The Border Has No Point
On 11 September 2026, the Pleiades export held 42,438 place records, including 7,607 marked unlocated. This page draws a region from the admitted neighbours of a selected place, then runs the same rule against known precise places, keeping Pleiades' related class out of the headline because its points can be made from connections. A region is an editorial graph result, not where an unlocated place was.
Parsika is a mountain in Pleiades described as on the border of Carmania and Gedrosia. Its record has two outgoing bounds connections and no usable published geometry. Give the record the neighbours Pleiades already keeps, then decide what counts as evidence and watch the answer refuse or take shape.
Try this first. Parsika is selected and its box is refused, because only one of its two neighbours, Gedrosia, is precise. Under admitted neighbour classes, tick related, circular: Carmania joins, a box appears, and the held-out coverage and median region area beneath it change together. Then set the region rule to buffered convex hull and change the radius.
Loading the frozen record.
The place table and the GIS package were exported hours apart, and the gap is measured two ways. 104 of the 42,438 legacy place ids are absent from the GIS package's own 42,334-row places table (42 rough, 33 precise, 26 unlocated, 3 related); the verifier counts these from the pinned GIS file, which this page does not load. 24 of the 14,983 GIS edges have an endpoint absent from the legacy place table; they stay in the shipped edge file and are counted in the check, but cannot join the graph.
I. Put a hand on the graph
Start with the selected record. The dotted question mark is not a guessed coordinate. The dots are its connected records at their published representative coordinates. A precise neighbour contributes the four corners of the bounding box of all its precise location rows; a rough or related neighbour contributes its published representative point. The shaded shape is labelled as where this rule's neighbours are. Grey dots are drawn for context but excluded until you admit their precision class.
Loading unlocated records.
Lookups read the frozen snapshot only.
42 of 42 connection types admitted
admit connection typesthe legacy connection class stays separate
These are the 42 connection types that occur in the paired edge file, including the legacy connection class. The type vocabulary file defines 43; garrisoned_by has no edge in this snapshot.
Loading the graph.
Waiting for the held-out audit.
Rough or related records are admitted in this view. Their published points can be derived from connections, so this is a circularity demonstration, not a headline truth test.
II. First reproduce the source's point
Pleiades' representative-points help page says that some places have points calculated from connections. The published geo.py source tries four typed stages in order and stops at the first that finds a precise neighbour: stage 1, inbound part_of_physical, at and on; stage 2, inbound part_of_regional, part_of_analytical and part_of_admin; stage 3, near, intersects, bounds, flows_into, route_next, abuts and communicates in both directions; stage 4, inbound member and dependent. Only then does it fall back to any connected place with a precise extent. The engine rebuilds that order with one difference it cannot remove: geo.py averages the centroid of each neighbour's convex hull of its precise geometries, while this page, which ships only bounding boxes, averages each neighbour's published representative point. The two coincide when a neighbour's precise geometry is a single point. So the page does not claim to reproduce the deployed service.
Lebadeia (540907) prints [22.87072996666667, 38.43363023333333]. Its three inbound at neighbours average to that point in the engine.
Acarnanian League (231467659) prints [20.976275666666666, 38.848851]. Its three inbound part_of_admin members average to that point in the engine.
Waiting for the engine.
Waiting for the aggregate anchor.
Waiting for the stage counts.
Headline. The admitted-neighbour rule places 1,566 of 2,152 eligible known precise places inside its reconstructed regions, or 72.8%. This page audits its own region rule rather than calling it Pleiades' algorithm.
The named anchors are a reproduction test. They are not evidence that a box around a connected place is where an unlocated place was. Pleiades says: "It is a very coarse-grained relationship, but it does not extend to cover more general notions of proximity."
III. The dismissal, answered by a further result
The dismissal is fair: a connection is an editorial relation, not a coordinate. So the deeper layer hides a precise place's own point, runs a connection rule from its neighbours, and measures how far the candidate lands from the point the table actually publishes. This is an error distribution on places whose answer is known. It is not a guessed location for any unlocated record.
Waiting for the held-out audit.
This panel reads each precise place's direct links only, never the place itself, so the base panel's choices do not change it.
largest misses
- loading
where the cascade fired
| stage | rows |
|---|---|
| loading | loading |
Loading the three rules.
Waiting for the frontier.
The frontier reruns whenever a base control changes and uses every base choice (direction, depth, neighbour classes, truth classes, buffer radius); this selector sets only its connection types, and by default it follows the base panel, so its box row matches the held-out panel. The box is the axis-aligned bounding box of the admitted points. The convex hull is their smallest convex polygon. The buffered convex hull is every point within the radius of that hull, measured in a local equirectangular projection centred on the points; its area is the hull area plus the perimeter times the radius plus pi times the radius squared, and its dashed outline on the graph is a drawn approximation. The base rate is the mean, over eligible regions, of the share of all 30,242 precise published points inside the region, not the held-out hit rate. Same-site hits are counted because a tiny box can be one archaeological site wearing several editorial names.
A related place's published point was computed from its connections, so testing a connection rule against it partly grades the rule against itself. This panel is the only place the page admits related truths. Its four rows are fixed at the default rule (all 42 types, both directions, one hop, axis-aligned box) and do not follow the base panel.
| control | truths | eligible | coverage | median area |
|---|---|---|---|---|
| not run yet | ||||
Not run yet.
Waiting for the transfer audit.
Like the error panel, this reads direct links and ignores the base panel's choices.
loading unlocated records have at least one precise neighbour in the frozen graph and still have no published representative point in the place table. The Tomb of Castor (303418080) is one worked case: it links at to Sparta, yet the source record remains blank. The page lists the result and does not guess why.
- loading
The check
Input record. The frozen snapshot is loading. The legacy table's 104 ids missing from the GIS places table are counted by the verifier, as stated above.
Named anchor. The engine is loading the source point. The aggregate related-place reconstruction is reported as a result, not treated as a publication truth. The verifier blanks every related place's published coordinates in a copy of the record and requires the same candidates, stages and points, so those coordinates are targets, never inputs.
Current free choices. The controls are loading. Admitted types, direction, depth, region rule, buffer radius, neighbour classes and truth classes are real engine parameters, and each one is declared in the page's record. The error, transfer and blank-list panels read direct links and ignore them; the circularity panel is fixed at the default rule.
Fault control. The verifier will doctor a copied neighbour row and rerun the unmodified cascade.
Refusals. The page refuses a box or unbuffered hull with fewer than two admitted neighbours, an unbuffered hull on collinear points, and any region with no admitted neighbour (the reach readout counts those records). It refuses an id outside the frozen snapshot and an id that already has a published point. It keeps related truths out of every figure except the labelled circularity panel, and it never labels a derived region as where an unlocated place was.
Uncertainties. The legacy place table and GIS edge table are a dated pair, not one atomic export, and both measures of the gap are shown above. Precise extents are the bounding boxes of every precise location row, derived from the undistributed 39.5 MB locations export, so the cascade averages representative points where the source averages hull centroids. The source code and the exported points disagree for some places, so the aggregate match rate controls the wording. A published related point, admitted as a neighbour, may itself have been computed from connections, and nothing in the export says which ones were. The place table's time-period fields are not used: no rule here asks whether two connected places existed at the same time. Areas and buffer distances use a local equirectangular projection, which is a comparison device, not a survey. The region is a statement about Pleiades' editorial graph. It is not a coordinate, a historical boundary, or evidence of physical proximity.
The instrument uses immediate redraws and has no motion loop.
Source and extraction ledger: research/on-the-border-of-carmania/README.md. Verifier: research/on-the-border-of-carmania/verify-on-the-border-of-carmania.mjs. Every runtime request is local to this page.