ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
UA
craft · 24 min read

Use a Source Span, Not a Confident-Sounding Answer

An archive's finding aid says that a collection documents a community radio project. An assistant tells a visitor that the collection contains recordings for…

An archive's finding aid says that a collection documents a community radio project. An assistant tells a visitor that the collection contains recordings for each year the project ran. Its citation opens the correct overview. But the overview only identifies the subject and broad period covered. A separate container note shows a few years of program schedules and no recording inventories. The source is real. The claim isn't supported by the span it points to.

A useful answer needs a piece of evidence that bears its actual weight. A document title, section number or search term match can help locate a source. It cannot substitute for reading which collection component a passage describes, whether it names formats or merely subjects, and what it leaves open. A good evidence review tests the claim against the relevant part of the source rather than trusting its citation.

All collections and finding aids in the worked examples are invented. This workflow is an editorial exercise for publicly shareable source descriptions under appropriate permission. It does not authorize access to archives, interpretation of access restrictions, digitization of source material, or copying real personal records.

AI disclosure: Drafted with AI assistance. All finding-aid text, holdings and answer examples are synthetic.

Understand what a finding aid can and cannot describe

A finding aid is a description of archival materials and how they are arranged, with detail that varies by institution and part of the collection. A collection summary, creator note, series scope, container listing and access note are different kinds of statements. The Library of Congress Encoded Archival Description resource gives background and documentation for marking up finding aids. It is a useful entry point for understanding the hierarchical descriptive setting; an encoded structure does not guarantee any particular institution itemized every record.

If an overview identifies a creator's broadcasting years, it may describe provenance or topic coverage. It doesn't necessarily prove an item was made or placed in the collection in each year named. A container listing can identify a particular record type in one unit and not say anything about a related collection. Avoid treating all headings as words of one flat description. A statement belongs to the part of the document where it appears.

Ask what claim the person actually needs. “Does this finding aid mention programs in 1977?” is a text observation. “Does the archive have the broadcast recording?” requires evidence of a holding with that format and date. “Can I listen to it online today?” adds access and serving state. The source can support the first while not settling the others. An assistant can help locate the next appropriate evidence without pretending all three questions are answered by a link.

Use controlled language to express this difference. “The series description lists paper program schedules” is a positive claim about what was described. “Audio recordings were not described in the section reviewed” is a bounded absence claim. “The archive holds no recordings” is substantially stronger. The first two can coexist in a good answer without forcing the third.

Write the answerable claim in a form a second reader can test

Read a draft sentence and ask what would make it false. “There are programs from 1977” might describe physical recordings, printed schedules, scripts, photographs or digital exhibits. Choose a precise subject and predicate before accepting its citation. Under an invented aid, the supported claim could be “The container list identifies program schedules dated 1977 in Folder 4 of the station paperwork series.” A second reader can inspect that exact text.

One evidence record can list intended answer, source edition, descriptive level, named scope, relevant span, nearby constraint, and support state. Where exact quotation isn't necessary, keep a short paraphrase with an evidence pointer. A space for competing interpretations lets a reviewer explain why an adjacent description doesn't answer the question. The reference should be a working route to the evidence, not a badge attached after the conclusion.

Make action claims separate. “Series 1 has a listing for these schedules” does not say a venue is currently open, the material is digitized, or a reader can order a copy. If a public finding aid points to a separate process, use the approved link or explain what should be checked. Never infer access rights from a successful load of a descriptive web page. Access evidence and description answer different questions.

The rule is not that every sentence needs a footnote. A sentence such as “Here is what the reviewed aid says” is an introduction. A claim about format, holdings, date, scope or access earns a trace. Keep the audit proportionate to the assertion. A helpful source span states the relationship actually being claimed rather than accumulating a long bibliography for a simple fact.

Compare two plausible paragraphs before deciding which one governs

Consider this brief fictional finding aid:

Evidence spanVisible words and relevant level
A, background noteThe Meadow Signal workshop produced radio programs and live interviews from 1975 through 1982.
B, series listingFolder 4, station paperwork, includes program schedules for 1977 and 1978.
C, format note for Series 1The described materials in this series are paper files.
D, related collections noteA separate oral-history collection is described elsewhere.

Ask, “Does this aid show that the 1977 broadcasts are held as audio recordings?” Span A mentions the programs and the broad time range. It does not say audio was transferred to this collection. B shows a named 1977 holding as schedules. C describes the materials in its own series as paper, not every archival holding at the institution. D offers a place to look, not proof an audio recording resides there. The supported response is that this aid identifies 1977 schedules in the paper series and doesn't establish that the requested recordings are here.

Try a narrower question: “Where is the 1977 schedule listing?” B directly supports a useful answer. Mentioning A alone would give context rather than a location. C doesn't prove there are no digital reproductions of paper material. Avoid using one paragraph to settle unrelated questions because it contains a familiar keyword. Each piece of evidence can be right within its boundary and irrelevant outside it.

Now ask where to look for related material. D supports a bounded next step: consult the separate aid to see what it describes. Do not imply that checking a reference and finding a collection title verifies specific contents. If the next source is outside your authorized access, say what you couldn't check. An assistant can stop with a useful route and a plainly named unknown.

A review exercise can shuffle the order of the four spans while retaining labels and hierarchy. If the output keeps choosing the first one mentioning 1977 no matter which action the prompt asks for, the fixture reveals a relevance failure. If it always ignores background, it will fail a question about the program's described years. A correct evidence choice responds to the predicate, not to a predetermined ranking of headings.

Anchor a span without losing what surrounds it

The W3C Web Annotation Data Model describes selecting portions of resources, including selectors that use an exact quote with surrounding text. The idea can help a small audit preserve where an extraction came from. A source ID, title or page link can locate the resource. An anchor and brief preceding or following cue can distinguish which of several repeated phrases the reviewer should read. You don't need to implement the full technical standard to make the pointer useful.

Include the descriptive level or parent heading when it materially limits interpretation. A bullet “Recorded interviews” may appear under a Related Holdings section and mean something different from a bullet under Box 7. The same words can appear under two series labels, one naming print transcripts and another describing media reproductions. Keep the words that attach the fragment to its proper object. If you can't verify the attachment in the authorized source view, mark the association uncertain.

A short span can contain the answer but not a needed limitation. Suppose a date range in a series title is qualified by a footnote noting that the exact years are not established for undated material. If you only quote the title, a reader could assume it applies precisely to every item. The evidence pointer should flag the relevant note without copying an entire aid. Keep all source use within approved access and reuse rules.

Check the pointer with a second reader. They should reach the relevant passage and explain whether it establishes the actual claim, not just find a related word. If the source has changed at the same URL, identify the reviewed edition or say the original is no longer reproducible. A still-working link to a different text is not a faithful handoff.

Keep hierarchy from turning subjects into holdings

A parent collection can cover many topics and several record types. Child descriptions may list only a subset. A headline claiming that everything about the creator is in one box confuses description with containment. On a real finding-aid site, search snippets can repeat the collection title above each match. That heading should not cause an answer to treat a result from a sibling series as proof about the series requested.

Construct a fictional example with two child groups. Series One contains paper schedules for 1977 and 1978. Series Two describes donated equipment catalogs from 1980. The overview mentions that the program continued into 1982. An answer saying Series Two holds 1977 schedules because its parent list contains both ranges has broken hierarchy. An answer saying the 1982 collections include equipment catalogs because 1980 is under a broader date label also overstates the evidence.

Document where relations come from. The word “part of” might describe a section relationship rather than an intellectual connection. The text “see also” can point to another collection without making it subordinate. Do not convert a reference into “this box includes a copy” unless the span supports it. The appropriate output might be a separate mention of where to investigate.

If an excerpt loses hierarchy during copying, don't invent it from indentation that may be an export artifact. Inspect the authorized original and its structural labels. A plain text caption alone may be insufficient. A good abstention can say “the text names the schedules, but the supplied excerpt doesn't establish which series contains them.” That is a narrower and more useful result than refusing to say anything or guessing the nearest heading.

Distinguish absence of evidence from evidence of bounded exclusion

Search misses can arise because terms differ, the aid is not itemized, a page wasn't fetched, or a system only indexes some portions. “Did not find” is a statement about a defined search and scope. It should not silently become “does not exist.” The same care applies to a text description that explicitly says a certain part contains no audio. It may establish a stronger localized exclusion without settling everything at the institution.

Use paired synthetic spans. One says “Records in Series 1 consist of paper files.” Another says “This folder does not include discs listed in Series 4.” The second rejects one specific relation and names another reference. It does not certify dates in Series 4. A question about recordings anywhere in the institution requires additional scope. Do not let a negative under one heading travel to another scope because the adjective is the same.

If only selected excerpts were supplied, make the boundary prominent. You can state “I do not see a recording inventory in the admitted excerpts” and explain what evidence could establish the requested result. A follow-up might name the related aid or the appropriate official staff channel if authorized to recommend it. Don't claim you contacted anyone or searched all holdings. A good next step doesn't require overstating what already happened.

Auditors can grade trained abstention separately. In an explicit positive fixture, answer with the exact supported scope. In a limited unmatched fixture, state the known context and the unresolved holding. In a localized explicit negative fixture, preserve where the exclusion applies. Treating all three as just unanswered throws away the core skill. Evidence support has different levels and they can be expressed naturally.

Work an answer where span and outcome match but the reason doesn't

Suppose the synthetic answer says, “Folder 4 lists 1977 schedules, so audio was never made.” The first part is supported by B. The second invents a historical conclusion from the absence of recordings in one described part. When reviewing a sentence joined by so, therefore or which proves, test the inference rather than rating the whole sentence by the accurate fragment. An unsupported inference often supplies the headline a reader remembers.

A different response says “The series is paper, so the 1977 schedules cannot be available online.” A paper original can have a digital surrogate. The span describing original medium doesn't establish whether anybody digitized some items. A properly bounded answer says the aid describes paper schedules and the supplied evidence doesn't establish online availability. If an official link explicitly points to a collection of surrogates, verify its actual scope separately.

Try a case where a new eligible source does provide specific access information. A fictional access note names a reading-room request process for a defined group. Now an answer can describe that process under the fixture's assumptions without claiming a request was filed or approved. Do not penalize successful grounding because previous fixtures taught caution. The benchmark should reward appropriate confidence when the required evidence is present.

The NIST Generative AI Profile discusses confident erroneous outputs and the risk of relying on them. A believable citation can make those outputs more persuasive when readers do not check whether the selected text supports the whole claim. Independent source review is a practical response. The framework does not certify any particular extraction or perfect protection from overreach.

Check whether a date describes material, coverage or work by a creator

A date can sit next to a proper noun and still have several meanings. The papers might have been created partly in one year, describe events in another and have been transferred much later. A background note records the years a person worked. A date field on an aid may describe the bulk of the material rather than the endpoints of every item. Consult the displayed labels and qualifiers before assigning a date to an unresolved object.

For a synthetic guide, say the workshop operated from 1975 through 1982. Its descriptive note says the paperwork was principally assembled in 1980 and 1981. A folder title says 1977 and 1978 program schedules. Those are compatible statements with different scopes. A generated response asserting that a 1977 schedule was filed in 1977 lacks that evidence. A statement that the guide lists schedules for those years is safer and more useful to a reader seeking them.

Test two questions that share the same number. “Was the workshop active in the time period named in the background note?” calls for that note. “Which folder names 1977 paperwork?” calls for the container listing. If a different sentence names the acquisition year, don't let it be mistaken for the event date. A short date-type label in the evidence register can prevent the most ordinary substitutions.

An explicit approximate date should remain approximate. “Circa 1977” represents an estimate in the described context, not proof the item belongs in a chronological sorting exactly at the year's beginning. Where date source and scope are uncertain, keep that uncertainty clear. Many useful archival answers can preserve “approximate,” “bulk,” or “undated” and still point the reader to a practical location.

Make a next step a recommendation with a verifiable prerequisite

People asking about a finding aid often care what they can do after reading the answer. They might want to look at the paper schedules, listen to an interview, compare series or confirm whether material is discoverable online. A source span can give a useful route to investigate a next step without saying it is already authorized. Think of source links as location evidence, not credentials or fulfilled requests.

An invented scope paragraph might say “Related: the Shore Station oral-history guide.” A responsible answer can recommend consulting that guide and checking its format descriptions. It should not tell a reader to expect the exact Meadow Signal interview dates unless it has verified them at the next source. You can convey why the reference is likely relevant without claiming it resolved the question. “Related collection named; content not verified” is an honest internal receipt.

If the reviewed material says conditions apply, state them as stated and send the reader to the official source for a live question. A finding aid can be publicly read while items require a documented process; the desk can confirm what is available and who may access it. Do not turn a general eligibility statement into personalized authorization. This is a question about evidence and scope, not instructions to bypass a collection's restrictions.

Give a model instructions to distinguish proposed next steps from observed outcomes. Quoting “request this folder through the archive's approved form” doesn't establish a request went out. If asked what remains to verify, it can identify the relevant form or permission boundary. The textual audit never needs to send an inquiry automatically. Any real contact or collection action needs its own appropriate authority.

Use a close reading when sources conflict, not a popularity vote

Two paragraphs can appear inconsistent because they answer different questions. A collection-level media note might say the group includes paper and audiovisual material. A specific series description says only paper files. Until you know whether they're talking about the same scope, neither needs to be wrong. If the question is about an exact folder, the broader statement cannot fill missing folder evidence.

Construct a second fixture. The general overview describes recordings among holdings, and the item list mentions a 1977 interview transcript in Folder 5. The question asks whether Folder 5 contains audio of that interview. It is tempting to join a known recording category to a matching interview title. The actual support is a transcript in the folder and recordings somewhere in the broader collection. An answer can say both, then explain why the association remains unverified.

Now put an explicit item description in the supplied list: an audio recording of the named interview under a different folder. That settles where it is described, not whether the user can listen without approval. A practical audit sees the evidence added and permits a more direct answer. If a system keeps answering “not enough information” despite the explicit listing, the fixture exposes an unhelpful handling problem. Tailor grading to the available evidence rather than always preferring caution.

When two authorized sources do directly disagree about the same scope, document both and identify plausible version or descriptive differences only as hypotheses. A current item authority might settle the contradiction if the proper source owner confirms it. Don't decide based on which paragraph contains more detail or which answer feels more useful. Preserve the genuine conflict as an outcome if you can't establish precedence.

Design the scope check to survive reformatting

A finding aid can be rendered as a website, a downloadable document, an excerpt pane or a structured file. Hierarchical relationships might be explicit in one and compressed in another. A table of search matches can repeat a title and omit parent labels. If your tooling flattens the structure, retain a reference to the original view rather than assuming the order of copied paragraphs carries all context.

Test a fixture where the same phrase appears twice, once in the actual holding list and once in a related group. A search-only reader may return both; a scope-aware reviewer should pick the span matching the question. Keep the exact authoritative parent label and any note necessary to understand the result. Someone reading the report later should be able to distinguish two relevant-looking occurrences without relying on your memory.

Watch out for document page numbers that don't match file positions. Page 12 printed on the source might be the 15th page in a reader because front matter is unnumbered. A working link with an incorrect notation can waste time or produce a mistaken substitute quotation. Record the notation you actually used, and add a short cue pointing to the heading where possible. This is especially useful for synthetic test definitions that might be generated in several output formats.

Verify the actual path a second reader will have. If the working view was private and the report reader only has public fragments, they might not be able to reproduce the full check. Keep unpublished or protected evidence in its authorized scope. An outward answer can be clear about the limited evidence that can be shared rather than exporting someone else's access. Provenance means preserving the relationship responsibly, not making all source text public.

Review a short paragraph by splitting inference, evidence and next action

Take a draft that reads: “The 1977 interviews are accessible in the papers series and digital copies can be requested. The station operated from 1975 through 1982, so this collection covers every year of broadcasts.” A novice reviewer might just look for the dates and the station name. An evidence-focused review sees several independent tasks: access, format, request availability, creator scope and comprehensive holding coverage.

Rewrite it from the supplied source: “The guide describes the station as active from 1975 through 1982. It lists paper program schedules for 1977 and 1978 in Folder 4. It does not establish that interview recordings are held or that digital copies can be ordered.” If an authorized related guide is known, name it as an unverified next search location. This version is not merely more cautious; it is more precise about the support for each sentence.

Find the strongest action a reader can take without the missing evidence. They can use the folder listing to locate the description. They can read the related collection guide, if appropriately available, to look for a recording description. They may need an official access answer before planning a visit. None of these suggestions requires speculating that a particular historical item exists. The answer stays useful by preserving actual next steps, not by pretending the evidence is complete.

A prose check can be assisted with a simple claim ledger rather than grading every word. List statements about the material, statements about description, statements about access, and next-step hypotheses. Assign each to the most relevant observed span or mark unsupported. The final reader may not need all those columns. The reviewer benefits because it is much easier to spot when a citation for source description has been borrowed to support access.

Give the evidence a date and a useful unknown state

A source span may once have been appropriate and become unavailable. Keep separate observations for source date, current verification date, and applicability where known. If a published aid uses a stable address while contents are replaced, compare the visible current revision before assuming an older cited paragraph still appears there. Retain only the snapshots and source details appropriate for the organization and permissions.

A report that says “source unavailable at review” is different from one that says “span found and does not support the assertion.” The first can often be resolved by finding a valid authorized copy. The second may require narrowing the conclusion or finding a different relevant source. A third state, “support found, surrounding limitation not yet verified,” deserves its own route. Don't treat every uncertainty as a generic missing citation.

Use a worked update with synthetic authority. The original aid identifies Folder 4 schedules; a revised aid renumbers them to Folder 9 after a rearrangement. If the published result must be current, verify the new relationship and revise the locator while retaining the reason for historical differences. Do not claim the schedules were new discoveries just because their display number changed. Nor should you silently use one path to assert the other under the old edition. State which version the answer covers.

For a one-time internal assignment, a concise uncertainty note can be enough. For a recurring editorial product, record what should trigger recheck: new source edition, changed access notice, a mismatch in parent heading, or unreliable locator. The next reviewer can narrow the work to the changed part. Proper maintenance supports confident answers when the evidence remains intact and honest holds when it doesn't.

Test with distractors that are relevant enough to be tempting

A poor test offers one archival paragraph and two lines about unrelated activities. Any system can appear excellent by looking for the right noun. A better synthetic fixture provides two plausible paragraphs about the same actor and period, with one establishing the requested relationship and the other supplying background. Ask reviewers to choose the span and explain why the alternate paragraph isn't enough. This reveals a common reasoning gap.

Use one case for positive support and one for a bounded no-match. In the positive case, the identified folder listing names the schedule. In the no-match case, background still discusses broadcasts while the archive list omits known recording evidence within the reviewed excerpt. Then add a competing description in a different series. A cautious assistant should cite the specific source when it can and describe the real gap when it can't. Vary the source order without changing scope so the test doesn't become a memorized first hit.

Evaluate whether the chosen span supports the requested claim, not whether the response was pleasant. A correct answer with a wrong reference is difficult to audit even if it landed on the correct conclusion by chance. A correct reference attached to a broader phrase is also unsafe. List both as distinct failures and plan targeted repairs. A result can be improved by fixing selection, scope interpretation or final wording, only one of which may involve a model change.

Keep the score small and diagnostic. A test of six artificial cases can show which seeded mismatch was caught. It doesn't establish institutional archival accuracy or predict real retrieval coverage. Describe the source forms tested and what reviewers could check. An honest finding with a narrow method is more useful than a perfect-looking score that never says what the system had to prove.

Record negative and positive tests on the same navigation path

An efficient reviewer doesn't just verify a promising result. They also show what the adjacent text does not establish. Start with the overview and record its description of activity. Follow the supplied container listing to the printed schedule. Then check whether the applicable child description actually names the medium requested. If it doesn't, document the boundary instead of searching every other collection by guesswork and saying no material exists. The same route should support both the known and the unknown parts of the answer.

A small audit sheet could have row labels such as “overview supports activity years,” “folder listing supports printed schedule dates,” and “no recording entry observed in supplied series scope.” Those are distinct facts. A reader can reproduce each with the same branch from the document root but different conditions. When the competing answer points to the overview for exact format, the audit makes clear why that citation isn't interchangeable with the container note.

Test a found passage that explicitly rejects one container claim. Suppose the invented note says “These are paper transcripts; recorded interviews, if present, belong to a separately described accession.” The source supports the paper format and the fact that another descriptive route is indicated. It doesn't guarantee recordings exist in that accession. If the note instead named an accession and cataloged exact audio items, it would be appropriate to report that additional supported evidence. A rigid workflow that always labels related notes irrelevant would miss the distinction.

Supporting proof can require more than one span when the claim is inherently composite. To say a described folder of paper schedules sits within the station-paperwork series, the container label and parent heading can both matter. To say whether that folder is available digitally, a separate explicit source is needed. An evidence register should allow multiple supporting pointers without letting nearby context serve as a general substitute for the missing fact.

Verify how the summary will read outside the original finding aid

A concise answer may be copied into a reference card for visitors who don't see the collection's hierarchy. Words such as “the series,” “here,” or “below” were clear in the reviewing interface and ambiguous on a slide. Use scope labels that travel. “The station-paperwork series listed by the reviewed guide” is more portable than “this collection has them” when the difference in coverage matters.

A report title can also accidentally universalize a narrow finding: “Where to Hear the 1977 Broadcasts” promises audio even if the text only identifies printed schedules. Review the snippet, heading and proposed next link together. Fix titles to describe what was located, not the hopes that inspired the search. If the article only maps sources for future research, call it that. A polished heading shouldn't make the background note into direct evidence of archived recordings.

Keep an approved way for a reader to redirect their question. An official archive link ordinarily makes more sense than asking them to send an entire personal request history to a public chat. If the site supplies a specific locator in the reviewed scope, preserve it. If contacts are outside the supplied evidence, say which official aid remains unverified. Authenticated procedures or restrictions are separate from the support audit and shouldn't be guessed to make instructions feel complete.

End by asking someone unfamiliar with the sources what they think was actually checked. If they answer “a survey of all audio holdings” while the reviewer only read one series, improve the wording. The purpose of preserving spans is not to make the writer look careful. It is to let the reader know which evidence they can use and where a further check remains necessary.

Hand off a provisional answer without disguising what was not checked

A good final answer might read: “The reviewed station-paperwork series lists printed schedules for 1977 and 1978 at Folder 4. The overview describes the radio project's activity during that period, but it does not document audio holdings for those broadcasts. The referenced oral-history guide is a possible next source and wasn't checked.” A reader can see the found material, the unavailable conclusion and a practical route for further verification.

Keep the unverified next source out of the confirmed finding. A corresponding audit note can name which exact span supports each important assertion. It can also identify what would change the answer: a current itemized recording listing with the relevant dates and collection relationship. Put that missing evidence into a narrow next step. Avoid asking a colleague to verify everything or to interpret a generic title as a green light for a new action.

This process is useful even when no AI suggests an answer. Human writers also see a familiar keyword and use it as proof of something it didn't say. Choose the specific relationship to establish, read the actual span in context, and keep what remains unverified visible. A source reference should help another reader reconstruct a conclusion, not simply make it look more confident.

Frequently asked
What is Use a Source Span, Not a Confident-Sounding Answer about?
An archive's finding aid says that a collection documents a community radio project. An assistant tells a visitor that the collection contains recordings for…
What should you know about understand what a finding aid can and cannot describe?
A finding aid is a description of archival materials and how they are arranged, with detail that varies by institution and part of the collection. A collection summary, creator note, series scope, container listing and access note are different kinds of statements. The Library of Congress Encoded Archival Description…
What should you know about write the answerable claim in a form a second reader can test?
Read a draft sentence and ask what would make it false. “There are programs from 1977” might describe physical recordings, printed schedules, scripts, photographs or digital exhibits. Choose a precise subject and predicate before accepting its citation. Under an invented aid, the supported claim could be “The…
What should you know about compare two plausible paragraphs before deciding which one governs?
Consider this brief fictional finding aid:
What should you know about anchor a span without losing what surrounds it?
The W3C Web Annotation Data Model describes selecting portions of resources, including selectors that use an exact quote with surrounding text. The idea can help a small audit preserve where an extraction came from. A source ID, title or page link can locate the resource. An anchor and brief preceding or following…
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room