A Bookshelf, a Dataset and the Design of NeoLibrary

By Rob Tyrie · Published October 25, 2026

A photograph of a bookshelf can expose a surprising amount of bad information architecture.

The obvious task is to identify the books. The interesting task is deciding what the photograph actually proves, how it relates to other evidence, and what we can responsibly build from it. That is why I am using my books as a working example of the NeoLibrary design at Ironstone Advisory.

The raw material is ordinary: Amazon receipts, Audible history, ThriftBooks purchases, Libby borrowings and photographs of shelves. Together, they describe a library. Separately, they tell different stories, with different omissions. Getting them into one spreadsheet is useful. Understanding their relationships is the real work.

Start with what the evidence says

As of October 25, 2026, our working dataset has 2,144 provisional work identifiers. Provisional matters. These are working groupings, not a certified count of unique books, and certainly not a claim about how many I have read. Buying a book and finishing it are different events. My shelves have been making that distinction for years.

A purchase receipt records a transaction. A borrowing record records access. A photograph records an observation at a particular time. None should quietly turn into evidence of something else.

We therefore distinguish the work from its edition, a physical copy from an acquisition, and an observation from the object observed. An audiobook and a paperback can represent the same work without representing the same transaction. Two photographs can show the same copy without either photograph being redundant.

This sounds fussy until the system starts making confident mistakes. A familiar cover does not establish which edition was purchased. A matching title does not establish identity. The name on a spine might belong to an editor or translator rather than the author.

The dataset needs room for uncertainty before it needs another attractive dashboard.

Improve the record without erasing its history

We are assembling the data in a private Google Sheet. It is a controlled prototype, not a finished NeoLibrary engine. Tables make the current work inspectable, but their columns are not the underlying model of the world.

Several AI agents are helping with identification and enrichment. ClaudeBot, Eve and GeminiBot propose work; Robot coordinates the process and acts as the sole writer to the dataset. I initially carried messages between them. That was an instructive way to discover that being the integration layer is not a particularly scalable occupation.

The process now assigns nonoverlapping groups of work identifiers, requires source evidence, and separates proposals from independent verification. Writes happen in small batches. The writer reads the current cells before changing them, reads them back afterwards, and compares untouched data. Uncertain matches stay open rather than being forced into a tidy answer.

So far, 428 blank author fields have been filled after independent verification. That is useful progress, not a declaration that every record is complete.

Corrections need a history too. Before replacing an existing value, we preserve it in a cell-referenced comment, alongside the source, time and reason for the correction. Otherwise, enrichment becomes a polite name for destroying evidence.

One proposed author attribution looked plausible and was wrong. Verification caught it before it reached the dataset. That small incident explains the control better than a paragraph about responsible AI ever could.

Coordination has also been messy. Entries in the shared agent log were replaced during concurrent activity. Status labels and summaries drifted. Some proposed work lacked visible review evidence. We have moved Robot's log writing to append-only operations and are recovering missing entries. The system is improving; pretending it never failed would teach us nothing.

The NeoLibrary connection

NeoLibrary is a semantic archive and publishing system. It holds disparate information objects, preserves their meaning and relationships, and supports transforming them into other objects without losing the route back to their origins.

For the books example, that means connecting works, editions, copies, acquisition events, photographs, people and time. A table is one view of those connections. A catalogue is another. A reading list or an essay could become a further expression of the same underlying material, with its own purpose and context.

For example, a thematic reading list should distinguish books I own from books I borrowed, and both from books I have finished. It should separate community ratings from my own judgement. Those distinctions let the same archive support different publications without quietly inventing a more impressive reader in the process.

The important design question is how evidence is stored and how it moves. An original photograph should remain available after a title has been extracted. An identification should retain its source and uncertainty. A corrected attribution should have a version history. A published result should carry enough context to explain what it represents.

This is recursive work. A new photograph can challenge an earlier match. A verified edition can clarify an ambiguous receipt. A proposed publication can reveal information we still need. We return to the archive, reinterpret, enrich and publish again. A neat diagram with arrows pointing only forward would conceal much of the actual process.

Agents can operate on a NeoLibrary, but the archive is not intrinsically autonomous. The distinction matters. The system must remain useful and intelligible regardless of which model happens to be doing the interpretation.

Ironstone Brain is the business instance of this design. Its scaffolding, archive and indexing exist; the integrated Brain is still being built. This books exercise tests the discipline that such a system requires.

What the bookshelf teaches the business

Organisations face the same problem with contracts, meeting notes, correspondence and reports. They accumulate records faster than they establish what those records mean, how they relate, or which interpretation deserves trust.

The books make those questions tangible. A person still has to resolve an obscured spine or authorise a consequential change. More agents do not remove that responsibility, and we have not run an experiment proving that this team is faster than one agent.

What we are building is a way to make information more useful without making its history disappear. If the system cannot explain how a bookshelf became a catalogue, I would hesitate to trust it with a company's memory.

© 2026 Ironstone Advisory. Licensed under CC BY 4.0 (Attribution 4.0 International).