SOURCE CURATION SPEC

Curation-only imagery — retiring Universal Diagrams

Written 2026-07-28, opening the stint after the Page Mode hardening pass. Supersedes the Wikimedia portion of SOURCE-EXPANSION-SPEC.md, which assumed the searched categories would stay.

---

0. The rule

Every image the system can show is a work someone chose. There is no keyword-searched imagery anywhere in R.I.S.E.

This is now a standing invariant, alongside reverent degradation and pinned sacred imagery. It applies to the sources that exist and to every source added hereafter.

The Universal Diagrams (Wikimedia) family is retired in full — not curated down, not partially salvaged. Retired.

---

1. Why the rule, and why total

The ten Universal Diagrams categories were live keyword searches against Wikimedia category trees. The museum collections are pinned accessions. A pinned work was chosen once and stays chosen; a searched category returns whatever the tree holds on the day you ask.

A live audit of all ten (2026-07-28, Commons API, 40 members each) found the searched shape failing in every way it can fail:

id	what the search actually returns
microscopy	n=0 — the category does not exist on Commons
solar	Category:Sun is topical: an .ogg audio file, French springs, the Luxor temple
sacred	contemporary hobbyist "sacred geometry" — the exact woo the system has outgrown
geometry	1690625457877-image.png, 15brezinys-2.png — scratch files
fractals	amateur 3D renders; we already GENERATE fractals, better
astronomy	mostly a Little of Earth PNG series, not diagrams
botany	real plates mixed with colonial-era ethnographic material
haeckel, anatomy, romantic	good today

The last row is the reason for total retirement rather than salvage. Haeckel and the Wellcome anatomical plates are genuinely fine right now — and that is exactly the trap. Their quality is an accident of what the category holds this week, not a property of anything we control. Keeping them would preserve the searched mechanism while pretending the quality problem was about taste. The mechanism is the problem.

If those plates are worth showing — and Haeckel plainly is — they are worth pinning, one accession at a time, like every museum work. That is a content task, not a provider task.

microscopy is the other half of the argument. It returned nothing for its entire life and nobody noticed, because reverent degradation did what it promises: a work that will not resolve is simply absent. The doctrine is correct and stays. But it means a searched source can rot invisibly, and a pinned one cannot.

---

2. Why this blocks the source expansion

NASA and the other planned sources were going to arrive in the same shape — a provider wrapping a live query. The audit says that shape does not meet the floor the museum collections set, so adding more of it multiplies the problem before fixing it.

Retire first; establish the pinned pattern; then expand into it. NASA imagery enters as chosen plates with attribution, exactly as a museum work does.

This is North Star §6 restated at the source layer — reverence over completeness. A source that cannot promise a good image should not promise one.

---

3. The migration constraint (do not skip)

These ids are load-bearing. Grepped 2026-07-28:

• src/components/Vault.js — vault presets cite geometry, fractals (line 81) and haeckel, botany (line 108).
• src/content/atrium/collections.js — 34 mappings across 82 total: geometry ×9, sacred ×9, astronomy ×5, botany ×4, microscopy ×2, anatomy ×2, solar ×2, fractals ×1.

A deletion is therefore not a deletion; it is a migration. Removing geometry outright would strip imagery from nine philosophy collections, and reverent degradation would hide the loss exactly as it hid microscopy.

Every retired id needs an explicit decision:

• re-point — to a pinned collection or an existing museum id; or
• drop deliberately — the collection reads with no sourced imagery. This is a legitimate and often better outcome; many readings are stronger plain, and a philosophy collection with no honest image should show none.

Dropping must be a decision on the record, never a silent consequence. No id is deleted until its dependents are re-pointed and a test asserts the new mapping.

---

4. Page Mode gap — the structural cause

Universal Diagrams do not render in Page Mode because Wikimedia categories are bare ids (haeckel, botany) while every other family is namespaced (aic-*, chapel-*, atr-*, dore:*). _getProviderForCategory dispatches on prefix, so a bare id falls through every namespaced branch.

Full retirement closes this gap by removing the only unnamespaced family. Anything later pinned from these plates enters under a namespace and works in Page Mode from its first day.

Constraint: aic-* ids are a vault dependency and are never renamed.

---

5. Absence must be observable

microscopy is the lesson. Reverent degradation is right for the reader — a broken frame is worse than silence — but it must not be silent to the maintainer.

A test should probe every registered collection and fail when one resolves to zero works. Cheap, and it would have caught this the day it broke. Same class of gap as the earlier lesson that a test's fixtures must not be better-behaved than real input.

Under curation-only this check becomes strong rather than advisory: a pinned collection resolving to zero is unambiguously a defect, because nothing about it depends on what a remote tree holds today.

Built 2026-07-28 as scripts/probe-collections.mjs (npm run probe:collections). It resolves all 17 pinned collections through the real service — the same resolveCollection and isDisplayable a reader gets — and reports EMPTY, DEGRADED, or THROTTLED per collection. Current state: 341/341 works resolve.

It is a script rather than a suite test on purpose. It issues real requests to four institutions, so in CI it would fail on any museum outage and train people to ignore it — the exact opposite of its purpose. The unit tests mock fetch and assert the CONTRACT; this asserts the WORLD, and the world is allowed to be briefly unavailable.

Two lessons are baked in, both learned the hard way while building it:

• Pace the requests. The first version probed 341 works in a tight loop, tripped the Met's per-IP rate limit, and reported chapel-resurrection at 20/46 — manufacturing the very absence it exists to detect. Every missing work was a Met pin; each resolved fine when asked alone.
• Distinguish CANNOT-ASK from IS-ABSENT. The Met's block is total while it holds (403 to every request) and lifts after ~45 seconds, so the count was identical every run and read like a real defect. A retry inside that window is refused identically. The probe now backs off past the window and reports a single-institution shortfall as THROTTLED rather than as absence.

A corollary worth stating: the Met is the largest source, not the smallest — 127 works of 341, ahead of Cleveland (86), Rijks (66), and AIC (62). Nine of the eleven Atrium collections are 100% Met, so removing it would empty Plato, Socrates, Aristotle, Marcus Aurelius, Stoicism, Cicero, Diogenes, Demosthenes, and Rousseau. Its per-collection counts look small only because philosopher collections are inherently small. The Chapel, by contrast, is well diversified (crucifixion 23/25 non-Met, patriarchs 81/83). The concentration risk is the Atrium's philosophers, and it is a curatorial matter, not a provider defect.

---

6. The Collections IA

The panel currently shows a flat list of 10+ categories under one accordion. The proposed regrouping — style vs subject — is the right instinct and is the same progressive-disclosure move as North Star §4 (stances).

Settle it after the retirement. With Universal Diagrams gone the surviving set is museum collections only, which may not want two headings at all — the grouping should be designed against what actually remains.

---

7. Order of work

1. Re-point or deliberately drop every dependent: 34 Atrium mappings, 2 Vault presets. Tests first — the mapping is the contract.
2. Retire the Wikimedia family in full, including the provider's category registry. Keep the provider module only if the Atrium's atr-* resolver still needs it; otherwise retire that too.
3. Add the absence check (§5).
4. Regroup the panel (§6).
5. Then the source expansion (NASA and beyond) enters as pinned works.

Tranche 2 — the Atrium's atr- categories (DONE, 2026-07-28)

The twenty atr- categories were searched too, and are retired. Two things made this smaller and clearer than expected:

• A successor already existed. imagery/collections.js holds eleven PINNED collections naming museum accessions with artist, title, and date — David's Death of Socrates, Testa's Plato's Symposium. Six ids lived in BOTH registries, and the cortex checks pinned first, so those were already resolving to real works. Emptying the searched registry removed a shadow, not imagery.
• A prior audit (2026-07-21) had already reached this conclusion and marked the module deprecated, recording the sharpest statement of the problem: Commons categories are FILING, not curation, and filename plausibility is not image quality.

That last line corrected this pass's own method. A file-type audit scored these pools at 90%+ "artwork" — Plato in art 91%, Marcus Aurelius 93% — which looked like grounds to keep them. Inspecting the actual titles showed the rasters were coins, genealogical charts, a Brussels building facade, an Esperanto book cover, and a Wellcome engraving of a woman with a bird on her head. The metric was measuring file extensions and calling it curation. Any future source audit must look at what the images ARE, not at what the filenames suggest.

No reading lost imagery: every affected record was already overridden by imageryPlanFor (conceptual → fractal, liberation → the Freedom field, mechanism → blueprint), and all 44 records still resolve to either pinned works or a procedural engine.

---

8. Invariants

• Curation-only: every image is a chosen work. No keyword-searched imagery. (New, established here.)
• aic-* ids are never renamed (vault dependency).
• Sacred imagery is pinned, never searched. Retiring the sacred category is this invariant finally applied evenly — it was a keyword search wearing a sacred name, which is precisely what the rule forbids.
• Reverent degradation holds for the reader; §5 adds observability for the maintainer without weakening it.
• Content authors; the runtime follows; the cortex renders. Re-pointing a collection is a CONTENT decision and belongs in the content layer.

---

Published from docs/vision/SOURCE-CURATION-SPEC.md. Edits made here are overwritten on the next push to main.