Develop - #60
Open
namedgraph wants to merge 72 commits into
Open
Develop#60namedgraph wants to merge 72 commits into
namedgraph wants to merge 72 commits into
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01C17WzuUXBuyAxZ3nQASCNg
- Line items table on order pages via forward ldh:view (detail gallery) - Order status (OrderDelivered/OrderProcessing/OrderProblem) derived from shipped vs required date; shipped date and freight re-emitted - Line item extended prices (schema:totalPrice) computed from the existing quantity/unitPrice/discount columns - Orders gallery first on the container page, newest-first, with customer/broker/date/status columns; narrative blocks rewritten; buggy sales-by-region chart (duplicate ldh:seriesVarName) removed - Inverse views upgraded to TableMode with richer columns; new "Orders shipped by this shipper" panel; showWhenEmpty false - KPI row (orders/revenue/avg order value) on the root dashboard - Fix unbound ?deliveryLocation triple in orders.rq The source CSVs stay stock Northwind - no orderTotal or productName columns are added. An order total is the sum of its line items and a product name lives in products.csv, so neither belongs in the source data; the row-at-a-time importer cannot aggregate or join, but that is a constraint to design around rather than to denormalise away. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01C17WzuUXBuyAxZ3nQASCNg
Page structure and copy: - Drop the frontpage block advertising LinkedDataHub's features and the low-code build story; an app does not explain its own toolchain - Lead the dashboard with the monthly sales trend instead of a two-value table stretched over the 400px chart canvas - Record lists come before analytics in each container, so a section opens on its data; narrative blocks run between sections throughout - Give the four reference containers (suppliers, shippers, regions, territories) the intros the others already had - Drop the revenue-by-category chart duplicated in products.ttl; it belongs to categories.ttl Labelling: - dct:title on every ldh:Object block. Without it the block heading renders the fragment identifier, so pages showed "kpi-block", "select-suppliers" - Northwind's own vocabulary in the ontology: Sales rep (not Broker), Reports to (not Sponsor), Territory, Region, Company name, Required by - Use .ldh-section for section headers; page-header/lead are Bootstrap 2 leftovers with no rules in the current design system Correctness: - Revenue charts are now net of schema:discount. They summed quantity x price and overstated by 6.7% (1,297,141 gross against 1,215,813 net), so no chart reconciled with the order totals - No hardcoded counts in prose. "all 830 orders" was also wrong: 19 rows are dropped at import because shipPostalCode is empty and the mapping requires it - Exclude the final, incomplete month from the trend. The data stops on 1998-05-06, which rendered as a collapse rather than a truncated month; the cutoff is derived from MAX(?orderDate), not pinned to a date - Order-scoped schema:totalPrice is no longer required by any query, following its removal from the import mapping Charts: - Add revenue by carrier to Orders, the first chart here with multiple series. Each ldh:seriesVarName names its own measure column, so the query pivots with conditional aggregation; the removed sales-by-region chart passed a label column plus a value column, which is why it never rendered Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01C17WzuUXBuyAxZ3nQASCNg
The checker maps every .ttl to <path minus .ttl>/, which is right for containers but wrong for root.ttl: install.sh PUTs it at $base itself, not at <dir>/root/. So a relative link in a root.ttl resolved one level too deep, and demo/northwind-traders/root.ttl's href="orders/" was reported as demo/northwind-traders/root/orders.ttl. No root.ttl carried a relative link until now, which is why this went unnoticed. The link is correct at runtime; the checker's base was not. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01C17WzuUXBuyAxZ3nQASCNg
Several schema.org properties were carrying values outside their declared domains. Each is moved to the term the vocabulary actually provides: - Order lines are now schema:OrderItem *and* schema:Offer. schema:OrderItem has no price properties at all, so schema:price/schema:priceCurrency were off-domain; typing the line as the offer that was accepted for it makes them legal without splitting the line across two nodes. - schema:Offer's own schema:price is the price of the offer as a whole, so it replaces schema:totalPrice (a Reservation/Ticket property) as the line total. The historical unit price moves to a schema:UnitPriceSpecification with schema:priceType schema:ListPrice. - schema:discount is dropped. It is an Order-level property in schema.org, and the list price, quantity and line total already pin the per-line rate exactly - unlike a stored rate, they cannot disagree with each other. - Product prices move from schema:Product to a schema:Offer reached through schema:offers. A product is not an offer, so this one needs its own node. - Freight moves off schema:ParcelDelivery, which has no price property, onto a schema:DeliveryChargeSpecification reached via schema:priceSpecification from a shipping schema:Offer. - The dispatch date moves off schema:availableFrom (a DeliveryEvent property borrowed onto a ParcelDelivery, and meaning "available for pickup" rather than "dispatched") onto a real schema:DeliveryEvent hanging off schema:deliveryStatus. Orders never dispatched get no event at all. - schema:orderedItem loses its rdfs:domain/rdfs:range. schema.org allows Order|OrderItem -> OrderItem|Product|Service, which a single domain/range pair cannot express; the narrowed pair entailed every OrderItem is a Product. Also fixes a pre-existing import bug: shipPostalCode sat in the required BGP but 19 orders have none, so those orders never entered the graph at all. Making it OPTIONAL restores them, taking the import from 811 to 830 orders and revenue from 1,215,812 to 1,265,792 - the canonical Northwind figure. Consuming queries follow the new structure. The six revenue queries drop their ?quantity * ?price * (1 - ?discount) recomputation for the materialised line total, which also settles a 0.1 discrepancy between the carrier chart and the rest. SelectProductsFromSupplier traverses schema:offers/schema:price. Verified by running all eleven import queries over the real CSVs and all 28 view and chart queries against the resulting graph: 830 orders, 2155 lines, freight totalling 64,943, and top product/employee/customer/category all matching canonical Northwind. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SYifJzEkGsTcJsuB4fGRAc
…and orders pages The schema.org remodelling added facts the content blocks were not using: per-order status, a materialised line total, the catalogue price each line was struck against, and a DeliveryEvent carrying the dispatch date. Four new charts read them, each with a narrative block ahead of it. Front page gains a discounting section. Revenue is what was billed; the ListPrice specification on each line is what was listed, so the gap is what was given away to win the deal. As a share of gross it runs 3.6%-11.2% and is decorrelated from volume, which a revenue chart cannot show. Orders page gains a fulfilment section - average days from order to dispatch, and the share of orders that beat the date the customer was promised - plus a late-shipments-by-carrier bar alongside the existing revenue-by-carrier line, so carriers are compared on reliability as well as volume. "Orders by country" is dropped to make room. The front page already tells the geographic story in revenue rather than order count, which is the more useful of the two measures. Both rate charts are normalised on purpose. Raw monthly status counts squash the late series (0-4 a month) against a delivered series climbing 21 to 69, and the 21 still-open orders are a dataset-cutoff artifact confined to the last two months - they are filtered out of the on-time and late rates, since not yet dispatched is not the same as delivered late. Also corrects the orders intro, which still promised unit prices in the line-items table; those moved to a nested UnitPriceSpecification. Verified against the full graph built from the real CSVs: 31 view and chart queries, none failing to parse, none returning no rows. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SYifJzEkGsTcJsuB4fGRAc
customers.rq required <#postalCode> in its main BGP, but HUNGO — Hungry Owl All-Night Grocers, Cork — has none. Ireland had no national postcode system until Eircode in 2015, so a 1996-98 Irish address legitimately has no postal code; this is correct data, not dirty data. The row processor emits no triple for an empty cell, so the BGP failed to match and the entire customer was silently discarded - company, contact, address, phone and coordinates alike. Its 19 orders kept their schema:customer link, which then resolved to nothing. Moving the column to an OPTIONAL, alongside region, fax, lat and long in that same query, restores the customer and takes dangling customer references from 19 to 0. HUNGO imports with its city and no postal code, rather than a fabricated placeholder. Same failure mode as the 19 orders dropped on an empty shipPostalCode. An audit of all eleven mappings against their CSVs confirms these were the only two required-but-sometimes-empty columns. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SYifJzEkGsTcJsuB4fGRAc
- command-line-interface: rewritten for the ldh executable — build and PATH setup, PKCS12 keystore auth, LDH_* environment defaults, stdout and exit code conventions, the command table (#scripts is now #commands) and a deprecation note for the bin/ HTTP API scripts - every page that invoked a script now invokes a command: upload-file, create-documents, create-content (--fragment is --uri), import-csv/rdf, change-model (create-construct.sh etc. never existed — they are ldh admin ontologies add-constructor/add-class/add-property-constraint) and the data-model CLI tables - manage-packages / administration/packages: packages are installed by declaring one ldh:import triple, composed at request time; drop the Actions dropdown, the packages/install endpoints and the docker-compose restart step - http-api: drop the removed package endpoints, document the settings endpoint, and correct the container hierarchy predicate to sioc:has_parent - admin app URLs use the admin. subdomain, not an /admin path; examples use localhost:4443 and a single example agent URI throughout - get-an-account: the CLI reads the .p12 keystore directly, PEM is only needed for curl Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014Tgf5DjbYiX3SPnCFSgCzd
The multi-stage Docker example built the WAR overlay on maven:3.8.4-openjdk-17, which cannot compile against a Java 21 platform - and the same page already told you to target Java 21 ten lines further down. The build stage now matches the platform's own Dockerfile. The pinned versions were stale in the same way: the base image at 5.3.2 and the overlay dependency at 5.5.3, both now 5.9.1, the current release. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014Tgf5DjbYiX3SPnCFSgCzd
Present Northwind as a business application, and fix its schema.org modelling
Merge the duplicated pages into their parents (content blocks, documents, resources, browse data), collapse the static ToC to the active trail, and fix every broken anchor. Add the versioning reference and the version-history guide; document the Memento options of ldh get, federation through the Linked Data proxy, and UI language negotiation. Sync the reskinned UI descriptions, the ORCID login, the fuseki/ data directory and the make up/drop targets. Import the tutorial from the docs-tutorial branch and migrate all of its examples from the deprecated bin/ scripts to the ldh CLI. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
The generated pages link the vendored design-system stylesheets (fonts, tokens, components, m3 skin, ldh-bridge) plus a docs-specific skin, and drop Bootstrap 2 and jQuery. The shell mirrors the app's: the sticky ldh-header carries the landing page's navbar content (version picker, sample applications, GitHub and social buttons), and the ldh-footer matches the app's footer. ttl-to-html.sh copies the CSS and fonts from a local LinkedDataHub checkout (LDH_SRC) or fetches them from GitHub (LDH_REF, default develop - flip to master once 5.11.0 ships). The UI/CLI tab strips run on vanilla JS; the rsync excludes now keep *.xsl and *.css out of the published files/ tree. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
docs/ and packages/ sit one level deep in the repository, but their Makefiles defaulted the owner certificate three levels up - pointing outside the sibling LinkedDataHub checkout, so accepting the make install defaults failed on realpath. The demo apps sit two levels deep and their three-level default is correct. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
One dataset runs through the whole documentation now: the examples that used SKOS concepts, countries CSV, Places, Friends and DBpedia all draw on the Northwind Traders demo instead, sourced from the demo app where an exhibit exists and linked to its file on GitHub and the live resource on northwind-traders.demo.linkeddatahub.com. The tutorial conforms to the demo: the revenue query sums the line totals the order-details mapping already computed (the old one multiplied quantity in twice), the composition sequence matches root.ttl, employees link portraits with foaf:img, and the dangling query and block references are defined. The demo grows the two pieces the docs lean on: a schema:Product constructor and missing-name constraint in ns.ttl, and categories/unesco-mappings.ttl mapping the eight categories to UNESCO Thesaurus concepts - the RDF import example and the seed of the cross-dataspace federation thread. The exhibit-starved pages get their first exhibits: an acl:Authorization and access request, the PATCH accepted/rejected pair, the PROV-O TimeMap with its link-format and Memento exchanges, the northwind-traders dataspace quoted from config/*.trig, and the stylesheets reference gains the full global parameter table, the author-facing function tables, and an input/output template pair. Every runnable SPARQL exhibit was executed against the live demo endpoints; screenshots are placeholder comments pending capture. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
The screenshot comments and the remaining Bootstrap-era media - six YouTube embeds, four webm recordings and eleven screenshots - become visible placeholder blocks: a dashed box with an image or play_circle glyph and a caption typed "Screenshot:" or "Screen recording:" that describes what the capture should show. The placeholder blocks are the shot list, in place; the orphaned media files are removed. Also corrects the Condiments pictureHash in the tutorial's CSV excerpt to the file's real content hash. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
The Resources reference and the Query data guide each carry an actual query block - an ldh:Object over an sp:Select bound through ldh:service to an sd:Service pointing at the Northwind Traders demo's SPARQL endpoint. Rendered in a LinkedDataHub instance, the block hydrates with the service picker and editor, and Run executes against the remote dataspace; the documentation demonstrates the platform inside itself. The static site renders such blocks as an "Interactive block" stub in the placeholder style instead of dropping them silently (the children view stays suppressed - containers list their children separately). These two pages' media placeholders from the preceding sweep ride along. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
…ioms, a fuller function reference Tab groups move from the dead Bootstrap tabbable markup to the design system's ldhc-tabs component with ARIA roles and hidden-gated panels, on all eight pages that use them; the static site's tab script follows. The customization guides drop retired forms - $ac:uri patterns become ac:absolute-path(ldh:base-uri(.)) - and the stylesheets reference gains ac:uri() and eight more public functions (label helpers, sort companions, ldh:link-targets, ldh:hash-code, value-set utilities) grouped into a new Utility table, with the closing note honestly covering everything left internal. Editing docs describe the dirty-gated autosave (no edits, no save, no version) and the tab-group limitation: panels edit in place, stacked while editing, but creating a group still takes the HTTP API. Guides move from user-guide/ to extending/, and prose gets a wording pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S1UnYpbW5aYLj7LbwtyUo1
Seventeen docs pages that presented mutually exclusive paths as sequential headings now use the design system's tab groups: the reference Management sections (UI/CLI pairs in data-model and versioning), the user-guide pages whose CLI content trailed as a final heading (add-data's three sources, version-history, create-resources), the triplestore authentication methods (HTTP Basic / Bearer token), the authorization example in RDF and CLI form, and the eight tutorial stages' in-the-browser/command-line pairs (insight's CLI path is its chart+view walkthrough, kept whole as one panel). The copenhagen and unesco-thesaurus demo apps shed their dead Bootstrap markup (page-header/lead, and unesco's collapsed row-fluid/span grid) for the ldh-section idiom, and gain tabs where their content branches: the copenhagen feature run-on becomes a five-tab tour with links into the app, unesco's usage-path dl becomes Browser/Editor/RDF Import (the import panel now names skos-import.rq), and the concepts intro's relationship list becomes Hierarchical/Associative/Matches with the live Northwind category mappings as the matches example. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S1UnYpbW5aYLj7LbwtyUo1
The tutorial's ten stages, setup's make targets, the Data stage's four BIND variables and the annotation dialog's fields were all parallel term-dash-description list items; each is now a striped table in the reference pages' idiom (the dialog's advanced fields keep their own table under the expander sentence, replacing a double-nested list). See-also lists and prose-heavy routes stay lists - they enumerate reading paths, not facts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S1UnYpbW5aYLj7LbwtyUo1
The Resources reference contrasted two SELECT queries by padding them into a single pre with spaces, so the pair read across the page instead of down it and neither query could be copied on its own. They are now two pre elements, each introduced by the sentence that says what it does; the "# renders category documents" / "# renders nothing" comments go, since the prose already carried them and their absence leaves the queries paste-ready. A sweep of the corpus for pre blocks with a second column of content turned up no other side-by-side pair - the remaining multi-column blocks are directory trees, aligned PREFIX declarations and end-of-line comments, all single examples. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016YxhcqMPK1kZR1pxxJqoge
Twenty-eight rdf:XMLLiteral lexical forms across the docs were not exclusive-canonical XML. Forty-four screenshot-placeholder spans wrote class before aria-hidden, and the stylesheets reference's wrapped function signatures used self-closing <br/>; canonical form wants attributes in alphabetical order and explicit end tags. RDF 1.1 makes canonical form the datatype's lexical space, and Jena 4.7.0 - which the Check Docs syntax check pins - enforces it while validating. The workflow runs under bash -e, so it aborted on the first offending file and reported only about.ttl. Jena 5 and later dropped the canonicalisation check and validate well-formedness alone, which is why make validate stayed green locally the whole time. Each literal is now its own xmllint --exc-c14n serialisation, so the change is confined to attribute order and empty-element tags; no whitespace, text or namespace moved. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016YxhcqMPK1kZR1pxxJqoge
check-xmlliterals.sh extracts every """..."""^^rdf:XMLLiteral lexical form, canonicalises it with xmllint --exc-c14n and diffs the result, reporting file:line and the offending lines. Run against the tree as it stood before the preceding commit, it returns the same verdict as Jena 4.7.0 on all twenty-eight offenders, with no false positives - so the rule survives independently of which Jena is installed. validate-documents.sh calls it after the riot pass, giving make validate a check that can actually fail. Check Docs runs it as its own step, which lets the syntax check move from the action's implicit 4.7.0 default to Jena 6.1.0 without losing the rule; Jena 6 needs Java 21 and the runner defaults to 17, hence the setup-java step. The check assumes each literal is a single-rooted fragment, which all ninety-five are, and compares the Turtle-escaped text - sound here because the corpus's only escape is \\, which canonicalisation leaves untouched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016YxhcqMPK1kZR1pxxJqoge
The matches tab pointed at concept3683/, concept1812/ and concept3495/ relative to the concepts container. Those documents are real once the app is installed - skos-import.rq constructs them from the thesaurus RDF - but they are never files in the repository, so check-links.sh, which resolves relative links to checked-in .ttl files, could only report them broken. The three now address the public demo instance, in the same form the Northwind mappings and the reference docs already use, and carry target="_blank" like the app's other absolute links. This surfaced only now: the RDF syntax check used to abort the Check Docs job before the link step ever ran. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016YxhcqMPK1kZR1pxxJqoge
Faceted search and parallax are rewritten around the view's control header: facet pills in a toolbar at the top of the view, loading their values with result counts on first open, and a chip row below it whose pivots are discovered by querying the whole result set in both directions - outgoing properties of the results and properties pointing at them. Applied steps read via/linked by next to the facets, and removing one rewinds the view to that step. The action bar section describes the three zones: Create and Add on the left, breadcrumb pills with document-type icons and the last-modified timestamp in the middle, and the Actions, layout-mode and export dropdowns on the right. The settings menu moves to the navigation bar section, where the button now lives. The guides that pointed at retired placements follow: Edit/Delete/Save as go through the Actions dropdown, File/Query/Import creation goes through the creation bar at the bottom of the page (the action bar's Create only offers Container and Item), and the editor toolbar mounts below the action bar. The right sidebar section becomes Backlinks: every block whose resource has a URI carries its own links popover, loaded on first open - in Properties mode that is one per resource in the document. The block data model sheds the left/main/right column convention for the card description, and the left sidebar is documented as the flyout it is, with the stale narrow-screen claim dropped. Graph-scoped updates state the PATCH contract: a single update operation, in the DELETE/INSERT-with-WHERE form or the DELETE WHERE shorthand, everything else answering 422 - matching DocumentHierarchyGraphStoreImpl. The hand-written "On this page" lists in the HTTP API and stylesheets references are dropped along the way. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WK1ssk2A67MYseeH3hWnhf
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S1UnYpbW5aYLj7LbwtyUo1
…syntax Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S1UnYpbW5aYLj7LbwtyUo1
$ldt:ontology, $ac:forClass, $ac:query and $ac:googleMapsKey survive in client.xsl solely as Web-Client compatibility leftovers, and $ac:method and $ac:httpHeaders are server-side layout internals - none of them is part of the contract a custom stylesheet writes against. The bs2:Form entry drops its $ac:forClass condition along with them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011j6GapYTmxZXEq3YK2Z53D
The dataset holds the current version of every document and nothing else, which the versioning reference now states next to the read-only notice rather than leaving readers to infer it from the Git-backed storage. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_011j6GapYTmxZXEq3YK2Z53D
Mirrors the bundled copy. A label added through the form came out untagged beside the tagged values already in the data, because the constructor declared [ a xsd:string ] and the added field is built from the constructor rather than from the data. A SKOS label is natural language, and SKOS's own rule of at most one prefLabel per language is not expressible without a tag. Requires the platform's rdf:langString form control (LinkedDataHub 48f2d7479, Web-Client 7abb022f); without it a langString marker renders as a resource picker.
…oncept you are reading. Mirrors the SKOS package stylesheet from the platform repository, where the change is described in full. In short: the tree roots at the concept scheme rather than at the document's own topic, and on a concept document it climbs to the scheme one hop per request - a constant-size SELECT over a frontier set, both assertion directions per hop, each hop in its own GRAPH because a property path cannot leave one - then descends expanding every child on the path, so a concept with two broader concepts appears under each of its parents. ReadMode only. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BSRUJBNKKjdgJeP3qRZTPd
…d in. Every hierarchy and membership view joined a prefLabel and filtered it to English or untagged. The label is the ORDER BY key and is never projected, so the filter selected nothing - it only dropped concepts labelled in another language, which is every concept created through a UI whose language control follows the browser. Which language a reader wants is decided at render time, where the label chain already picks per value; a stored query cannot know it. The join stays as the sort key; what is given up is the order among one concept's own labels. Synced from the platform's bundled copy (src/main/resources/com/linkeddatahub/packages/skos/ns.ttl), which this file is kept byte-identical to. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BSRUJBNKKjdgJeP3qRZTPd
106 p.lede become p.ldh-body-lg in 57 documents, and docs.css drops the .docs-main p.lede rule that carried them. The class was documentation-local: it resolved only under the static site's .docs-main wrapper, so the same rdf:XMLLiteral bodies rendered a lead paragraph on the generated site and plain body text inside LinkedDataHub, which links no docs.css and has no delivery path for one. .ldh-body-lg is defined in colors_and_type.css, which both contexts link, so one class now serves both. It also corrects a size drift. docs.css set the lead to --fs-lg, which the scale's own role map reserves for headings; --fs-md is the size that map assigns to lead paragraphs, and the m3 skin maps .ldh-body-lg to Material Body Large 16/24. The hero's .lede stays as it is: app.css scopes that rule to .ldh-hero, where linkeddatahub.com's homepage nests it correctly. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01BSRUJBNKKjdgJeP3qRZTPd
…ulary it speaks. packages/skos/ becomes packages/editor/taxonomy/, so the package URI is now https://packages.linkeddatahub.com/editor/taxonomy/#this, and layout.xsl becomes skos.xsl. The old name was the vocabulary, not the capability. What the package ships is a taxonomy editor: a concept tree beside the content, hierarchy views that read both assertion directions, and constructors and constraints for concepts, schemes and collections. SKOS is what it speaks. The directory now says the first and the stylesheet says the second, matching the platform's own rule that vocabulary-typed templates live in a module named for the prefix. editor/ opens a namespace for the editors that follow. update-folder.sh derives the published URI from the filesystem path, so the move IS the URI change; packages/editor.ttl is the container document that makes https://packages.linkeddatahub.com/editor/ exist for taxonomy/ to be created under. Nothing else about identity is a convention: the ontology is found through ldt:ontology and the stylesheet through ac:stylesheet, which is why the stylesheet could be renamed at all. Both READMEs and the packages reference page said layout.xsl was part of the layout contract and drew a single-segment packages/<package-name>/; both claims are corrected. skos.xsl is also brought up to the platform's bundled copy, which had been ahead by the concept tree's heading wrapper and its ReadMode hrefs (ldh:conceptree-href, ldh:tree-node-uri). The copies are meant to be byte-identical, and ac:stylesheet points at this one on develop, so the drift was what a non-bundled install would have fetched. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…m now dresses. docs.css loses its content-typography layer — headings, code/samp/kbd, pre, dfn and the reference-table skin — because those are element rules in the kit now (colors_and_type.css, surfaces.css), which both this site and LinkedDataHub link. The layer existed because the design system covered app surfaces and prose not at all. That meant the same rdf:XMLLiteral body rendered correctly here and undressed in the product, which links no docs.css and has no delivery path for one: about 2,500 elements, the same defect that moved p.lede to p.ldh-body-lg, two orders of magnitude larger. A rule keyed off an element rather than off .docs-main is the only kind that can reach both, since neither stylesheet can know the other's wrapper. What stays is genuinely site-only: the layout grid, nav, child cards, version dropdown, header and footer chrome, the media rules, and the relaxed line height this column wants. The table pair stays too — flow spacing and topped-out cells are what a table in PROSE needs and an app surface does not — but it keys off the element now, because the content carries no class: the 37 tables have dropped class="ac-table dn-comfy" and are plain <table> again, which returns every one of those files to the state it was committed in. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…ads. The 37 reference tables gain a visually-hidden <caption> and all 204 <th> gain a scope. The 43 alerts put their running text in the component's own .ac-alert-text slot, and the two package documents trade class="lede", which resolves in neither context, for the .ldh-body-lg role that resolves in both. A table is the design system's DataTable, and that component requires a caption: a table needs a name, and an aria-label on a grid is less reliable. The captions say what each table lists rather than repeating the heading above it, because a caption is read on its own. scope is native table markup — it is what survived when the component dropped every role= attribute — and the reference tables need all three kinds: col in the head, row for the term column, rowgroup for the colspan bands that open a tbody in the CLI and HTTP API matrices. Cell decorations (.num, .iri, .mono, .var) are deliberately NOT applied. They dress cells holding bare text, and this corpus marks its tokens up as <code> and <samp>, which the kit's element layer already dresses; adding them would put mono on the cell and a chip inside it. The one numeric column, the tutorial's stage number, is an ordinal row label rather than a quantity, so right-aligning it would be wrong. The finding dissolved on inspection. Alert titles are left alone: the component makes the title optional, the pass that harvested the obvious ones already ran, and inventing 38 more would be writing, not conforming. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…nstalled. Three pages disagreed. reference/command-line-interface.ttl described the `packages` group in prose while omitting it from its own command table; both reference/administration/packages.ttl and user-guide/manage-packages.ttl asserted, word for word, that "packages have no command of their own" and that `ldh patch` on the settings document is the whole interface. The commands exist, and the corpus already documents the release they belong to — it marks the `bin` scripts deprecated and records the retirement of the packages/install endpoints, both of which land in the same version. So the CLI page's prose was right and the other two were stale. The table gains a Packages group, and both pages now install with `ldh packages add` and uninstall with `ldh packages remove`, keeping what is still true: the declaration IS the installation, it takes effect on the next request without a restart, and the permanent form is the same triple in config/dataspaces.trig. The READMEs were further behind than the pages they link to. packages/README.md still documented "installation-time composition via JAX-RS endpoints" with a tick-and-cross list claiming no dynamic loading at request time — the opposite of what the reference page says — plus install-package.sh, the master-stylesheet mechanism, POST /packages/install and /uninstall, an instruction to restart Docker, and an unresolved TODO. The root README pointed at the deprecated bin scripts as the prerequisite. Both now describe the mechanism that exists. Also corrected: create-resources.ttl posted schema:category at a document URI where every other example in the corpus uses the #this fragment, teaching the exact confusion between a document and its topic that the data model is careful about. The demo's own import mints categories as <graph>#this. Not fixed here, because it is a migration rather than a correction: demo/unesco-thesaurus/install.sh still calls install-package.sh, and the install tooling has 24 call sites on deprecated bin scripts and none on ldh. Converting one line would leave a single ldh call among them, and would fail on authentication — ldh takes a PKCS12 keystore where these scripts pass a .pem. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…intro. Copenhagen and UNESCO were written in a register the design system names as its anti-example: "Copenhagen is renowned as one of the world's most bicycle-friendly cities", libraries as "a cornerstone of community life", concepts as "a distinct unit of thought or terminology". The system asks for confident, technical and unmarketed copy that states the mechanism rather than the outcome, and Northwind's own blocks already do — each paragraph earns the chart it introduces. Rewritten from the data. The bicycle racks carry a capacity, a street address and a district, so the map reads as supply rather than as points. The libraries carry a postal district and a link out to the city's own page. The toilets carry opening hours and step-free access, which is what makes their map answer availability and not just location. Counts are the real row counts of the imports. Both apps also told the reader to "explore the map below" and to "use the interactive map below", which the block underneath was already showing. Structurally, seven of Copenhagen's ten containers had no prose at all while two had hand-written intros — charging stations, parking, places, playgrounds, toilets, schools and sports centres opened straight into a map with nothing said about it. All ten have one now, and every demo block opens at h2 inside .ldh-section, the shape Northwind's 23 blocks use; the five that opened at h3 with no h2 above them are gone. 37 blocks, 37 h2, 37 .ldh-section. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
… that were really sentences become sentences. Converted, because each term carried the same positional facts and the list was hiding a grid: the triplestore service properties (property, endpoint, required, example — exactly three dd per dt, in that order every time), the dataspace service fields (required-ness parenthesised into every description), and the shared CLI parameters, whose entries carried two dt each for the short and long form and sat beside an environment-variable table doing the same job. Dissolved into prose, because a list of one is a sentence: ENABLE_LINKED_DATA_PROXY, JAVA_OPTIONS, the applications' Stylesheet property, the Title step in create-documents, and the two byte-identical Value steps in create-content. Deliberately NOT converted. The twelve environment-variable lists in configuration.ttl stay lists: most terms carry several dd — google_client_id has two, credentials has three — and their defaults live inside the prose rather than in a column, so a two-column table would flatten notes into one cell. The layout modes in user-interface.ttl stay too: one description each, of wildly unequal length, where a table would make one enormous cell. With the kit now dressing dl in both contexts, appearance is no longer a reason to convert anything — only information shape is, and these are shaped like definition lists. The ten <pre> blocks flagged as "not code" are all left as they are. Every one holds a literal artifact and is framed as such: the CSV samples and the imports manifest are the contents of named files, each with a Source link to that file directly beneath, so a table would lose the commas the reader needs in order to write the mapping; the rest are directory trees, a shell session and response bodies whose line structure is the content. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
tutorial/insight.ttl put <h2 id="chart"> and <h2 id="view"> — 78 of the page's 110 content lines — inside the "From the command line" panel, which is not the one selected on load, while the default panel held a single paragraph. A reader arriving at the stage saw that paragraph and then "What you now see", which refers to a chart and a view they were never shown; a link to #chart landed on a heading inside a hidden panel. The nesting was inverted. Those two headings name what the stage BUILDS, which both routes build, so they belong to the page and not to one route. The sections own the tabs now, one strip each, and the browser and command-line routes sit inside the section they are routes through. The other three pages with headings inside tab panels are left alone, and the difference is what the headings name. get-an-account's tabs are three mutually exclusive ways to get an account, and its headings are the steps of one of them; version-history and create-resources tab a user-interface route against a command -line one, and their headings — "In the History dialog, click Restore" — belong to the route they sit in. In all three the headings are in the DEFAULT panel, so the page reads correctly on arrival. Inverting them would also have meant writing command-line content that does not exist: version-history's CLI panel is one paragraph covering all three Memento roles and does not cover restore at all. Five sections whose only child was another section are merged: "Management" over a lone "Command line interface" in the ACL, ontologies, packages, CSV import and RDF import references. Nothing linked #management-cli; #management keeps its id, which five pages do link. The action bar's reference said its controls fall into three zones and then spent four paragraphs on them, because the timestamp — "next to the breadcrumbs" — had a paragraph of its own. It joins the middle zone, so the three paragraphs now open with "On the left", "In the middle" and "On the right". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…ally too. reference/dataspace.ttl introduced its Northwind block as "the Northwind Traders dataspace, as declared in config/dataspaces.trig" — a claim about shipped configuration. The entry it was quoting exists, but as a testing leftover: the out-of-the-box stack populates nothing behind that origin, and the tutorial does not create it, because the tutorial writes everything to $LDH_BASE, which is the root dataspace at localhost:4443. So the block is reframed as what it is and what the corpus needs it to be: the worked template that manage-dataspaces#create-dataspace asks the reader to follow and never shows. The Turtle is unchanged, origins included — a reader adding a dataspace to the default stack really would write localhost:4443, and the <app>.demo.<host> shape is the one the demo site deploys, so the same declaration carries from local to public. The tutorial deliberately keeps building into the root dataspace. Requiring two TriG edits and a restart before a first document would put configuration in front of someone whose goal is to run the thing out of the box. reference/http-api.ttl keeps its federation example on the public demo hosts, which are reachable with no setup and make the cross-INSTANCE case concrete, and gains a sentence saying the mechanism is about origins rather than instances: two dataspaces on one LinkedDataHub federate the same way once both are declared. Going local there would have meant importing the UNESCO thesaurus first, which is a heavy prerequisite for one example. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019EptZ6w7rd883q8N4CMorR
…ote behind. Scripted Playwright scenes against a running instance, plus the tooling to pace, review and still them. A scene is a workflow rather than a tour: it opens on a question somebody would ask, each step follows from the one before, and it ends by answering it — by writing the finding into a ContentMode document, so what survives the recording is an artefact and not a memory. The five here build their pages out of interleaved prose and evidence, which is how the Northwind pages are built already, so the result belongs where it lands. Five rules turned out to matter more than any amount of scripting, and every one of them came from a scene doing something a person would not. A URI is never known in advance: to reference a resource the scene navigates to it, presses the app's own copy control and pastes, because a value appearing in a field from nowhere is a step the audience cannot reproduce — synthesising a ?mode= URL is the same violation with the address bar. The copy takes the inner resource and not the ldh:Object wrapping it, or the embed nests and renders two stacked headers for one piece of content. No gesture is performed whose state is already current, so a view already in Grid is not switched to Grid for the camera. Presence in the DOM is never availability — a collapsed details, the ContentMode-gated add buttons, the inert hover-opened drawer, canvas-drawn map pins and hover-revealed copy controls each need a real gesture first. And every locator is scoped to the active pane, since inactive tabs stay in the document and an unscoped one will happily drive a tab nobody is looking at; that single mistake accounted for more failures than the rest together, and it hid well because it only bit where a comma selector let one alternative escape the scope. Pacing is measured rather than declared. A uniform speed-up scales pointer travel and typing along with the waiting and reads as fast-forward, so the transcode detects static stretches with freezedetect and speeds only those, taking 163s of capture down to 65s with every gesture still at 1x. The marks sidecar each scene emits — an edit list keyed by beat — is remapped onto that timeline piecewise, which is also what lets the reviewer pull frames from the paced file without every late beat landing past its end. The reviewer exists because three defects in a row were invisible to the beat log. A page whose prose said twenty territories beside an evidence block rendering nine employees, since facet state is not persisted and the embedded view reverts. A run that died after two of sixteen beats and reported green, because the beats that ran had succeeded. And an embed that nested an Object in an Object while every step reported success. So it now requires the terminal beat, counts the document's blocks, and resolves each embedded value against its source to say whether it reached the resource or its wrapper. FINDINGS.md records what the scenes ran into that is the product's rather than theirs: facet values that never load on large containers, where the whole cost is a label lookup over an unbound graph that cannot match; a chart axis bound to a URI-valued variable rendering raw escaped markup; and the parallax step chip's ×, which is a dead control that throws.
…her than where it showed. A form control is found by the property name beside it and never by position. LDH's inline forms carry no label elements — each field is a .ldh-prop-group holding its own name — so filtering anything coarser matches the container that holds every name and then takes its first control. Three call sites got away with it and the fourth did not: on the View form the first control is the query combobox, so filling "Title" typed the title into the middle of the URI that had just been pasted, leaving a corrupt reference that still looked plausible. field() is now the only way in, and fill() reports which control it wrote to, so a mistargeted write shows up in the beat instead of only in the video. The mode guard never once fired. currentViewMode matched \bGrid\b against a toggle that reads "grid_viewGrid" — a Material ligature running straight into the label, so there is no word boundary in front of it — and quietly returned unknown every time, which meant the "already in this mode" branch was dead code and every switch went ahead. It compares against the tail of the string now. The toggle text had been sitting in an early probe's output the whole time. A query typed on camera is multi-line and shaped the way somebody writes SPARQL, because the query is the artefact the scene is making: if it looks machine-generated the claim that a person could have written it goes with it. It is typed deliberately unindented, since the editor auto-indents and supplying our own would double it, and closing braces are typed normally because CodeMirror types over the ones its auto-closing inserted. All of which fails silently, so the text is read back from the editor and compared — the beat says "11 lines, read back and matching" or it says nothing of the sort. And the copy takes the inner resource rather than the ldh:Object wrapping it, with an optional name when a page holds several. Two scenes had been embedding block URIs instead of view URIs, which nests an Object in an Object and renders two stacked headers for one piece of content; N2 was copying the same URI for its query and its view, so the embed showed the query. The reviewer now resolves every embedded value against its source and says whether it reached the resource or the wrapper, because every beat passed while this was wrong. FINDINGS gains a fourth: YASQE asks prefix.cc for its prefix list over http from an https page, so the browser blocks it as mixed content and the editor comes up without prefix autocompletion, telling nobody.
…of the bytes that ship. Every shot is cropped to its subject: a chart block is a picture of that chart block, not of a page with a chart somewhere in it. A page-level shot stays page-level — the application layout, the landing page, the unauthenticated app — and everything else names the element it is about. Uploads are content-addressed, so a reference is only valid for the bytes that are served. Stills are published as WebP at 2240px, twice the docs' content column, which stays sharp on a retina display at a quarter of the PNG's weight; the clips already carried CRF 20, yuv420p and a leading moov atom, so they are copied rather than re-encoded. A clip is a <video> and never an <object>. Item.java serves every upload with default-src 'none'; sandbox, the LNK-011 fix — an <object> loads the file as a nested document, so that CSP governs its own media load and Chrome blocks it while Firefox does not. A <video> is a subresource of the page and loads normally. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
… page. The runner took viewport screenshots, so a picture captioned "a chart block" spent half its pixels on the navbar and the block was cut off at the fold. A shot now names its subject — a selector, a locator, or a pair whose union it spans — and the frame is that element's box. Two things that cost three attempts each are written down in the code. Sticky chrome paints OVER an element capture, because Playwright scrolls the subject into view and the header is still on top of it; every fixed or sticky element that is not the subject, inside it, or containing it is hidden for the shot. And a union has to measure its boxes in page coordinates atomically: scrolling to the second subject moves the first one's viewport rect, so boxes read at different scroll positions cannot be combined. Shots that document an act of authoring — annotating a word, composing blocks — opt in with writes: true, and the run provisions a scratch container, uses it and removes it. It resets rather than creates, so a run that dies before teardown does not poison the next one, and nothing depends on a document some earlier session left behind. fill.mjs is the step that was missing: it writes the published files into the .ttl sources and rewrites them when they are re-shot. The caption is the key — it is already in the markup as @alt or @aria-label — so a slot is found either by its placeholder or by the element carrying it, and a re-shoot just swaps the hash. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…next person will look. The rig had no reader-facing account of the docs shoot: what the manifest is, why a shot asserts a want before it captures, why masters are not web assets, and why optimisation has to happen before hashing when the hash is the address. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…a person would say. A view without a title is only addressable by URI, which means a scene or a shot has to know the URI in advance — and copying one out of band is the thing the rig is not allowed to do. With a title, the same view is found in any combobox by typing what it is called. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…arn what the app actually does. Each scene is a workflow answering a question, so every step has to be one a person could take: a resource is found by the name it carries, a mode is switched with the control that switches it, and a block's URI is copied with the button that copies it. The helpers grew where the app turned out not to work the way the scenes assumed — collapsed disclosures, hover-gated controls, inactive panes still in the DOM. Adds annotating a word with a statement, which is how an arbitrary triple is written through the UI: the resource form is ontology-driven and offers only what the class declares, while prose annotation is not constrained by the subject's class. Adds the graph helpers, the cold open that lifts a scene's payoff to the front, and a sixth scene. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
RULES.md is the set a scene kept breaking and rediscovering: no URI is known in advance, copy the inner resource rather than the Object wrapping it, presence in the DOM is not availability, scope every locator to the active pane. FINDINGS.md gains the product defects these runs walked into — the endpoint that could not render HTML, and a document-scope Map that draws Null Island when there is nothing to plot. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A selector belongs to the app, not to a guess, so each of these dumps some corner of the running instance — the chart create form, the mode buttons, the typeahead, the query block, federation, the graph. probe-viewform stops asking about a form it no longer needs. inventory.mjs lists what a dataspace actually holds, which is what a scene has to consult before inventing a question about it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
ldh:container was deleted from the platform - the property, its constructor input, the ontology query's binding, the data-container stamp and the RDFa fallback read. The Views reference was the last place in either repository that still mentioned it: a paragraph presenting it as the way to get a Create button, and an example carrying the one absolute https://localhost:4443/products/ value that deletion found. A reader following either would have declared a property nothing reads. The replacement text corrects two further things the old paragraph got wrong independently of the property. Inline creation is for views attached to a property in the application's ontology, not for any view that carries some metadata: the class to construct comes from the attaching property's rdfs:range, or its rdfs:domain when the attachment is inverse, and a view block authored by hand in a document has neither. And the container is now derived from the view's own results - a second query built from its SELECT with LIMIT, OFFSET and ORDER BY removed, projecting the containers of the documents the existing solutions are described in. One container renders the button, none and more than one render nothing, and the three cases are spelled out because the two silences are the ones a reader will otherwise read as a bug. The showWhenEmpty example keeps its shape minus the deleted line. It stays valid for a document-authored view, since ldh:showWhenEmpty does have the RDFa fallback read that data-for-class never had. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The concept tree's three promises passed their failures to the one-argument ldh:promise-failure, which LinkedDataHub replaced with one that takes the element to report in and a headline (937a215f8): a refused parents query reports on the tree's root row, a refused step on the row it was stepping into, and a refused children query in the list the children would have filled. The package bundled in LinkedDataHub changes identically. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…ints at a category, a territory and an employee. schema:City gets a constructor (title, name, identifier, coordinates, region), so the Create button on a region's city list produces a form instead of a bare type chip. Three inverse views join the six that existed: products in this category, employees serving this territory, and an employee's direct reports. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…a loop where they can. The scenes are rewritten from the rules: each starts on a document whose loaded render is the money shot, explains that state in a content block before any work, and moves by visible clicks from there. Where the app lets a scene write into the data — Houston, created from the region's own list — the closing beat re-reads the evidence it opened on. The rig learns what the takes taught it: pins are located by resource URI through the captured OpenLayers maps, graph nodes are aimed at by the canvas's own hover event rather than a projected point, queries are typed line by line past the editor's completion popup, the head of each cut is trimmed to the first beat instead of a cold open being spliced in, and a scene resolves its inputs over SPARQL at start. Retired scenes and the probes that answered the questions are kept. FINDINGS gains the empty constructor form, the restart that drops package imports, and table headers that are not sort controls. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…tead of winning the walk. The platform split ldh:RenderRow into a sealed walk and the open leaf mode ldh:RowHook it asks at every element, so the rule that reveals the open concept moves to that mode and drops the priority it needed to outrank the walk. The README now describes where a package stylesheet is composed - right above hooks.xsl, the open modes' module, and below everything else - and what a rule in each open mode owes. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Person, Corporation, Order, OrderItem, ParcelDelivery, DeliveryEvent, Offer, the three price specifications, PostalAddress, ContactPoint, ProductGroup and Place join Product and City, each with the literal fields its instances carry — dates as xsd:date, quantities as xsd:integer, prices as xsd:float — and typed lookups for what they link to. DeliveryMethod is an enumeration and stays without one; enumerated values such as order status are set on the instance afterwards. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…. One fuseki service holds an end-user and an admin dataset for every dataspace, named after the dataspace origin with the deployment host dropped and the role appended (northwind-traders.demo.end-user, and so on; the root dataspace keeps plain end-user and admin), declared in config/fuseki/config.ttl and stored one TDB2 directory per dataset under fuseki/. The configuration reference replaces the fuseki-admin and fuseki-end-user sections with the single fuseki service and documents the new egress forward proxy: outbound SPARQL SERVICE and LOAD from both Fuseki and the platform pass through it, public destinations allowed and loopback, private and link-local refused, so federation with public endpoints works while a query cannot reach another dataset, the cache or the platform. The dataspace reference's service example points at a per-dataspace dataset, the dataspace-management guide gains a step to declare the datasets and bind each application to its own service, the setup tree notes one TDB2 directory per dataset, and the triplestores reference's quad-store example follows the single-fuseki endpoint form. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Docs: single Fuseki with a dataset per dataspace, and the egress proxy
… property or datatype wins. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.