the original transfer. In our invented packet, the amounts are explicitly authorizations and recorded payments; no refund entry is supplied. The defensible statement is therefore the two totals and their difference. Calling $8,500 the project's total cost would require a completeness premise that the packet does not provide.
A further fictional complication makes the issue sharper. One letter requests a $1,000 transfer; a later ledger entry records $1,000 paid; an annual summary repeats that same payment. There are three documents supporting one payment, not three independent payments totaling $3,000. Conversely, identical amounts on two dates need not be duplicates. The appropriate join uses a transaction identifier or a documented relationship, not amount alone. A researcher who merges solely on dollar value can erase genuine transactions just as easily as a researcher who sums every reference can invent extra spending.
The strongest outcome of this reconciliation may be a narrower question: which document would explain the remaining $3,500, and who would have created it? That question is actionable without assuming misconduct. It can also expose a genuine accountability gap if no explanation can be found after a well-described investigation. The gap matters on its own terms. It does not become more important by assigning it an unsupported answer.
5. A destruction claim needs a defined object and a defined event
NARA's current scheduling guidance explains that federal records require approved disposition authority and distinguishes temporary records slated for destruction from permanent records destined for transfer. It states that unscheduled records must be treated as permanent until scheduled. These current instructions help readers understand why destruction and unlawful destruction are not interchangeable categories. They are not, by themselves, a retrospective legal determination about conduct in 1973 or any other historical case. National Archives scheduling guidance.
In the Alder exercise, D19 and D20 are covered by an item-specific completion notice. That premise supports our classification of those two objects as documented destroyed. It says nothing about D22 merely because D22 has a nearby identifier. If the notice instead listed a storage carton, the mapping from carton to documents would become an additional evidentiary step. If the carton inventory was prepared after the destruction, its basis would need examination. The strength of the conclusion depends on the chain of identifications, not the confidence of the final verb.
Imagine three separate documents: a proposal to clear a room, an instruction authorizing disposal of a specified carton, and a completion notice recording the disposal. They describe different stages. The proposal may never have been approved. The authorization may have been canceled. The completion notice may omit attachments or use a broader label than the original inventory. Treating any one as automatically equivalent to the others can move an event backward or forward in the account, or describe an event as completed when only an intention was documented.
Dates deserve their own reconciliation. Suppose Alder's authorization is dated April 1, its signature April 3, the completion notice April 12 and the archival receipt May 1. These are four dates associated with different actions, not four competing dates for a single moment. A report could accurately say the authorization document bears an April 1 date and was signed April 3. It should not silently use the earliest date as proof that all later steps had already occurred. If the sequence is impossible, preserve that problem instead of normalizing it away.
Scope matters just as much as chronology. A notice might refer to the local office's originals while recipients retain copies. A destruction entry might cover working notes but not a final report. In Alder, the item-specific notice establishes the fate assigned to D19 and D20 within the example, but our register does not say whether an unlisted recipient holds another copy. We cannot announce that all versions vanished from the world. A finding about a bounded collection is still valuable without that universal claim.
Finally, motive is another question, not a synonym for destruction. The hypothetical documents above can establish that an action was proposed, authorized or recorded as complete without settling why it occurred. Inferring motive would require additional evidence and careful attribution. Where records support a serious allegation, the article should state that evidence clearly; where they do not, dramatic language should not supply it. Neither institutional reputation nor public suspicion can substitute for the missing link.
6. An allegation register is not a list of established findings
NARA's unauthorized-disposition case index says it creates case files to track allegations and communications until an issue is resolved. The displayed records distinguish statuses including founded, unfounded and pending review or follow-up, with letters available in some cases. The existence of a row therefore does not, by itself, establish that every allegation in its description was substantiated. A case's correspondence and stated disposition matter. This article makes no finding about any individual current case in that index. National Archives unauthorized-disposition case index.
A fictional complaint about Alder could say that D22 was deliberately destroyed. The register would record that allegation as an allegation and separately record D22's object status as unresolved. If the two fields were combined, a search for destroyed files would return D22 as if the complaint had established its fate. The next analyst might then cite the search output as corroboration of the complaint. An allegation would have traveled through a database and emerged wearing the appearance of an independent finding.
To prevent that circularity, preserve who asserted the proposition and what evidence they supplied. A complainant may have a shipping receipt, a photograph, a firsthand observation or only an inference from an unsuccessful search. Those are not interchangeable inputs. A later reviewer may accept part of the account and reject another part. A conclusion that a box was misplaced does not automatically confirm an allegation that its contents were intentionally destroyed, even if both concerns began in the same complaint.
Consider an invented resolution letter stating that D22 was located in a different storage room. That changes D22's location status. It does not automatically establish whether access had been improperly delayed, whether the original search was adequate, or whether other files remain unaccounted for. Equally, an unresolved case is not evidence that the institution has been cleared. The update should state the actual result rather than forcing every procedural outcome into a binary contest between complete vindication and complete condemnation.
This is particularly important when summarizing several cases. Imagine ten complaints: four founded, three unfounded and three pending under the fictional review system's stated labels. Four of ten complaints have founded dispositions at that snapshot. Among the seven resolved complaints, four have founded dispositions. Those are 40 percent and about 57.1 percent, respectively. Neither measures the share of all record losses caused by wrongdoing, because complaints are not a complete inventory of losses and the categories may concern different issues.
The arithmetic is simple; the editorial mistake is not. A reader seeing “most cases confirmed” may assume the denominator is all ten, while the writer used only seven. A reader seeing “only four confirmed” may overlook that three are unfinished. Writing the numerator, denominator and status date into the sentence is often more illuminating than choosing an adjective. It also makes later updates straightforward when one of the pending cases changes status.
7. Reconstructing a fact is not recreating a lost document
The National Archives' July 2023 account of the NPRC fire describes the July 12, 1973 destruction of roughly sixteen to eighteen million military personnel files and the use of auxiliary records to reconstruct service information. It identifies other record series, including morning reports and unit rosters, as useful when original files are unavailable or too damaged. The example concerns a documented archival disaster, not intelligence-file destruction, and it shows the possibility of recovering particular facts without recovering the entire lost original. National Archives account of the NPRC fire and reconstruction.
In Alder, suppose the missing D23 was a meeting report. A room reservation records that a room was booked for May 6. A travel reimbursement places one attendee in the city that week. A later letter says the sender discussed Project A at a meeting on May 6. The three items can contribute different pieces to an account. The reservation alone does not establish attendance; the reimbursement alone does not establish the meeting's agenda; the later letter may describe a discussion without preserving every decision.
A reconstruction should therefore be written at the level of the supported proposition. “A room was reserved” is different from “the meeting occurred,” which is different from “the meeting approved the project.” If several records jointly support the middle proposition, explain the connection. Do not present the resulting paragraph as a quotation from D23. The paragraph is the researcher's reconstruction, assembled from identified surviving sources. Readers should be able to distinguish that synthesis from the words of a document that no one has produced.
Copies can help, but their relationship needs attention. Suppose a surviving letter quotes two sentences from D23, then comments on them. The quotation may be evidence of what the letter writer represented D23 as saying. If another recipient preserved the same passage independently, that may strengthen confidence in the quoted wording. It still does not restore the report's unquoted qualifications, surrounding discussion or attachments. A two-sentence extract should not quietly become an imaginary full transcript.
Witness accounts require the same precision. An invented witness recalls seeing a delivery arrive at a building. That is not direct observation of the later use of its contents. Another witness who learned about the delivery from the first is not a second independent observer. A third person's contemporaneous receipt may support a transfer through a separate record path. The account can be evaluated without ranking witnesses by whether their memories fit the desired story. The relevant distinctions are opportunity to observe, timing, specificity and dependence.
A useful reconstruction can finish with a mixed result. The packet may establish the room reservation and payment, make a meeting plausible, and leave the supposed approval unresolved. That is not a failed investigation. It identifies the point at which evidence stops carrying the argument. Publishing that boundary lets a future discovery improve the account without requiring the entire narrative to be withdrawn.
8. Independence is about information paths, not document count
Suppose Alder document D13 describes a disputed incident. D15 summarizes D13. A later newsletter quotes D15. Three physical objects now contain versions of the story, but the central assertion has one identified information path. Counting them as three independent confirmations would exaggerate corroboration. Their agreement might simply show faithful copying. Conversely, a disagreement between the newsletter and D13 might show editorial compression or error rather than two witnesses describing different events.
Draw the relationships explicitly. An arrow from D13 to D15 means the latter relies on the former for this proposition. Another arrow from D15 to the newsletter completes the copying chain. Now introduce D08, an independently created equipment receipt. It may corroborate the presence of an item mentioned in D13 without corroborating the incident's cause or consequences. Independence is proposition-specific: a source can be independent for a payment date and dependent for an account of what happened afterward.
The practical editorial choice is to avoid using a broad phrase such as “multiple sources confirmed the incident” when the sources confirm different components. A more informative account says the equipment receipt independently supports delivery on the stated date, while the narrative description appears to originate with D13. That sentence preserves the genuine corroboration and the remaining limitation. It does not demand that every supporting source be completely unrelated to every other source in every respect.
This also changes how contradictions are investigated. If D15 changes a date from May 6 to May 8, first check whether it is describing the same event or a follow-up. If the newsletter copies May 8, its agreement with D15 adds little evidence for that date. A room booking for May 6 does not automatically resolve the dispute because the meeting might have moved. The next useful item could be a cancellation record, an attendee's contemporaneous note or a dated response, rather than a fourth retelling.
An archive can contain extensive paperwork and still have thin independent support for one consequential assertion. It can also contain only a few records that establish a narrow fact strongly because their origins and contents line up. The number of pages is therefore not a substitute for a map of evidentiary relationships. A hundred repeated summaries should not overwhelm one clearly identified contradiction merely by occupying more space in a search result.
For Alder's publication, the evidence note should show which claim depends on which item. It need not reproduce the entire investigative notebook. A concise description of the copying path and the independent receipt would let readers understand why one part of the story is better supported than another. That is a more honest use of space than listing every derivative mention in a long reference section designed to look conclusive.
9. Missingness changes the denominator and the uncertainty
Among Alder's twelve complete digital documents, four mention a safety concern and eight do not. For this calculation, assume the concern can be evaluated only from complete documents; the other twelve entries remain unclassified on that question. Four divided by twelve is one-third, or about 33.3 percent. That is the fraction of evaluable documents mentioning the concern. It is not the fraction of people harmed, experiments conducted, or all documents that would have mentioned it if every page survived.
The twenty-four-entry register allows a deliberately limited range. If none of the twelve unevaluable entries mentions the concern, the total would be four of twenty-four, about 16.7 percent. If all twelve mention it, the total would be sixteen of twenty-four, about 66.7 percent. These are extreme arithmetic bounds under the exercise's definitions, not a confidence interval or an estimate that the truth is equally likely anywhere between the endpoints. We have not supplied a random sampling design or a model of why material survived.
The observed one-third could be too high or too low for the full register. Perhaps records about problems were specially retained, increasing their representation among survivors. Perhaps those records were especially vulnerable to loss, decreasing it. Either hypothetical mechanism could fit the same observed count. Choosing one without evidence would convert uncertainty about preservation into a numerical conclusion about the underlying activity. The proper response is to identify the sensitivity, not to manufacture a correction factor.
Our local exercise also tested a common data error: replacing an unresolved entry's missing safety value with false. The checker rejected that substitution. A blank or unknown field is not the same as a recorded negative. If software silently fills every unavailable item with zero, a later chart will display an apparent clean record for material that was never evaluated. The chart can look precise because the uncertainty was removed during data preparation rather than resolved through investigation.
Document-level counts have another limitation even when complete. One report can discuss many people, and many reports can discuss one person. Four documents mentioning a concern do not establish four affected people. A serious account needs a separate, appropriately supported unit of analysis before making claims about individuals or events. The register here intentionally contains neither participant identities nor incident totals. Our calculations cannot answer questions that were not represented in the inputs.
That limitation should remain visible in the final story. A reader should be able to tell whether the percentage describes files, allegations, payments or people without hunting through an appendix. If the available data cannot support the quantity readers care most about, state that rather than substituting a nearby number. A less satisfying but correctly labeled measurement is more useful than a precise answer to a question the evidence never addressed.
10. Describe the search without claiming it could find everything
The Justice Department's FOIA procedural guide, listed as updated August 15, 2025, distinguishes the reasonableness of search methods from whether every responsive record was located. Its text also notes that inability to find each responsive item does not necessarily defeat an otherwise reasonable search. The chapter principally covers authorities through June 2023, with some later material; it is not an assurance about every subsequent case. This is background, not legal advice about a particular request or a judgment that any specific agency search was adequate. Justice Department procedural guide.
For the Alder investigation, write a search log that another researcher could understand. The first pass examined the project-name index and located D01 through D12. The second examined an identified correspondence series and located D13 through D15. A physical inventory established the locations of D16 through D18. Separate notices supplied the classifications for D19 through D21. These are invented steps in the exercise, not reported actions by an actual agency. They explain why the result is more informative than a bare statement that “the archive was searched.”
The unsearched territory matters. Suppose the team has not examined a regional office's accession list, private recipients' collections or an unprocessed carton mentioned in correspondence. The report should not state that all possible repositories were exhausted. It can state exactly which locations were checked and which leads remain. A specific lead is stronger than an unlimited demand to search everywhere, because it connects a proposed location to a reason the missing material might be there.
Search terms need context too. If Alder used an abbreviation in its accounting records and a full name in its correspondence, searching only one spelling would miss predictable material in this fictional system. But a failed keyword search would not establish that the papers were destroyed. The next action could be a different index, a date range or manual examination of an identified folder. The point is to diagnose what the search method could encounter, rather than treating a zero-result screen as a physical inventory.
A stopping decision should be reported as a resource and evidence decision, not disguised as omniscience. Perhaps the team has inspected every identified relevant folder available to it and recorded the remaining access limits. That can justify publishing a bounded account now. It cannot justify writing that no additional document exists anywhere. A clearly stated cutoff allows later researchers to add a newly processed collection without making the original account dishonest; the original claim was about the work actually done.
11. Make the uncertainty useful to the next reader
The most valuable final product may be a short list of unresolved propositions with different next steps. In Alder, whether D13 contained a final recommendation turns on its missing third page. Whether the $3,500 authorization-payment difference has an explanation calls for a financial reconciliation or additional transaction records. Whether D22 survives calls for location evidence. These questions should not be bundled into one dramatic statement about a hidden truth, because each could be resolved differently.
Priorities can be explicit without pretending to know the outcome. Suppose examining a newly identified index would require one unit of research time, requesting a known physical copy two, reviewing a restricted-file response four, and visiting a distant repository eight. Those invented costs do not tell us which lead will succeed. They do help distinguish an inexpensive, specific check from a broad, costly expedition. A project can pursue the first two while acknowledging that the latter questions remain open, rather than announcing premature closure.
Publish the evidence needed to audit the central claim, while considering privacy and relevance. A fictional receipt containing an unrelated employee's home address does not need to expose that address to establish a payment date. Removing an irrelevant identifier from the public illustration is different from pretending the underlying item lacked it. The retained research record and the public explanation should make the scope of any editing clear without releasing sensitive information merely to decorate an article with authenticity.
Corrections should also be proposition-specific. If D22 is later found, update its status and any totals that depended on it. Do not assume the discovery resolves D23, D24 or the contents of D13's missing page. If the recovered record changes the financial total, show which transaction was added and why it was not a duplicate. The purpose of a revision is to improve the evidentiary account, not to preserve the emotional shape of the first published version.
For example, locating a surviving copy of D22 would raise Alder’s confirmed surviving count from nineteen to twenty and reduce unresolved entries from three to two. The destroyed count would remain two. That update changes the inventory without telling us what D23 contained or resolving the separate payment discrepancy.
The strongest conclusion is often a set of bounded findings: these objects survive, these two have documented destruction, these three remain unresolved, and these particular claims are supported to different degrees. Such an account does not minimize lost evidence. It makes the loss legible. Nor does it grant an institution the benefit of an invented clean record. It reserves judgment where the necessary material is absent, while being specific about what is established. An archive with holes can still support serious history. It cannot responsibly support whatever the writer wishes had been on the missing pages.