Every rule below was learned by getting it wrong here first. They live in
who-iris/skills/iris-dspace.md — read the skill —
and this page is generated from that file, so the two cannot drift. There are
22 of them, against 222 item(s) in the
catalogue, 12 of which this repository actually holds the bytes for.
Why a page and not just a skill. A skill is read by an agent about to act. A reader deciding whether to trust what was ingested needs the same facts and will not open a skill file to get them. Same text, two audiences, one source — the skill.
| # | Requirement | Why, in one line |
|---|---|---|
| R1 | Resolve by Handle where one exists; treat UUID and govdoc as secondary | only the Handle is guaranteed outside WHO |
| R2 | Never deduplicate dc.identifier.uri | the second one is evidence of a host that vanished |
| R3 | Preserve repetition on every field, not a known list | four fields repeat in a single record; assume more do |
| R4 | Treat an absent language as unasserted, never as en | both states occur in one record |
| R5 | Carry dc.subject.mesh as a controlled value with its authority | it can be resolved and subsumed; a keyword cannot |
| R6 | Model bitstreams per bundle, with size and format on the bitstream | an item is not a file |
| R7 | Model containment as a list of paths | four levels deep, and multiple mappings are legal |
| R8 | Never infer metadata from the PDF when a record exists | OCR off a title page catalogued the WHO Editorial Style Manual as Abies |
| R9 | Record what the source says even when it contradicts the file | dc.description says 30 p.; the PDF has 33 |
| R10 | An item with no known collection carries parents: [] — never a guessed one | held-and-unplaceable is a STATE; inventing containment is unfalsifiable |
| R11 | Key a node by path when its UUID is not in the capture | an invented UUID resolves to nothing while looking authoritative |
| R12 | Keep the page captured and the identifier apart | /items/<uuid>/full is what was fetched; the Handle is what the item IS |
| R13 | A materialized claim must name bytes that EXIST, and something must check it | an unchecked claim can name no bytes and still pass as held |
| R14 | Read a storage figure in the units that round-trip its own bytes | IRIS prints "GB" for what its byte counts make GiB — a 7% error |
| R15 | Never fall back to a parent URL when an item has no upstream URI | "no upstream URI recorded" is the answer; the front page is a lie that resolves |
| R16 | A partial transcription must SAY it is partial | a truncated abstract that does not declare itself cannot be told from a short one |
| R17 | Hold what you hold: model the bitstreams whose bytes are here | the handbook lists 8 upstream; this repository has 1 |
| R18 | Do not "correct" a record that looks wrong | the English-titled editorial manual really does carry an Italian ISBN |
| R19 | A derived figure must round-trip the figure it was derived from | round(361.55 × 1024³) bytes must render back as 361.55 GiB |
| R20 | An item count and a file count are different denominators | 273,559 against 1,057,223 — a fraction built from the wrong one is off by 3.86× |
| R21 | Mint stable URIs at the extractor (Upstream Skolemization) | minting https://iris.who.int/entity/item/{handle}#{prop}_{idx} at extraction leaves no blank nodes for a later pass to resolve |
| R22 | Append-only streaming partition writes | A 300,000 item corpus cannot accumulate uncompressed in working memory; write quads directly to partition file streams |
yl5wEvery ORIGINAL localPath in this catalogue named
who-iris/uploads/ while the bytes sat in cat-harness/uploads/,
and check:catalogue did not look at localPath at all — so
three materialized claims resolved to nothing while the gate printed
clean. The three sources now sit beside their own intake records at
who-iris/uploads/<slug>/, the page generator's fallback is deleted,
and check:catalogue verifies every localPath — with
could not determine as its own answer, because an unreadable file is not a
missing one.
iris.who.int is egress-blocked from the environment that generates this
page, so nothing here was fetched from IRIS. Every record was transcribed from a capture
the owner supplied. Where a transcription is partial, the record says so.