Why this block: healthcare has no textbook for what we are building, because nobody has built it. The pieces live in computer science. Each reading below maps to one box of the diagram. Read them in this order; the first two take one evening.
Links checked 2026-10-11. Five are free papers; two are books you buy or borrow.
1. Pat Helland — Immutability Changes Everything (2015)
Box: claims store, append-only. 6 pages. Why you never overwrite, and why it is affordable now. Read it twice; it is the whole stance of the claims store in one paper.
2. Pat Helland — Data on the Outside versus Data on the Inside (2005)
Box: owner's walls, bubbles, "records stay, answers leave". 10 pages. Data that leaves a service is immutable, versioned and carries its own identity. Our bubble principle, twenty years early. Adrian Colyer's summary is a good warm-up.
- Online: https://www.cidrdb.org/cidr2005/papers/P12.pdf
- Summary: https://blog.acolyer.org/2016/09/13/data-on-the-outside-versus-data-on-the-inside/
3. Kleppmann & Riccomini — Designing Data-Intensive Applications, 2nd edition (O'Reilly, 2026)
The backbone: the whole diagram. 670 pages; buy or borrow. Read the chapters on stream processing and event sourcing (the log, change data capture, state derived from events) and the closing chapter on derived data, correctness and auditability. Chapter numbers changed from the first edition, so find them by topic.
- Book page: https://martin.kleppmann.com/2026/03/24/designing-data-intensive-applications-2e.html
- Where to buy: https://dataintensive.net/
- Note: O'Reilly's own catalog page blocks automated checks; it may still open in a browser.
4. Mokhov, Mitchell & Peyton Jones — Build Systems à la Carte: Theory and Practice (2020)
Box: dependency index, corrections follow forward. 54 pages, the hardest and the most ours. How a changed input finds every result that depends on it. Read §3 "Build Systems, Abstractly" (p. 9) and §5 "Rebuilders" (p. 22: dirty bit, verifying traces, constructive traces). Swap "file" for "claim" and "build output" for "issued answer" as you read.
- Online: https://github.com/snowleopard/build/releases/download/jfp-preprint/build-systems.pdf
- The Microsoft Research PDF links block automated checks; the preprint above is the same paper.
5. Cheney, Chiticariu & Tan — Provenance in Databases: Why, How, and Where (2009)
Box: origin of every value, "what did this answer read". 98 pages; reference, not cover to cover. Why-, how- and where-provenance are the formal theory behind our vocabulary's ORIGIN and "derived by". Start with the introduction and the section that compares the three kinds.
6. Richard Snodgrass — Developing Time-Oriented Database Applications in SQL (1999)
Box: two clocks, true-when and learned-when. 528 pages; reference. Valid time (when it was true), transaction time (when we recorded it), and bitemporal tables (both). Read the opening chapter, then the valid-time, transaction-time and bitemporal chapters as the two-clock layer gets built. Out of print; the author hosts the PDF.
- Online: https://www2.cs.arizona.edu/~rts/tdbbook.pdf (the older smsf.cs.arizona.edu link is dead)
7. Peter Christen — Data Matching (Springer, 2012)
Box: identity ledger. The standard book on record linkage: preprocessing, blocking, comparison, classification, quality evaluation, and privacy. Buy or borrow. Read Part II while the matching layer gets built.
- Publisher: https://link.springer.com/book/9783642311635
Written exit test
- Draw the diagram from memory. Under each box write the one reading it stands on.
- In one sentence each: why append-only (Helland 2015); why records stay inside (Helland 2005); what a "constructive trace" is and why it is how a correction finds old answers (Mokhov); the difference between valid time and transaction time (Snodgrass); where-provenance versus why-provenance (Cheney).
- Say aloud: "Models propose; a named, versioned rule decides." Then explain with one reading why a replay must read the stored claim and never re-run the model.