How a shelf of books becomes a thinking library — the five-step engine that reads it, takes it apart, and finds the lines that run across it, wrapped by four quiet layers that measure it, keep it honest, keep it running, and hold every judgment only a person can make.
A shelf goes in. A photograph of spines — or an import — becomes a candidate list of real books, matched to editions and covers. The raw intake.
Each book is filed into the lattice — the order of knowledge. This fixes what a book is: its discipline, the voice it is written in, its form, where it is set, how demanding it reads. These are the facets a book carries with it, later read off for free.
The Samudra Manthan — the churning of the ocean. Manthan runs in two acts. First it grows the facets nobody recorded — the people and works a book is about, the thinkers it descends from, the stances that recur — by churning the whole library's ideas until they precipitate out. A guardrail drops anything that merely restates a fact the library already knew, and an independent exam checks the rest before it becomes permanent vocabulary. Then it passes every book through the finished prism, refracting it into every facet at once and recording one grounded membership — a rare, whole-library pass for the first churn, a cheap per-book pass for every book added since. A book lands on every shelf it truly belongs to, not just one.
see it →The mirror. Everything the churn surfaced, held as the curator's complete self-knowledge — Lekha-Jokha (the numbers) and Darshan (the meaning). From this full record a chosen few findings are pulled forward.
see it →The window. What the world sees — the beautiful views: the lens galaxy, the library over time, where the books come from.
see it →Manthan’s instrument. Rather than ask which one shelf a book goes on, the prism refracts it into every facet it carries — so a book belongs on as many shelves as it truly does. Some facets a book is given and we read off for free; a few must be discovered by churning the library itself, because no one ever wrote them down.
Every book is then passed through the whole prism in one step — Slot — which records one grounded membership keyed by the book’s id: 1,252 books, ~11,474 memberships across the thirteen facets, not one of them ungrounded. Why we built a prism →
Five engines relay a book from a spine on a shelf to a sentence in the window — each orbiting it as it works. Beneath: Taal blinks each engine’s beat; Parakh glows as each output is checked; a queue holds what only a person can decide; Hisab keeps a running receipt of every call.
Beside the pipeline sit four things that aren’t steps — they wrap every step, and together they close a loop. One measures every run, one keeps it honest, one keeps it alive, one holds what only a person can settle: Hisab senses what each run costs, Parakh checks that nothing invented slips through, Taal decides when the engines fire again, and a Console gathers every question the machine cannot answer for itself into one place a person can actually work through.
Hisab is the reckoning — the account kept. Every model call leaves a receipt: tokens, cost, time, outcome, written to one step-tagged ledger. One store, many views — including the one no separate log can answer: the true cost of a single book across the whole pipeline. It is the sensor every other layer runs on, already watching every call — and reading its receipts is how the machine learns which of its own crutches it no longer needs.
Parakh is the assayer’s word — to test for genuineness. Every engine’s output is independently checked by another, never by itself. The core is one reusable assay — does this claim actually trace back to the evidence? — applied at every step, weighted to where the risk lives: data errors at the start, hallucination in the middle, leakage at the end. The vocabulary Manthan grows is examined the same way before it becomes permanent: a sample is graded independently against a rule fixed before any result exists, and a finding survives only if it holds up against three separate attempts to refute it.
Taal is rhythm — the beat, reading Hisab’s ledger and every engine’s own records to decide what is due: a new book waiting to be slotted, a finding gone stale, a page that has drifted from what the engine now knows. It plays the steps that need no judgment by itself, on a standing weekly beat, and hands anything that costs money or needs a person’s call to the Console below — so drift is caught from the records, never simply remembered.
Every question the engines cannot settle alone — a book cited that may not exist on the shelf, a stance that might overreach its name, a shelf grouping close enough to an existing category to ask twice — lands in one queue, not scattered across a dozen tools. Each carries its evidence and a short set of honest options. The engine proposes; only a person disposes — and every verdict writes itself back into the record the moment it is made.
Experiments on the hard part — getting a machine to read a shelf and see a library the way a reader would.
The wider search — reading not just this library but the shape of whole fields and book-cultures.