AI uses for records
Document authority retrieval dataset: the records that outrank a date
By SourceX Editorial · Updated
9 sections
Short answer
A document authority retrieval dataset labels, for each question and date, which version governs and which record proves it: approval status, the owner's designation of the current version, supersession links and effective dates, each kept as a separate field. W3C's browser-specs list shows how often the newest text is not the authority: in 49 of 61 multi-level specification series, a newer level is listed but an older one is current.
Retrieval
Document authority retrieval
Key takeaways
- Label the governing version for each question, reader and date, and cite the record that makes it govern; a date is never enough.
- Keep approval status, owner designation, effective period and successor link as separate fields with their own sources.
- In W3C's browser-specs list, 49 of 61 multi-level series list a level newer than the current one: 26 because the newer levels are amendments, 23 because an older full level is designated.
- In all 49 the current level is the older one; its W3C status is more mature than the newest level's in 40, the same in 9 and less mature in none.
- Run a grounding test on any sample, and report how often the newest candidate governs so the set's value over a date sort is visible.
What is a document authority retrieval dataset?
It is a set of questions, each paired with every version of the documents that could answer it and a label saying which version governed for that reader on that date, together with the record that proves it. The proof is never the date alone. It is a status value, the owner's designation of the current version, a link from a replaced document to its successor, or an effective period, each kept as its own field.
Retrieval systems rank by how closely text matches a question, and many break ties by recency. Both habits fail in the same place: a draft revision, a narrow amendment or an unadopted new edition can match a question better than the approved text and carry a later date. A document authority retrieval dataset gives a model the cases where that happens, with the evidence needed to get them right, and the cases where it does not, so the model learns to check rather than to distrust everything new.
This page measures the problem in a public registry that records authority explicitly, then turns the measurement into a map of traps and a grounding test a buyer can run on a sample. Closely related questions, such as how to score retrieval of the controlling version or how to label relevance and authority in one pass, are answered on the pages linked at the end.
Why approval status, owner and effective date need separate fields
Because each answers a different question, and a single current flag answers none of them well. Approval status says whether a body accepted the text and at what level. The owner's designation says which version the responsible group wants people to use. The effective period says when the text applies, and the supersession link says what replaced it. Collapse them into one flag and a model cannot tell an approved document that is not yet in force from a draft, or a replaced version that still governs older work from one that is simply wrong.
W3C's process shows why replaced does not mean void. It allows W3C to declare a Recommendation superseded when a newer version exists that W3C recommends for new adoption, and to mark one obsolete when it lacks the market relevance to keep recommending it, with the possibility of restoring it later (both paraphrased from the process document). A superseded text can still be the right answer for a system built to it, so the dataset needs the status, the successor and the scope of each, not a yes or no.
Audience matters as much as status. Microsoft's documentation for SharePoint describes libraries in which minor versions, the drafts, stay hidden from people who can only read until a major version is published (paraphrased). An index built with an editor's permissions sees drafts that most readers never saw, and the newest file version in an export may be one of them. Every label should therefore say which readers it applies to as well as which date.
What W3C's specification list shows when newest and current part ways
It shows the newest text losing more often than it wins wherever several levels of a specification are listed side by side. The browser-specs project in W3C's GitHub organization keeps a curated, machine-readable list of web specifications that names, for each series, the level to treat as current. At commit 5fbea4e, read on 27 October 2026, the list holds 831 specifications in 737 series, and 61 series list two or more levels. In 49 of those 61 (80.3 percent), a level newer than the current one is listed; only 12 have their newest level as current.
The reasons are recorded, not guessed. In 26 of the 49, every newer level is what the project calls a delta spec, written as changes over the previous level, so it cannot govern alone. In the other 23, a full newer level exists, yet an older full level is named current. The project's own default rule would have picked the newer level in each of those 23, so every one is an override held outside the version numbers; the build script takes the current level from the W3C API unless a manual flag in the source list says otherwise.
Approval marks point the same way as the overrides. In 40 of the 49 series the current level has a more mature W3C status than the newest level, in 9 the two share a status, and in none is the newer level more mature. The current level is a full Recommendation in 15 of the 49, and 11 of the newer levels have no published snapshot at all, only a nightly draft: 10 Editor's Drafts and one unofficial proposal.
Retirement and addressing add two more traps. Of 71 discontinued specifications, 50 carry obsoletedBy pointers (54 in all, every one naming a specification in good standing on the list) and 21 carry none. In one series, Compositing and Blending, the current level (Level 2) has no published snapshot yet while Level 1 has one, the situation in which the project's documentation says the unversioned TR address may return an earlier level. Across the list, 433 specifications have a published snapshot and all but one of them also point to an Editor's Draft, a separate working document with its own status; 58 of the 433 (13.4 percent) are Recommendations.
The knowledge authority map: seven traps and the signal that settles each
Each trap is a way the newest or most similar document misleads, and each has a field that settles it. The map below pairs them with the count from browser-specs and an enterprise analogue. The counts show the traps occur in a carefully kept public registry; the analogues and labels are illustrative, and a company would replace them with its own document types.
A dataset should carry at least one labeled example of every row, so that a model is tested on each trap rather than on the most common one:
| Trap | Signal that settles it | browser-specs count | Enterprise analogue (illustrative) | Label in the dataset |
|---|---|---|---|---|
| Newer text is an amendment, not a replacement | Composition: full, delta or fork | 26 series where every newer level is a delta | A policy addendum that rewrites two clauses | Read with base |
| Newer full version not yet designated | Owner's current-version pointer | 23 series with an older level named current | Version 5 circulated for comment while version 4 is in force | Older version governs |
| Draft newer than the approved text | Approval or maturity status | 11 newer levels with only a nightly draft | A minor version visible only to editors | Draft, not citable |
| Temporary branch with proposed changes | Fork marker and base pointer | 18 forks in 2 WebAssembly series | A redline negotiated off the master template | Branch, not authority |
| Retired with a named successor | Supersession link | 50 specifications, 54 pointers | A procedure marked replaced by a new number | Follow the link |
| Retired with no successor | Retired status and no link | 21 specifications | An archived help article with no replacement | Abstain or escalate |
| Stable address serves an older version | Versioned address or snapshot ID | 1 series | A latest link still pointing at last year's file | Cite the dated version |
The evidence-grounding test: does every authority label cite a record?
Only if the label can be traced to a record that existed on the as-of date; otherwise it is a guess with a confident format. The test below is meant for a sample of labeled pairs, each made of a question, an as-of date and the version labeled as governing.
At a glance
The evidence-grounding test: does every authority label cite a record
A pair that fails stays in the set as unresolved instead of being forced to the newest candidate. The seven checks:
- Grounding: the governing version is justified by a record other than its date or file name, such as a status value, an owner's designation, an approval entry or a supersession link.
- Separation: approval status, owner, effective period and successor sit in separate fields, each with its own source and timestamp.
- As-of discipline: the label uses only records that existed on the as-of date; an approval granted later is not read back into an earlier question.
- Audience: the label names the readers who could see the governing version, since drafts and restricted pages differ by permission.
- Composition: amendments carry the read with base label and point to their base; no amendment is labeled as governing alone.
- Abstention: a question whose document family has no designated current version, or whose only candidate is retired without a successor, carries abstain or escalate rather than the newest candidate.
- Recency baseline: the sample reports how often the newest candidate is the governing one, so a buyer can see what the set adds beyond a date sort.
The strongest objection: the newest version is usually the right one
Often true, and the dataset should be built to measure it rather than deny it. browser-specs leaves out levels that have been fully superseded, so its 80.3 percent overstates how often newer levels lose in general: a registry that keeps every retired edition would show many more families where the newest edition is the one in force. A company policy library is likely to look like that, with one live version for most documents.
What the registry does show is the direction of every exception. All 23 overrides named an older, full level over a newer one, and no newer level outranked the current one on maturity. Newest unless a status says otherwise is a sound rule only when the status exists in the data, which is the case for keeping it. On a buyer's sample, the recency baseline settles the objection: if the newest candidate governs in nearly every pair, the set should be narrowed to its exceptions, sampled by trap with the rule written down before sampling.
How buyers would use authority records in training and evaluation
As retrieval and question-answering cases with the authority evidence attached. Each case gives a model a question, an as-of date and an audience, plus a candidate pool that includes the traps from the map, and asks for the governing version and the record that makes it govern. Grading checks both, so a right answer drawn from a draft does not pass, and abstention is scored as correct where the label says abstain.
For training, the same records teach a model to read the fields that carry authority before it reads the body text: a status line, a superseded-by link, an effective date in metadata. Cases should be cut by date, with the index rebuilt as it stood on each as-of date, because a later approval leaking into an earlier question teaches hindsight. Metrics for ranking the controlling version, and a full relevance-and-authority labeling guide, are covered in the linked retrieval benchmark and source authority pages rather than repeated here.
What company document systems add, and how to request them
Company systems keep the authority evidence that a plain export drops: approval workflow entries, major and minor version history, document control registers, effective-date metadata and links from replaced documents to their successors. A useful request asks for those records alongside the documents, for a window that spans at least one change of version in each family, and states which permission level each record reflects.
SourceX finds such records at businesses after a buyer describes the need, and holds no inventory of them. The request form collects scope, volume, language, format, timing and licensing needs, plus a dataset description, source systems, historical timeframe and de-identification requirements; a request for this dataset would name the document systems, the families in scope, the seven traps the sample must cover and the grounding test as the acceptance check. The company that holds the records approves the scope and terms and signs before anything is delivered.
Rights deserve their own line. SourceX's FAQ notes that owning records does not automatically give a company the right to license personal information or data belonging to customers, vendors or other third parties, and that permissions are checked with the company before any deal. Document libraries often hold vendor manuals, customer contracts and copies of published standards, which usually need separate treatment or exclusion. Nothing here implies that matching records exist today or that a sample will pass.
Methods note: browser-specs data, counting rules and limits
Data: index.json from w3c/browser-specs at commit 5fbea4e (committed 27 October 2026), read by a script (specs_authority.py) that loads only that file; no repository code was run. A series is the set of entries sharing a series shortname, and a multi-level series has two or more entries. Levels are compared as tuples of integers parsed from seriesVersion, with missing versions treated as zero (the project's own code compares the version strings; no series in the list orders differently under the two). The current level is series.currentSpecification. A newer level is an entry whose level is greater than the current level; forks share their base's level and are counted separately.
A series counts as newest-not-current when it lists at least one newer level, with the reason all newer levels are deltas when every one has delta composition, and full newer level exists otherwise. The default rule is the one in the project's compute-currentlevel.js (the highest level that is not a delta or fork); a series where the listed current level differs from it counts as overridden. Maturity compares TR statuses on this order: Recommendation, Candidate Recommendation Snapshot, Candidate Recommendation Draft, Working Draft, First Public Working Draft, no snapshot. Percentages are rounded once, from exact counts.
Limits: browser-specs is a curated list of specifications relevant to the web platform that leaves out superseded specs and fully superseded levels, so its shares describe this list, not enterprise archives. It holds no dates, so effective periods were not measured, and it records W3C status, not adoption. The enterprise analogues and labels in the map are illustrative. Not measured: errata, translations and how often readers cite the wrong level.
Frequently asked questions
What is a document authority retrieval dataset?
A set of questions paired with every candidate version of the documents that could answer them, labeled with the version that governed for a given reader and date and the record that proves it. The proof is a status, an owner's designation, a supersession link or an effective period, never the date alone.
Why is the latest document not always the authoritative one?
Because a newer document can be a draft, an amendment that only changes part of the base text, a temporary branch or a new edition its owner has not adopted. In W3C's browser-specs list, 49 of 61 multi-level series list a level newer than the one named current.
Which metadata fields signal document authority?
Approval or maturity status, the owner's designation of the current version, links from replaced documents to their successors, the effective period, the composition (full text, amendment or branch) and the audience that could see the version. Keep each in its own field with its own source and timestamp.
How do you test whether authority labels are reliable?
Check that each label cites a record other than the date, uses only records that existed on the as-of date, names its audience, points amendments to their base and abstains when no version is designated. Report how often the newest candidate governs, so the set's added value is visible.
Can a company license its document library for AI retrieval training?
Sometimes, but owning the files does not settle the right to license them. Libraries mix the company's own documents with vendor manuals, customer contracts and copies of published standards, and personal data needs separate review. The company approves scope and terms before any record moves.
Sources
- browser-specs index.json at commit 5fbea4e lists 831 specifications in 737 series with their current level, composition, standing, successor pointers and TR and nightly statuses; the source of every count on this page. Source
- browser-specs documentation defines the current specification of a series, delta and fork composition, standing and obsoletedBy, notes that the unversioned TR URL may return an earlier level, and says fully superseded levels are not listed. Source
- browser-specs computes the default current level as the last level that is not a delta or fork spec unless a level is explicitly flagged. Source
- browser-specs updates the current specification from the W3C API unless its source list imposes a specific level. Source
- The W3C Process Document lets W3C declare a Recommendation superseded when it recommends a newer version for new adoption, or obsolete when it lacks market relevance (paraphrased). Source
- SharePoint libraries can keep minor (draft) versions hidden from people who can only read until a major version is published (paraphrased). Source
- SourceX's request form: datasets are sourced on request, not held in stock, and the form collects scope, volume, language, format, timing, licensing, source systems, timeframe and de-identification needs. Source
- SourceX's FAQ says owning records does not automatically give the right to license personal or third-party data and that permissions are checked before any deal. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.