shuffle-on
New Books translated page by page, layout kept — and OCR for the scanned shelves The Library · offline knowledge

A library that reads, answers and proves every page.

72,304 works you can search by meaning, ask questions of, and read in your language — every file fingerprinted in a signed ledger anchored in Bitcoin, plus 400 offline reference books. No cloud. One machine. Yours to verify.

See how it works How the proof works
72,304Works catalogued
123,113Files, all fingerprinted
11.0MMultilingual passages
364,517Entities in the graph
7,810Books with study guides
414Offline reference archives

Figures refresh from the live system every 30 minutes.

What it is

Not a folder of PDFs. A reader that never sleeps.

Most digital libraries stop at storage. This one keeps going. Every book that arrives is identified, deduplicated, catalogued, split into passages, embedded by meaning, read by a local language model for its people, places and ideas, summarised — and then sealed with a cryptographic fingerprint.

The result is a collection of 72,304 works in 123,113 files across 53 shelves and 68 formats that you can talk to. Ask it a question and it answers from its own books, citing the passages — or tells you honestly that the answer isn't on its shelves.

Search by meaningAsk the LibraryStudy guidesKnowledge graphBitcoin-anchored provenance41 MCP tools
The pipeline

From a raw file to an answer you can trust

Seven stages, all running locally, all resumable. A book dropped into the inbox comes out searchable, summarised and provable.

  1. IngestOne import queue for uploads, the NAS inbox and harvests. Metadata suggested with a confidence score; only confident items apply themselves.459,846 audited edits
  2. ReadBorn-digital text is extracted; scanned pages go through the new vision OCR with layout and tables.118,779 docs indexed
  3. EmbedEvery book split into passages and embedded twice — English and multilingual (1,024-dim bge-m3).11.0M passages
  4. UnderstandA local model extracts summaries, subjects and named entities, and links them into a graph.9.34M entity mentions
  5. TeachSection-by-section study guides with key points, validated for quality before they are kept.327,561 sections
  6. AnswerKeyword, semantic and hybrid search fused by reciprocal rank; grounded Q&A that cites or abstains.3 search modes
  7. ProveEach file's exact bytes fingerprinted, logged in a signed Merkle tree and anchored in Bitcoin.0 errors / 123,113
A real record

Follow one book through the machine.

Numbers are abstract. Here is what the pipeline actually produced for one public-domain classic — its files and their seals, the summary a language model wrote, the people and ideas it found, the study guide, and where the library sent it next.

Work 30433/1859/3 files · EPUB · LIT · TXT/Public domain

On the Origin of Species by Means of Natural Selection

Charles Darwin

evolutionnatural selectionspeciesheredityadaptationvariationscientific method

Summary machine-written

Groundbreaking scientific treatise examining evolution through natural selection, variation, heredity, and adaptation. Darwin presents arguments on how species change over time through selective pressures and environmental competition, illustrated with extensive examples from domesticated organisms, wild populations, and geological evidence. The work establishes that divergence of character and extinction drive biodiversity, and refutes claims of special creation through detailed analysis of sterility in hybrids and transitional forms.

People found in the text: Charles Darwin · W. Whewell · Francis Bacon · Kölreuter · Gärtner

Study guide written by a local model · in reading order

  1. §0

    Introduction: Historical Context of Evolutionary Theory

    • Shift in scientific opinion regarding species, contrasting the dominant view of immutable creation with the emerging theory of modification. It traces the intellectual lineage from Aristotle's accidental adaptation to Buffon's early scientific treatment, highlighting Lamarck's pivotal role in proposing descent with modification…
  2. §5

    Preface and Introduction to Natural Selection

    • Darwin explains that he published this abstract due to his poor health and the simultaneous arrival of Alfred Russel Wallace at nearly identical conclusions regarding the origin of species. He acknowledges the assistance of Sir Charles Lyell and Joseph Dalton Hooker, who facilitated the publication of Wallace's memoir alongside brief extracts from Darwin's own manuscripts…
  3. §12

    Evidence for Single Origin of Domestic Pigeons

    • Darwin argues that all domestic pigeon breeds descend from a single ancestor, the rock-pigeon (Columba livia), rather than multiple wild species. He refutes the theory of multiple origins by highlighting the improbability of half-civilized humans domesticating seven or eight distinct species and then losing them while retaining specific traits…

Opening sections of the guide · regenerated on 13 Oct 2026 by the local pipeline

A real record from the library, unedited except for length. This book is in the public domain.

Where it came from

Twenty years of digital chaos, put in order.

72,304Works
79,268Editions
123,113Files · 68 formats
459,846Audited edits

For more than twenty years I kept everything I could find. Scans, ebooks, papers, manuals, courses — in every format there is, under every naming mess you can imagine. Duplicates everywhere. Wrong authors. Books split across folders. The same title in five editions, none of them labelled.

Putting it in order took three kinds of work, and never just one of them:

1By hand — reading, comparing, deciding.
2By scripts — fingerprints, ISBNs, filenames, metadata.
3With AI assisting — never alone. Every suggestion carries a confidence score, and only the confident ones apply themselves.

Every change is audited — 459,846 of them so far. When duplicates are merged, the losers are quarantined, never deleted. Today it is 72,304 works on 53 shelves.

Now that it is in order, it is sealed — so the order itself can be checked.

— Cristian

New · October 2026

Scanned books, finally readable — and accountable.

Thousands of the most precious works on the shelves are photographs of pages: old manuals, out-of-print treatises, school books. Text search could not see them. Now a vision-language model reads them the way a person does.

Layout, not just letters.
Each page becomes typed blocks — headings, paragraphs, tables, footnotes, figure captions — with their positions, in the right reading order even across columns.
Tables stay tables.
Every row and column comes out as a clean Markdown table, so the numbers in a 1950s manual become searchable data.
Faithful by rule.
It is forbidden to translate, summarise, correct or invent. Unreadable spans are marked [illegible], never guessed.
An attestation for every result.
Each result records the source file's fingerprint and IPFS address, the text's fingerprint, the model and the time — "this text came from this exact file".
Two modelsFast or carefula 4B reads quickly, an 8B reads hard pages; any page can be re-read with the other
Page-exact proofWhat the model sawthe deskewed 200-dpi image, the file's fingerprint, the model and the time stay with every run
People firstGPU etiquettea page you ask for jumps the bulk backlog; the backlog pauses when a person needs the model

An "OCR this document" button on every work, a bulk queue for the whole backlog, a REST API — and a phone app: take a photo of a page and watch the model read it live, block by block. Every run is kept in a history with its attestation; any page can be re-read with the other model; a result can be shared as a read-only link or an encrypted paste. Every page is saved the moment it is finished — stop, resume, restart, nothing is lost — and the backlog steps aside whenever a person needs the GPU.

Memory firstTranslation memorya sentence translated once is reused everywhere it recurs; the model only sees what is new
Named onceGlossary by the big modelone call per book decides every name and term; the 9B then translates with that list
Scan or textSame patha page without a text layer is detected and read by the OCR's layout mode first; diacritics verified in the fonts
New · October 2026

Read any book in your language. Same page, same place.

A treatise in French, a manual in German, a photographed page in Portuguese: the library translates it page by page and sets the translation exactly where the original text stood — same box, same size, same emphasis. Figures, ornaments and the paper itself stay untouched.

Layout kept, not approximated.
Typed pages lose only their original glyphs; scanned pages go through the OCR's layout mode and each block is repainted in the paper colour around it. Text shrinks only as much as the longer language needs.
A glossary before the first page.
Names and terms are decided once for the whole book, by the larger model, so chapter forty says what chapter one said. You can check and edit the list before any page is translated.
Draft, then proofread.
A second pass reads the draft against the source and fixes it. Sentences already translated are reused from memory. Three tasks — translate, proofread, glossary — each with the model you choose.
Your corrections stick.
Compare pages side by side, correct any segment and the page is re-rendered. Redo a page, translate five more later, pause and resume — a finished page can be read while the rest of the book is still running.
From the phone, too.
Photograph a few pages with the camera and they become one job; the result opens on the same screen. Source language detected automatically, target among twenty.
Back on the shelf.
Started from a library work, the translation is pushed back as a new edition of that work — with a note of the models used and the pages done — searchable, readable and fingerprinted like everything else.

The same local models do the work: the 9B for the pages, the 27B for the glossary. A "go faster" switch pauses the other pipelines that share the GPU for the length of one job and resumes only what it paused.

Provenance

Every file has a fingerprint. Bitcoin is the witness.

Digital archives quietly rot: files get swapped, edited, corrupted. Here, that would be visible. All 123,113 files — 928.9 GB — are fingerprinted from their exact bytes, written into an append-only log, and the log itself is sealed in the Bitcoin blockchain.

Fingerprint

IPFS address + SHA-256

Each file gets a content address (IPFS CIDv1) and a SHA-256 hash. Change one byte and both change. 123,113 of 123,113 fingerprinted, zero errors.

Log

Signed Merkle tree

The same design that keeps the web's TLS certificates honest (RFC 9162). Each daily epoch root is signed with Ed25519 and chained to the last — history can be appended, never rewritten.

Anchor

OpenTimestamps → Bitcoin

Each epoch is timestamped through four independent calendars and confirmed in a Bitcoin block. To fake the library's past you would have to rewrite Bitcoin's.

The ledger so far

EpochSealedFiles in treeBitcoin blockStatus
#05 Oct 202649,749970,063confirmed
#16 Oct 2026120,894970,105confirmed
#27 Oct 2026122,157970,263confirmed
#310 Oct 2026122,158970,698confirmed

122,158 distinct contents (identical copies share one leaf). A new epoch is cut daily when something changes; proofs upgrade every three hours; the signed epochs are mirrored to an append-only git history. The library's own IPFS node holds all 914 GB with zero mismatches.

Latest signed root (epoch #3): 32f8262798d2b6777ebeea48fa0515c216ea84779c480922900b62de71eaa24b
Signing key: Ed25519 · fingerprint 29512140b6c3c727 · pinned 5 Oct 2026

How anyone can check a book

  1. Hash the file you have.SHA-256 and the IPFS address of your copy — any standard tool works.
  2. Get its inclusion proof.The audit path from your file's leaf to the signed epoch root.
  3. Check the signature and the chain.The Ed25519-signed root, plus a consistency proof to the latest epoch.
  4. Check Bitcoin.The OpenTimestamps proof resolves to a real block anyone can look up.

A standalone verifier script and a ledger browser exist; access is available on request. Every docset also exports an MLCommons Croissant dataset card listing each file's hash, address, epoch and block — validated with zero errors.

Search & Ask

Ask in plain words. Get the page, not a guess.

Three ways to find anything, fused into one ranking — and a question box that answers only from the shelves when you want it to.

Keyword · Meaning · Hybrid

Exact words when you know them, semantic search across 11 million multilingual passages when you don't, and a hybrid that fuses both with reciprocal-rank fusion. When the embedder is busy, it degrades gracefully to keyword search instead of failing.

Grounded or exploratory

Grounded mode answers only from the library — or abstains. Exploratory mode may reason beyond it. A fast model answers in seconds; a deeper one takes on the hard questions.

Read it your way

Search inside chosen works, follow a series in order (one runs to 663 volumes), discover related works found both by meaning and by shared subjects, and convert any book to EPUB, PDF, MOBI, AZW3, DOCX, RTF, FB2 or TXT.

ask · grounded
# question
What did people believe about species
before Darwin, according to the book?

# retrieved
[1] On the Origin of Species · §0–§1
    Historical Context · Antecedents

# answer
Most held that species were immutable,
each created independently. Lamarck had
proposed change through inherited
acquired characteristics, and earlier
thinkers such as Erasmus Darwin and
Goethe foreshadowed the idea [1].

↳ 1 source · answered only from the shelves

Example based on a real study guide.

Understanding

It has read the books, so you can start anywhere.

A local language model works through the collection day and night — in English, Romanian and some forty-five other languages — turning books into summaries, subjects, entities and lessons.

Summaries80,561

documents summarised, with 537,169 subject tags.

Entities9.34M

mentions of people, places, organisations and ideas.

Knowledge graph364,517

canonical entities linked by 174,097 co-occurrence edges.

Study guides7,810

books guided section by section — 327,561 sections, searchable on their own.

Study guides that respect the book

Each section gets a summary and its key points, in reading order. Quality gates reject copied, repeated or weak sections, check the language, strip filler, and validate the whole document before it is kept. Medicine & Health goes first — 6,155 of 13,251 documents done — and a queue of ~875,000 sections keeps two workers busy, stepping aside whenever a person is chatting.

Languages in the reading room

English53,226
Romanian26,186
~45 others2,100+
MCP connector

Plug the whole library into your AI.

The library speaks the Model Context Protocol, so Claude or any MCP-capable assistant can search it, read it, follow its references and verify its files — with 41 tools, each labelled read, write or critical, per-user tokens, and every call logged.

Find & read 4

get_work · find_work · read · book_guide

Search 5

search · search_semantic · search_hybrid · ask · docset_search

Browse 6

authors · series · facets · subjects · entities · work_facts

Reference shelf 7

reference_search · reference_read · reference_books · work_references · reference_for · collections · collection_search

Provenance 4

provenance · verify · ledger_status · dataset_card

Curate, import, export 15

related_works · edit · relocate · merge · docsets · import queue (dry-run first) · convert · export

Find Darwin's Origin of Species in the library and check that the file is authentic.
library_find_worklibrary_verify
Found it — work 30433, three files (EPUB, LIT, TXT). The EPUB's fingerprint is in the ledger and sealed in Bitcoin block 970,105 on 6 Oct 2026.
Walk me through it, chapter by chapter.
library_book_guide
Here is the guide in reading order. §0 traces the views before Darwin — immutable creation, Buffon, Lamarck; §1 the thinkers who anticipated natural selection, from Erasmus Darwin and Goethe to Wells's 1818 essay… Shall I continue from §2?Example using real tool names and a real record
The reference shelf

Wikipedia, Stack Exchange and hundreds more — inside every search.

Next to the books sits an offline reference shelf of 400 curated Kiwix books (414 archives on disk). They are not a separate island: the first page of every library search also returns the matching reference articles and public-domain ebooks — the library looks beyond itself for you.

Programming198

books — 184 DevDocs manuals, Python docs and PEPs, Linux documentation, cheat sheets.

Questions & answers107

Stack Exchange sites, fully offline.

Science & education29

LibreTexts, PhET simulations, Khan Academy, school manuals.

Encyclopedias17

Wikipedia in English and Romanian plus subject selections; Wiktionary.

Survival & practical17

Survival guides, iFixit repair manuals, Wikivoyage.

Health, literature, music, talks25+

MedlinePlus and medicines references, Gutenberg, Wikisource, Wikiquote, music and talks.

Collections you can search by title, author or subject

CollectionItemsSpan
Project Gutenberg69,918public-domain ebooks
Survivor Library14,3661684–2014, 163 categories
US Army publications4,3731947–2024
Algorithms course library696lecture notes & exercises
Moldovan school textbooks~74grades 1–9

Books that know their own Wikipedia

Works, authors and entities are matched once to their Wikipedia articles — by Wikidata identity where possible — and cached, in English and Romanian.

Works2,818 EN · 1,453 RO
Authors1,614 EN · 1,082 RO
Entities2,799 EN · 1,348 RO
The engine room

One machine. Local models do the heavy lifting.

Everything above — embeddings, summaries, OCR, answers — runs on hardware in one room in Romania. In September 2026 the lab moved from an 8 GB consumer card to a 32 GB workstation GPU, and replaced its old model server with a single smart router.

AMD Radeon AI PRO R9700

32 GB of GPU memory, tamed from 100 °C to a steady ~88 °C with a 210 W cap and an extra fan — for only ~3% less speed. Beside it: two 22-core Xeons (88 threads) and ~512 GB of RAM.

One router, many minds

A llama.cpp router serves every model on demand: idle models free their memory and wake in about 2.5 seconds from a RAM disk. Around 75,000 requests a day, with people always served before batch jobs.

The right model for each job

Qwen3.5-9B with speculative decoding (67 tokens/s) for everyday reading and answers, Qwen3.6-27B for deep questions, Qwen3-VL for OCR, bge-m3 for meaning. It replaced the earlier Ollama setup in October 2026.

Where it all lives

Owned, not rented.

The library and its index live on the lab's own disks: NVMe for the hot models and indexes, a 54 GB RAM disk for instant model wake-ups, and a 58 TB NAS for the collection itself. The content-addressed IPFS store holds a second, self-verifying copy of every file. Nothing depends on a cloud account that can be closed, repriced or switched off.

928.9 GB fingerprinted914 GB on IPFS58 TB NASKubernetes on one node
When the internet goes dark

Censorship. A severed cable. A long winter.

Whatever the reason, this machine keeps answering. Encyclopedias in two languages, medical references, repair manuals, survival guides, school books, seventy thousand public-domain classics and a library that can explain them — all on local disks, all answerable by a local model.

Knowledge

The entire English and Romanian Wikipedia, Wiktionary, Stack Exchange, LibreTexts and Khan Academy.

Health & repair

MedlinePlus, medicines references, military medicine, iFixit — the pages you need when nobody can be reached.

A teacher

Local language models that can read any of it to you, explain it, translate it and quiz you on it.

No subscription. No permission needed. No cloud. Just electricity and knowledge.— the design brief, from day one
Curious?

Want to read, research or build on it?

Access to the library, the verifier and the MCP connector is by invitation. Researchers, educators and builders of sovereign infrastructure are especially welcome.

Request access on LinkedIn Meet lsn-agent-v3