Trust Depth Over Data Volume: An Independent Audit of LUX Light Archive

Most heritage databases answer one question: where is the object, and what does it look like today? LUX Light Archive was built to answer a harder one: what remains the same thing while becoming something else?

In August 2026 we commissioned an independent, AI-assisted research audit of this archive — a structural review carried out by an external model (Google Gemini), working only from the public site and from published external sources, with no privileged access to our internal editorial tools. We are publishing an adapted, source-linked version of that audit here, deliberately, as a demonstration of the same principle the audit itself describes: evidence should be checkable, not just asserted. Every internal claim below links to the live page it describes. The full source list is in the Sources section.

The core design choice: trust depth over data volume

The audit's central finding is architectural, not cosmetic. Most digitization projects optimize for coverage — how many records, how many photographs, how complete the inventory. LUX optimizes for something else: trust depth. Every claim in the archive carries a confidence level, a review status, and a direct source link. Evidence (what is documented) is kept structurally separate from narration (how it is told).

That separation is visible directly in how the site presents itself: every Heritage Journey page renders in two modes over the same underlying data — Story, an editorially guided narrative, and Research, the raw evidence table, source list, and open questions behind it. A reader can move from a sentence in the story to the exact claim, source, and confidence level it rests on, on the same URL. Governance for that dual-mode contract is public — see the site's own Story Mode principles: Story is not a simplified or "tourist" version of Research; both are equally accountable to the same underlying claims.

The practical consequence is that uncertainty is shown, not hidden. Where a coordinate, a date, or a demolition is not confirmed, LUX marks it as an open question instead of quietly rounding it into a false-precise fact.

What the audit found in the data itself

The audit's numbers, gathered from the public site in August 2026, sketch a database that is deliberately small and dense rather than sprawling: roughly 1,241 heritage records built from 518 core entities — objects, publications, persons, and media collections — connected by some 370 typed relationships (mentions, worked_at, cites, documents, and 156 transformation links specifically tracking construction, splitting, or destruction events). Object types include lighthouses, range/leading lights, optical apparatus, lighthouse churches, and lightships. Around 660 heritage assets — lenses, lanterns, optics — are modeled as independent entities with their own identity and provenance, separate from the buildings that once housed them, so an apparatus can be dismantled, relocated to a museum, and still keep a traceable lineage.

Mapping is treated with the same discipline. Locations carry an explicit role — current light position, physical structure, historical position, approximate zone, museum display — rather than a single flattened point, and precision is stated per record, from roughly 50 meters for well-documented structures down to wide approximate zones for early or contested sites. Where a historic coastline cannot be reconstructed with confidence, LUX leaves it explicitly unresolved rather than interpolating a shoreline that isn't backed by evidence.

How that compares to classical lighthouse registries

The audit places LUX against the two dominant global inventories — Russ Rowlett's Lighthouse Directory and the ARLHS World List of Lights — and its conclusion is not that one replaces the other. They answer different questions.

Classical registries answer "where is it and what are its current characteristics." LUX is built to answer "how did this object become what it is now, and what exactly survived its transformation" — a different job, closer to a semantic-web research tool than a directory.

Reading the identity-continuity cases directly

The audit's strongest evidence is not the statistics but the individual cases — and every one of them is a live, checkable page rather than a claim in a report. A reader does not have to trust our summary; they can open the record.

• The Great Light — a hyper-radial lens built for Tory Island in 1887, later split by Chance Brothers into two biform branches at their Smethwick works. One half returned to Tory Island; the other went to Mew Island until 2014, then to a public display in Belfast's Titanic Quarter. Eighteen graph relations tie three optic identities across five locations into one traceable lineage.
• La Coubre and Roches-Douvres — French cases where the navigational function survives repeated physical destruction: wooden towers, stone towers, wartime loss, and reconstruction, tracked as one continuous warning function across structurally distinct buildings.
• Vasilisin and Zmiinyi — lights destroyed outright, where identity survives as memory rather than structure: scale models, family testimony, and postcard imagery, modeled as legitimate evidence carriers rather than folklore. Vasilisin's continuity runs directly through Mayachnik, the Russian lighthouse archive this project grew out of, active since 2008.
• Cape North and Lightship Columbia — objects that moved from active navigation aid into museum life without losing their identity chain.
• Pencarrow Head and Clare Island — cases where the navigational function relocated to a new site while the original tower and optic remained on record as separate heritage objects.
• Women Who Kept the Light — a reminder that "who carried the function" is itself part of the evidence chain, not a footnote to it.

That is 23 such cases in total, each readable in Story or Research mode from the same URL. Traditional catalogue formats — a text field inside one object card — cannot represent a split lens or a relocated function without collapsing it into prose. A typed relationship graph can.

AI's role, deliberately constrained

The audit noted something specific to this project: AI is not used here as an author. Under the site's own Guiding AI principles, a model is not permitted to invent facts outside the evidence base, upgrade a weak graph proximity into a stated causal link, publish an unreviewed interpretation, or convert private memory into public heritage claim. Every AI-authored proposal must cite an existing relation, source, or open-question ID, or it stays an internal review prompt rather than a published fact.

This article is itself an instance of that boundary in practice. It originates from an external AI-generated report, but nothing in it is treated as canonical; the underlying facts live in the governed lighthouse records and journeys it links to, not in this prose. A companion project, ALManac, applies the same evidence-before-narrative discipline to institutional decision histories rather than lighthouses.

Where the audit sees the most room to grow

Three directions stood out as the highest-value next steps, and they are consistent with work already underway rather than a change of course:

1. Linked Open Data. The archive already exports a CIDOC-CRM 7.1.3 aligned preview with LUX IDs, external alignments, and citation metadata (including BibTeX, via the how-to-cite page). Moving from a preview export to a stable open endpoint would let aggregators such as Europeana or ARIADNEplus pull lifecycle histories automatically. 2. A distributed model for small institutions. Local museums and coastal historical societies rarely have the infrastructure to run a graph database. The European Coastal Digital Heritage Lab model — contribute raw material, LUX applies editorial and ontological rigor — is a way to extend this method without asking every small archive to rebuild it themselves. 3. Coastal-erosion monitoring. LUX already has the ontological machinery to record erosion-driven loss — La Coubre and Cape North are direct examples. Pairing that with climate-risk geospatial datasets would turn "lost lights" records into a longitudinal dataset on coastal change, not just a heritage curiosity.

Why we're publishing this as evidence, not marketing

We could have summarized this audit in a paragraph on the homepage. We didn't, because the whole point it makes only holds if it is checkable. Search for 1,111 records across seven filter groups, open any linked journey above in Research mode, and compare what you find to what is claimed here. If something doesn't hold up, that is exactly the kind of open question this archive is built to record rather than hide.

Sources

This article adapts an independent, AI-assisted research audit (Google Gemini, August 2026) into our own source-linked format. The audit's own citations, and the internal LUX pages it references, are listed below; none of the internal links above are asserted as canonical fact by this prose — they are pointers to the governed records that are.