Ninety-one artifacts.
Not a reading list.

Taken as the maxi build rather than the 46 GB or 7.9 GB slices. The 69 GB saved would have cost the long tail — which is precisely where a niche repair procedure or an uncommon diagnosis lives. At 2% of the drive the trade was not close.

Taken with images rather than the smaller text-only build: at a single additional language the storage saved is not worth the lost diagrams. That calculation reverses at seven languages, where the English edition already holds the same illustrations.

Kept after an explicit argument. Against: it is overwhelmingly about web frameworks and cloud services that presuppose the internet, and the coding model already encodes much of it in 18 GB of weights. For, and it won: the one obscure answer about a serial protocol or a compiler flag cannot be predicted, and that unpredictability is the whole case for a full corpus over a curated slice.

Ten times the size of the engineering site. Somebody repairing a charge controller, a pump timer or a radio is the intended use case more directly than most of the corpus. If only one Stack Exchange site survived, this is the one.

Question-and-answer suits physics teaching better than encyclopedia prose, because the reader arrives holding a specific confusion rather than a topic.

Small for its domain, and complementary to Wikipedia rather than duplicative: the questions are the ones a reference work never thinks to answer.

The laboratory-safety material carries more weight offline than on: there is nobody to ask, and the failure modes in this domain are not forgiving.

Not named in the specification; added deliberately. In a scenario where this node matters, keeping a truck running is closer to the purpose than several things the specification does name.

Not named in the specification; added deliberately, to serve the applied side of the mathematics corpus. Surprisingly small for the domain it covers.

Serves trade and market mechanics at the practical end — the questions somebody actually has, rather than the theory behind them.

Question-and-answer suits this better than a treatise, because people arrive with a situation rather than a doctrine. It is not the law of any particular country, and the node says so when asked.

The first material ever placed in this directory, which until 2026-08-25 was completely empty.

An index, not a shelf — these books are not on the drive. The catalogues that are here list thousands of titles and pass no judgement on any of them. This is the judgement: which open textbooks a review board considered good enough to teach from. Worth keeping because it decays faster than the books do — several entries already resolve only through the Internet Archive.

It answers the problem the mathematics section opens with: Wikipedia is a poor mathematics teacher, because its articles are written for readers who already hold the concept. Question-and-answer is the inverse — someone stuck, someone explaining, at every level from arithmetic to research.

Known as Cross Validated. The agricultural trial and the quality-control chart are the two places statistics earns its keep in this deployment, and both are well covered.

A direct hit on the requirement for worked examples that Wikipedia does not provide. Five years stale, and it does not matter — mathematics does not rot.

Complements the someone-stuck-someone-explaining register of the Q&A sites with the precise version of the same result.

In no specification list; added deliberately. The build's second artifact concerned with transmitting knowledge rather than holding it — which is what a node meant to teach ultimately requires.

The first bilingual dictionary in the build, and open all the way down — the source is editable text, not a database only one program can read. The two directions are not equal. English to Spanish is 64,258 entries, current. Spanish to English is 116 kilobytes and two years older. Both are here; only one is finished.

A catalogue, not a library. The works are not here. What it carries is the imprint: "En Madrid : por Iuan de la Cuesta …, 1609" — the press that printed Don Quixote. A record of what was printed, where and when, which is not reconstructible offline. The Biblioteca Virtual Miguel de Cervantes reserves rights over its digitised texts and we respected that; the same institution publishes its catalogue with no restrictions at all, so that is what was taken.

The difference between a scanned transcription and a book someone would choose to read. Carried twice over — one format to read, one to search — because the searchable form is what the model actually retrieves from. This is the cleanest licence in the archive: public domain throughout, with no non-commercial clause, so it can travel with a node under no condition at all.

With the Spanish Gutenberg holdings this is the Spanish literature shelf. It is not the Cervantes library, which remains the outstanding item — a distinction worth keeping, because Spanish literature at real depth is what makes Spanish a first language here rather than a translation target.

Its value exceeds the description. Those older technical books were written when the subjects were taught procedurally, assuming the reader intended to do the thing rather than understand it. Modern references assume you can buy the part; these assume you must make it. It also carries the literature that makes this a civilization backup rather than a survival tool.

That is a property of the public domain corpus itself, not a packaging decision: far less pre-1928 Spanish-language material has been digitised. It is also the only place in the build where the Spanish edition is newer than the English — by two and a half years.

Extracted by tag sweep rather than editorial selection, which is why the curated MDWiki build is taken alongside it rather than instead of it.

A Spanish-language clinical query has materially less to retrieve from than an English one. That bears directly on the instruction to verify any clinical answer against the archive before acting on it.

Overlapping content with the Wikipedia medical subset, different selection method. Both were taken deliberately: one is swept by tag, the other chosen by clinicians, and the disagreements between them are informative.

Closer to the intended use case than a reference work is. Found by reading the directory listing rather than searching it — a keyword search for medicine would not have surfaced it. It grew from 46 MB to 357 MB in three months, so it is actively expanding.

It shrank from 76 MB to 53 MB over nine months, which is unusual for a wiki. Flagged for re-checking at the next refresh, in case content was lost rather than deduplicated.

Standard medical references are overwhelmingly illustrated with light skin, and the presentation of rashes, infections and inflammatory conditions differs visibly. A reference showing only one is a diagnostic hazard, not merely incomplete.

Livestock health is a food-security question in the target scenario, not pet care.

The zimgit collections are deliberately assembled rather than mirrored, which makes them dense for their size and worth carrying alongside the large sweeps.

The first artifact here that addresses a limitation the design openly admits: the model's vision will not reliably identify plants or animals, and staying geography-neutral cost the regional field guides that were the stated mitigation. This is not a field guide and no substitute for one — it is the largest identification resource obtainable.

Carried with a caveat. 1.42 GB is large for a single blog and it is probably image-heavy; whether the content is reference-grade or content marketing is an open question, and it is flagged for assessment rather than assumed good.

Nothing else in the build serves this. Power is the first-ranked design constraint, and until this arrived the corpus said nothing at all about the system the whole node depends on.

Appropriate technology across solar, water systems, small-scale construction and sanitation — arguably a better fit for the practical section than several sources the specification does name.

Repairability is a design priority, and a node that cannot be fixed has a shelf life. The photographed teardown is a format nothing else in the corpus reproduces. Both languages, and the Spanish build is the newer of the two — the only place in the archive where that is true.

The largest single item on the drive, and the only one that teaches by watching someone work a problem rather than by reading. Textbooks assume a reader who can already study alone. This is for the one who cannot yet.

A repair community rather than a manual: diagnoses, what actually fails, and fixes people have tried. Small at 17 MB, and it complements the iFixit manuals rather than standing in for them.

Not strictly medical, but continuous with the disaster-response use case, where the question is usually about water, shelter and sanitation before it is about injury.

Trivially small for the consequence it carries, and one of only two things partially mitigating the absence of Spanish environmental-health material.

Preservation is the part that matters here — a harvest that cannot be stored is a harvest that is mostly wasted.

Purely practical, trivially small, and exactly the kind of thing nobody thinks to include until they need it.

The register matters as much as the content. Most technical references are written for people who already work in the field; this one assumes a beginner and a teacher, which is the mode this node most often has to operate in.

Spanish coverage has been thin in every corpus acquired. Ten megabytes of Spanish practical instruction is not much, and it is more than most of the practical shelf offers in Spanish at all.

This one is about keeping the node itself alive. The system runs Linux, and there is no other systems documentation anywhere in the build — so when the machine rather than the question is the problem, this is the only thing to read.

This is what substitutes for regional curation when the deployment site is undecided. Per-country reference is the geography-neutral answer to a corpus that cannot know where it will end up.

A placeholder more than an answer: it supports the agriculture shelf until the FAO and extension-service material arrives, which is where the geography-neutral agricultural corpus actually comes from.

The largest artifact in the build, and the one that does not go stale — the ground does not move. Terrain answers where water flows and pools, which slope faces the sun, whether a route is walkable, where a small hydro drop exists, what floods. The finer regional archives were declined: the global source is 30-metre data, so past this zoom nothing new exists except in a handful of wealthy countries. Taking the planet file keeps the archive geography-neutral instead of quietly committing it to one continent.

Answers what grows here, anywhere on earth — which is what the agriculture material needs and what a road map cannot give. The four historical periods are the part worth having: the same place classified in 1901 and in 2020 shows whether the growing conditions have moved, which no single snapshot can.

A duplicate that earns its place. The maps say Bogotá is Csb and Mérida is Aw with equal certainty. This layer says the models agreed 50% of the time about Bogotá and 100% about Mérida — and that Mérida's apparent century-long shift begins from a class half the ensemble rejected. Without it the node reports a change it cannot support. It is also the only claim in the archive stored twice in independent formats, so the two can be checked against each other instead of trusted.

The planet basemap is data; this is the documentation for reading and rebuilding it — which is what matters when the map server needs troubleshooting and there is no internet to search.