<!-- ARTICLE_BODY_START -->
1. The commission creates a decision problem
Austin Little's commission began with an expansive ambition: select knowledge worth carrying forward, across practical life, science, institutions, culture and the ordinary activities that make a life recognizable. The requested form was two fresh inventories, each containing two thousand coherent knowledge packages. That numerical instruction supplied a boundary. It did not supply a theory of everything worth knowing, a tested reconstruction curriculum, or a finding that two thousand is the correct size of a collection.
The work therefore required a series of editorial decisions before any large count could become meaningful. What would qualify as one item? How would the two selections differ? Would a book, a skill, a diagram and an entire field each consume one slot? How would familiar subjects avoid crowding out less immediately recognizable needs? What would happen when two apparently different titles turned out to describe the same learning outcome?
These questions are the subject of this article. They concern the design and interpretation of the selection exercise, rather than the technical architecture of a future preservation system. Storage formats, redundant copies and recovery procedures matter later. They cannot rescue a selection whose units are undefined or whose claims exceed its evidence.
The experiment is best understood in the everyday sense of a deliberately structured attempt to learn by doing. It produced editorial proposals and review evidence. Calling it an experiment does not automatically turn it into a randomized study, a blind comparison between competing models, or a demonstration that the proposed knowledge has been acquired and can rebuild anything.
2. The missing earlier list changes the claim
A previous two-hundred-item selection was referenced in the commission, but its contents were not available through the authorized recovery routes used for this work. The recovery record documents unsuccessful searches and the absence of a working authorized route to the historical local file. That evidence supports a limited statement: the earlier inventory could not be inspected here. It does not establish that the file was deleted, that its entries were poor, or that none of them overlaps the new proposals.
The new selections must consequently stand on their own. They cannot honestly be described as an expansion that preserves every earlier choice. They also cannot claim to be deduplicated against the unavailable original. A title remembered from a prior conversation would not be enough to reconstruct the full rationale, dependencies, permissions or intended scope of an earlier package.
This limitation affects evaluation. Suppose a later reviewer recovers the earlier list and finds a topic in both versions. That would not retrospectively prove a failure of the documented comparison process, because the older material was outside the available comparison set. It would create a new reconciliation task with new evidence.
A useful inventory makes this boundary visible instead of hiding it in a footnote. The historical selection remains a separate object, to be preserved unchanged if recovered. The new work is a fresh editorial version. Treating unavailable evidence as unavailable is part of the method, not an embarrassment to be concealed by confident language.
3. A package needs an intended outcome
The smallest consequential decision was the unit of selection. A topic such as water, mathematics or music is too large to tell a future reader what would actually be included. A single unexplained noun such as microscope is more specific, but still leaves nearly everything important unresolved. Does it mean the history of microscopy, a diagram of an instrument, an optics course, an operating manual, or the evidence needed to interpret an image?
The selection contract instead asks for a coherent knowledge package. Each proposal needs a title, a reason for inclusion, concrete proposed contents, dependencies, a relevant source lead and safety or rights qualifications. The package is still a proposal, but it should describe an identifiable learning or reconstruction task.
That makes comparison more demanding. Two entries with different titles may offer the same outcome, while two entries sharing a word may teach different things. Measuring temperature and interpreting a temperature calibration certificate belong to the same broad area, yet they answer different questions. Conversely, two maintenance passports that both contain an equipment index, service contacts and unresolved defects may be substantially the same editorial idea despite naming different appliances.
The unit also establishes a test for vague ambition. A rationale should explain why this particular package earns a place. Contents should describe what a future compiler would gather or create. Dependencies should identify what the learner must already know or obtain. If those fields remain generic, the selection has not yet progressed far beyond naming a subject.
4. The taxonomy was an intervention
The work was organized through a shared taxonomy before the full inventories were assembled. Practical rebuilding and daily life, science and understanding, health and institutions, and culture and expression were assigned separate editorial areas. Each area was divided into named categories with fixed quotas. The two sets were designed to be complementary rather than independently generated answers to an identical prompt.
That coordination has advantages. It gives culture and everyday dignity explicit space instead of hoping they survive a race among more obviously technical subjects. It also makes responsibility visible. A category has an authoring assignment and a quota, so missing work is easier to detect than in a single unstructured stream of suggestions.
But the taxonomy shapes the outcome. Once a category is allocated a fixed number of slots, its author must make decisions within that space. The resulting balance reflects an editorial design, not a discovered distribution of all valuable knowledge. A category with many natural subdivisions may feel constrained; another may require unusually fine distinctions to reach its quota without repetition.
Institutional collecting also begins with an articulated purpose. The Library of Congress describes its collection policies as setting scope, intensity and goals in relation to its mission. That is a useful precedent for making selection assumptions explicit, not an endorsement of this particular taxonomy. Library of Congress collection-policy introduction.
The right question is therefore whether the chosen structure served this commission transparently. It is not whether the structure emerged spontaneously from an impartial census of human knowledge.
5. Complementary sets cannot establish a winner
The labels A and B can tempt readers to imagine a contest. That interpretation would be misleading here. The sets were given different responsibilities and were edited to reduce overlap. A difference between their contents is partly an intended consequence of the assignment.
If one set contains a foundational mathematical package and the other contains an instrument-calibration package, the contrast does not show that one selector values theory while another values practice. It may simply show that the taxonomy placed those outcomes in different locations. Likewise, a set containing more care coordination cannot be declared more humane without considering what it was asked to include.
NIST's description of completely randomized designs illustrates a different kind of evidentiary structure: specified experimental units, assigned factor levels, a response variable and randomization. This exercise did not use that design. Nor did it define a measured outcome capable of ranking model performance. NIST/SEMATECH experimental-design guidance.
That distinction does not make the editorial work useless. It changes the question the work can answer. The inventories can show what a coordinated process proposed under stated constraints, which defects review found, and how those defects were repaired. They cannot isolate the causal contribution of a model, a prompt, an authoring role or a reviewer.
A future comparative study could be designed differently. It would need matched tasks, declared scoring criteria, independent outputs and an evaluation procedure suited to the claim. Those are possible next experiments, not events that occurred during this commission.
6. The relationship to Karpathy is specific
Andrej Karpathy's April 4, 2026 LLM Wiki idea file describes a maintained knowledge layer between raw sources and a user. It is a pattern for organizing and developing knowledge with an assistant, not a supplied preservation corpus. Karpathy's author-published idea file.
The relevant comparison is the location of the decision. A maintained knowledge system needs material to work on. This commission asks an earlier question: which knowledge packages should be proposed for inclusion, and what evidence should accompany that choice? Organization and selection meet at the source boundary, but they are not interchangeable tasks.
A neatly organized collection can still have serious omissions. An ambitious selection can still lack usable documents. Neither problem is solved merely by invoking a well-known person's name. The relationship should remain concrete enough that a reader can see what is being compared without inferring a collaboration, endorsement or shared authorship.
Austin Little commissioned this independent work. The inventories are this assistant's coordinated editorial proposals, developed through assigned authoring and review tasks. They are not attributed to Karpathy, to an unseen earlier author, or to independent models whose outputs were never actually compared.
That modest attribution is more useful than borrowed prestige. It directs scrutiny to the actual artifacts: the selection contract, the category assignments, the authored rows, the source registry, the replacement decisions and the final validation records. Those are the objects capable of supporting claims about this exercise.
7. A rationale should discriminate
A weak rationale says that a subject is important. Water is important, language is important, measurement is important. Repeating that observation under a hundred titles does little to explain why each package deserves a separate place.
A discriminating rationale identifies a particular failure, capability or dependency. A water-meter interpretation package can address confusion between a reading error and a change in consumption. A package about reference standards can address the loss of traceability when equipment changes hands. A garment repair package can preserve a distinct stitch geometry rather than merely promise longer-lasting clothes.
This matters when the quota becomes difficult. The easiest way to expand a list is to take one useful idea and substitute nouns: a maintenance passport for a bicycle, a maintenance passport for a generator, a maintenance passport for a receiver. Some local records are genuinely necessary, but necessity alone does not make each one a distinct learning package in a constrained selection.
The method therefore asks what new reasoning the next entry contributes. Does the learner encounter a different mechanism, a different interpretation problem or a different prerequisite? Or does the package simply move an existing template to another object?
The answer is not always binary. A specialized domain can change a generic procedure enough to justify its own unit. But that difference should appear in the contents and dependencies, not be assumed from the title. A useful reviewer can demand that the author either make the distinction explicit or release the slot for something genuinely different.
8. Contents expose empty specificity
A title can sound impressively precise while its contents remain vague. Consider a hypothetical package called Precision Bearing Restoration. If the contents are only diagrams, instructions and examples, the proposal has not said which bearing types, what failure evidence, what inspection limits or what expertise would be required. The specificity is largely cosmetic.
A more useful proposal might focus on reading bearing designations and distinguishing interfaces needed for replacement. Its contents could include code fields, seal types and supplier-confirmation records. Its dependencies could require the actual component markings and manufacturer documentation. Its safety boundary could exclude improvised installation of critical assemblies.
The second description is narrower, but it is easier to evaluate. A compiler can investigate whether the proposed materials exist. A reviewer can compare it with neighboring units. A future user can see what the package would and would not enable.
This discipline also prevents a common count inflation: treating a bundle of unspecified materials as a complete course. The inventory does not contain those diagrams or exercises merely because the contents field proposes them. The field is an acquisition or authoring brief. It establishes a target against which later work can be checked.
For this reason, detailed metadata should increase accountability rather than create an illusion of completion. Every additional promised component is another thing that would need to be found, created, reviewed and maintained before the package could become a usable educational object.
9. Dependencies challenge the self-sufficiency story
A list about preserving knowledge can drift toward the fantasy that enough documents remove dependence on people, institutions and equipment. The dependency field is a direct challenge to that story.
Many practical packages require more than literacy. They depend on suitable measuring instruments, known materials, current local rules, qualified instructors, safe workspaces or access to a laboratory. A document explaining how to interpret a water test does not supply the test. A diagram of a pressure system does not supply safe components or the competence to alter them.
Dependencies also include contextual knowledge. A repair technique may require identifying the actual material, knowing an object's history or understanding whether the work is culturally appropriate. A translation package may require community review, not merely a dictionary. A clinical communication package may require an individualized professional assessment.
Recording these requirements has two effects. It limits exaggerated claims about what a single package can accomplish, and it reveals connections among packages. A measurement unit may support several workshop units. A language-access unit may make otherwise inaccessible instructions usable. Those relationships can later guide the assembly of smaller learning pathways.
They do not prove that the pathways are complete. The inventory's dependency statements are proposals requiring review. Still, they make incompleteness visible in a way that a bare list of titles cannot. The selection begins to describe a network of capabilities rather than a collection of supposedly self-contained answers.
10. Source leads need a narrower status label
An opened institutional website is evidence that a relevant resource or collection can be inspected. It is not evidence that a custom package has been acquired from it, that every promised component exists there, or that the rights allow the intended reuse.
The source registry therefore uses acquisition leads. A national metrology institute can be a relevant place to look for measurement guidance. A sewing extension program can be a relevant place to look for practical instruction. The proposal still needs a later work-level decision: which exact document, edition, diagram or exercise will be used, and under what conditions?
This boundary is particularly important for commercially published technical education. ARRL's Courses & Training page provides an authored route into radio learning, while its page also carries an explicit reproduction restriction. A successful open establishes neither membership access to all courses nor permission to copy a curriculum into another collection. ARRL Courses & Training.
The general rights issue is broader than one publisher. The U.S. Copyright Office distinguishes protected expression from underlying ideas and describes rights that can attach to fixed works. A publicly reachable page should therefore not be treated as an automatic redistribution license. Actual legal conclusions depend on the work, jurisdiction and proposed use. U.S. Copyright Office overview.
The inventory can preserve a useful lead while leaving acquisition and permissions unresolved. That is an honest intermediate state, provided it remains visible to the reader.
11. Documentation makes the process inspectable
The usefulness of a large selection depends partly on whether someone else can reconstruct how it was made. The final prose explanation should therefore be accompanied by structured artifacts rather than serving as their substitute.
For this commission, the meaningful records include the selection contract, the shared taxonomy, the raw authored rows, the deterministic conversion scripts, source registries and validation outputs. Review notes add another layer: they show what an independent reading challenged and how the proposal changed afterward.
A related documentation principle appears in Gebru and colleagues' Datasheets for Datasets. Their proposal calls for describing a dataset's motivation, composition, collection process and intended uses. The present inventory is not being declared compliant with a dataset standard; the useful analogy is that downstream readers need context about how a structured artifact came to exist. Datasheets for Datasets, author record and abstract.
Documentation also protects against retrospective storytelling. Without intermediate versions, it is easy to present a corrected result as though every decision had been obvious from the start. Preserving earlier rows makes it possible to say that a duplicate was found, a source mapping was too broad, or a technically imprecise phrase needed correction.
Those admissions improve the evidence. They show that review altered the work in identifiable ways. They do not prove that review found every problem, but they allow a reader to distinguish an actual correction process from a final document that merely claims to have been checked.
12. Mechanical validation answers limited questions
A deterministic validator can count records, check required fields and compare normalized titles. It can confirm that identifiers follow the promised range, that each category has the prescribed quota, and that every source key resolves to a registry entry. These are useful tests because a large artifact is easy to damage through omissions, inconsistent spelling or incomplete conversion.
The tests remain narrower than editorial quality. A nonempty rationale can be uninformative. A valid source key can point to an irrelevant collection. Two unique titles can describe the same lesson. A plausible dependency list can omit the expertise that matters most.
The distinction can be illustrated with a hypothetical record that passes every structural check while proposing an unsafe improvised repair. Its identifier may be unique, its fields complete and its source URL syntactically valid. None of that establishes the safety or truth of its contents. The validator has answered the questions it was programmed to ask, not every question a reader cares about.
Mechanical validation is strongest when its output names those checks plainly. A report that says unique identifiers verified is more trustworthy than an unexplained quality score. The former can be reproduced and challenged. The latter invites readers to assume that many different judgments have been collapsed into a single authoritative number.
The selection process therefore needs both structured checks and substantive reading. Their roles complement each other, but neither should borrow the other's authority. Passing a schema is a necessary production gate for this inventory, not a certificate that the knowledge collection is complete or correct.
13. A duplicate can hide behind a different object
The practical inventory's review revealed why title matching was not enough. Two initial proposals covered a trouser-knee patch and an elbow patch. Their locations differed, but their instructional cores overlapped: choosing a patch shape, accommodating movement and abrasion, and checking placement. Keeping both would have let a body-part substitution consume a second slot without adding a comparably distinct learning outcome. The elbow entry was replaced with pleat reconstruction and retention, a different clothing problem.
A similar issue appeared in workshop proposals about bench-vise jaws and soft-jaw fixtures. Each addressed protecting a delicate surface while holding it securely. Different wording did not erase the shared outcome. One became a package about left-handed tool access and visibility, changing the problem rather than merely changing its name.
These are examples from the local editorial record, not measured discoveries about how often all knowledge inventories contain duplicates. They show a repeatable review question: if a learner completed one proposed package, what substantially new capability would the other add? An answer can justify retaining two closely related entries. The absence of an answer can justify consolidation or replacement.
The test should not become a demand that related subjects never appear together. Bicycle steering geometry and bicycle gearing share an object but address different mechanisms. A conceptual introduction and an advanced applied procedure might also be distinct, provided their prerequisites and outcomes make that difference explicit. Conversely, entries in different categories can duplicate one another. A dimensional-capture worksheet for a replacement building component can overlap a nominally separate repair-parts measurement worksheet.
The meaningful unit of comparison is therefore the learning and decision content, not the noun in the title. This is harder to automate than matching strings. It requires reading the rationale, contents and dependencies together. It also requires a reviewer willing to challenge plausible-looking entries rather than treating the first complete table as finished.
14. Administrative templates can crowd out knowledge
A particularly revealing cluster involved passports, handovers, manual indexes and service registers. Such records can be useful. Their repeated appearance nevertheless created a risk that the inventory would fill with nearly identical instructions to record an identifier, a date, a contact and an unresolved issue. Replacing equipment names does not necessarily create a new knowledge package.
The repair process did not ban documentation. It asked whether each retained record represented actual domain relationships. The water handover was strengthened around sources, treatment barriers, bypasses, storage and the authority to interpret sampling. The kitchen handover was strengthened around cold-chain dependencies, raw and ready-to-eat separation, allergens and sanitation. These are not simply different labels on the same contact sheet.
Other generic records were replaced outright. A vehicle owner-manual recovery proposal became a package about tire tread and water displacement. A bicycle maintenance passport became a gear-ratio and cadence workbook. A household energy handover was replaced with controlled-fire fundamentals and permitted-use boundaries, responding to a specific gap the commissioner had identified.
This distinction matters because documentation can be both necessary and easy to overproduce. It offers an attractive appearance of responsibility: every topic has a checklist, every device has a passport, every process has a handover. Yet an inventory dominated by those objects could preserve the fact that a manual is needed without preserving much understanding of what the manual must explain.
A strong selection therefore asks two separate questions. Is a record needed to manage this activity? And is the proposed package teaching a distinct body of knowledge about the activity? Sometimes one package can do both. Sometimes the record belongs as a component inside a richer package rather than occupying a separate numbered position.
15. Source matching is an editorial claim
Several corrections were not about the proposed learning outcome at all. They concerned where an eventual acquisition would begin. A water-meter dispute evidence packet initially pointed to a private-well resource. Both concerned household water, but the relationship was too broad. The lead was changed to an EPA resource about understanding a water bill.
A texture-adapted meal proposal needed a specialist lead rather than general food-safety guidance. The replacement was the American Speech-Language-Hearing Association's resource on dysphagia diets, whose discussion situates texture modification within individualized professional assessment and care. The proposal retained clinical oversight rather than turning a collection lead into a self-treatment instruction. ASHA: Dysphagia Diets
Scaffolding procurement questions likewise needed a scaffolding-specific OSHA lead, not merely a page about hand and power tools. Tire sidewalls and spare-wheel readiness were remapped to NHTSA material rather than a general consumer-safety collection. Hearing assistance, communication boards and physical accessible signage received specialist sources instead of relying on a web-accessibility tutorial for every kind of accessibility.
The general lesson is not that broad collections are useless. They are often the right place to begin discovery. The problem arises when a registry's wording makes broad relevance sound like verification of a particular package. A source may be authoritative in one domain and still be the wrong source for the stated outcome.
This is why the registry describes collection leads rather than claiming that each custom package already exists at its linked address. The proposed unit must still be assembled from specific works, with their scope, editions, permissions and suitability checked. A reviewer can verify that a lead is relevant without claiming to have acquired or tested the future package.
16. A familiar authority can carry an old edition
The measurement entries exposed a subtler version problem. NIST's page for Special Publication 811 explicitly warns that the publication has not been updated for the SI changes effective on May 20, 2019. The registry therefore retains that limitation. Uses concerned with notation or conversion guidance cannot silently turn the source into a current account of every SI definition; an updated SI lead is available separately. NIST: Special Publication 811
This is a small correction with a large methodological implication. A familiar institution and a recently maintained webpage do not establish that every document on the page reflects the newest underlying standard. A selection record needs to distinguish the date of a work, the date of a webpage and the date on which the curator inspected it.
For the same reason, the inventory's preparation date is not a universal evidence date for all its sources. An older educational resource may remain useful for a stable concept. A current collection page may link to historical material. Neither situation should be hidden by labeling everything simply current.
Edition checks also change the eventual acquisition task. The curator may decide to keep an older work because its explanation is valuable, but pair it with a note identifying superseded definitions or requirements. Alternatively, the mismatch may be serious enough to reject it. The present proposals do not claim that these item-level decisions have already been made.
A responsible registry makes the problem visible early. Doing so is more useful than presenting an apparently seamless catalog whose authority depends on readers never examining dates closely.
17. Safety is a boundary, not a decorative warning
The commissioned inventory was intended to include foundational practical knowledge, including fire. Removing every potentially hazardous subject would not produce a balanced account of everyday capabilities. But adding a dangerous procedure merely because it is foundational would not solve the selection problem either.
The resulting controlled-fire proposal is bounded around a combustion model, fuel and moisture, ventilation, containment, extinguishment concepts and local restrictions. Its acquisition lead is National Park Service educational material on campfires. That material emphasizes site rules and responsible control; the proposed package does not claim that a general educational page authorizes a fire in a particular place. National Park Service: Campfires
The proposal excludes accelerants and chemical recipes, does not authorize indoor use or fires during prohibitions, and does not claim to be validated operational instruction. These restrictions are part of its intended scope, not a sentence added after an unrestricted procedure. A future editor would need to preserve them when selecting actual material.
The same logic applies to other practical subjects. A conceptual package about ceramic electrical insulator deterioration should not become a live-equipment repair guide. A package on optical clarity in transparent repairs should not imply approval for protective eyewear or other safety-critical optics. A lesson in hydrostatic weighing should not invite immersion of reactive, porous or valuable objects without qualified judgment.
Safety review therefore examines the relationship between the promised outcome and the dependencies. If a package promises a capability that requires a laboratory, trained supervision or a regulated professional, those requirements should be explicit. Calling the package introductory does not make the dependency disappear.
Nor can a general warning redeem a package whose central purpose is inappropriate. Some material should be excluded or redesigned. The useful question is not whether a safety field exists, but whether the selected content could remain within the boundaries that field describes.
18. Review independence has limits
The work used separate drafting and reviewing roles. Reviewers read records they had not authored, questioned overlaps, opened acquisition leads and requested specific changes. That separation added friction to easy acceptance. It is a meaningful production practice, but it is not evidence that the selections were independently generated by different underlying models or independently validated by outside experts.
The distinction is important to the experiment's public description. A coordinated assistant workflow can involve multiple agents while sharing an instruction hierarchy, a workspace, an available toolset and a common commissioning context. Calling that process blind comparison would erase information that readers need to interpret the result.
The review also had a defined object. It assessed the proposals and their stated source leads. It did not acquire every proposed package, run every exercise, inspect every edition in a future archive, or observe people using the material under real constraints. An independent reading of a proposal is not independent testing of the knowledge it hopes to contain.
This does not make the review worthless. Its corrections are visible: overlapping outcomes were replaced, misplaced sources were repaired and misleading quantity language was narrowed. Those are concrete improvements. They should be reported as such rather than inflated into a general assertion that the vault works.
A useful report names both the intervention and its limit. Separate-role editorial review helped identify identifiable defects in the submitted inventory. It did not establish universal coverage, cultural representativeness, instructional effectiveness or safe applicability in all contexts. Those would require different evidence and, in several cases, different participants.
19. Replacements should leave a visible trail
A quota-preserving replacement can look trivial in the final table. One title disappears and another takes its identifier. But without a retained earlier version, the reader cannot distinguish a considered correction from an accidental change or quietly altered promise.
The production files preserve earlier authored rows and item exports before substantive review changes. That allows an auditor to examine what was replaced and why. The current selection can remain clean and readable while the working record retains the history needed to explain it.
This matters especially when a proposed replacement creates another collision. The dimensional-capture repair proposal was first replaced with press-fit and clearance-fit mechanics. A further cross-worker check found that the science inventory already covered that relationship. The practical entry was replaced again, this time with refractive-index matching in transparent repairs. The second revision is evidence that duplicate review remained fallible and iterative, not that the first repair should be hidden.
Keeping quotas stable also makes the cost of inclusion visible. Adding a missing topic need not inflate the total and quietly abandon the commission's boundary. It can require identifying a weaker or redundant selection to surrender its place. That is a useful editorial discipline, although the quota itself remains a commissioned constraint rather than a scientific optimum.
The history should not become a performance claim based on counting edits. More changes do not necessarily mean better work; fewer changes do not necessarily mean cleaner work. What matters is whether the recorded changes address substantive defects and whether the final version incorporates them consistently.
20. The next experiment would ask different questions
This selection exercise can support a future evaluation, but it should not pretend that evaluation has already occurred. One next step would be to acquire a small, carefully chosen sample of packages and ask intended readers to perform bounded learning tasks with them. The task would need clear success criteria, appropriate supervision and permission to stop where safety or expertise becomes limiting.
For example, a low-risk hypothetical evaluation might ask readers to interpret a tire sidewall from a supplied image, distinguish a source lead from an acquired work, or explain why two nominally different proposals duplicate the same learning objective. These tasks would test specific comprehension outcomes. They would not establish that participants could maintain an entire transport system or reconstruct a workshop.
Another proposed evaluation could compare alternative curations under the same scope and budget. If the purpose were a model comparison, the design would need to specify what was held constant, which systems produced which outputs, how assistance and revision were controlled, and how reviewers were prevented from favoring a known source. Those features were not present in this coordinated complementary-selection commission and should not be retrospectively invented.
A community review would ask different questions again. People who use particular tools, teach particular traditions or live with particular infrastructure constraints could identify omissions and misframed dependencies that a general editorial pass misses. Their participation should be described accurately, with actual contributions credited only when they occur. A future invitation is not evidence of consultation.
The central principle is to match the evaluation to the claim. Structural checks answer structural questions. Editorial review can identify textual and conceptual defects. User tasks can test bounded usability. Expert appraisal can assess particular domains. No single layer should stand in for all the others simply because its results are easier to count.
21. What the completed counts establish
The combined export inspected for this article contains exactly 4,000 proposal records: 2,000 in Set A and 2,000 in Set B. There are 40 categories containing 100 records each, with 20 categories in each set. Each of the four authoring assignments contributed 500 records to A and 500 to B. These are measurements of the saved inventory, not projections from an unfinished plan.
The separate set exports match that allocation. Counting their records establishes that the numerical commission has been met in those files. Checking unique identifiers, normalized titles and complete rationale-and-contents pairs provides additional structural evidence. It does not establish that every pair of differently worded entries is conceptually distinct. That is why semantic review and its correction trail remain part of the deliverable rather than being replaced by a duplicate-title test.
The global mechanical checks passed, and both independent editorial review reports subsequently closed their documented corrections. The review covered the proposed records and source-lead scope, with the practical/science and health/institutions/culture assignments recorded separately. This is a completed proposal-field review, not expert certification of every subject, proof of rights clearance, or evidence that the proposed packages have been acquired or tested. Its scope matters as much as its passing status. The culture-source audit also records a specific access qualification: the independent reviewer received a 403 response from Smithsonian Folkways, although the author reported an earlier successful opening. The report does not convert that author report into an independent retrieval claim.
The count should also be read with its unit attached: 4,000 proposals, not 4,000 acquired manuals, completed courses, licensed collections or tested capabilities. A proposed package may ultimately require several works, original diagrams or expert review. Another may prove infeasible to assemble in the intended form. The inventory creates a structured queue of such decisions; it does not make them disappear.
That is the defensible result of Austin Little's selection experiment. It has turned an expansive request into inspectable editorial commitments, exposed collisions that a bare list would conceal, and preserved limitations that a promotional account might omit. The next stage can begin with those commitments and improve them. It should not begin by renaming an inventory as an archive and treating the change of noun as evidence that the hard work is finished.
<!-- ARTICLE_BODY_END -->