# Feasibility of testing milestone–road alignment · v9

17 October 2026. **A reproducible geometric discrepancy screen is feasible now. An independent historical test of “low alignment” needs stronger location and source evidence than the three labelled screenshot points currently provide.** A visible gap can arise from a misplaced locality anchor, a moved stone, layer selection, incomplete network coverage or a reconstructed road course. Distance alone cannot choose among these explanations.

This release implements a bounded pilot for EDCS-31400384, EDCS-60100066 and EDCS-70200077, preserving all source coordinates and road geometries. The first screenshot has no readable identifier; it is not assigned an identity by guesswork. The pilot uses the entire local road snapshot, independently of current atlas filters. It does not recreate undocumented screenshot layer settings.

## What the pilot actually measures

| Screenshot catalogue record | Imported anchor to nearest modeled land road | Nearest source class / ID | Location evidence and eligibility |
| --- | --- | --- | --- |
| Lukavac, EDCS-31400384 | 0.109 km | Hypothetical /28738 | Ballif says original findspot unknown; linked EDH locality is Lukavica near Milići. Raw point withheld. |
| Karayakup, EDCS-60100066 | 18.8 km | Conjectured /24744 | French records removal from Kilise Mevkii to a village house. Retain village context; no original-position comparison. |
| Bağlıca, EDCS-70200077 | 11.1 km | Conjectured /24162 | Published origin near Boyabat and later Sinop Museum are separate assertions. Conflicting raw point withheld. |

These distances are calculations on the unchanged imported anchors, not offsets of verified ancient milestone settings. The Lukavac point is also 31.2 km from the nearest Conjectured line. Karayakup is 33.4 km from the nearest Hypothetical line and 21.2 km from a Certain line. Bağlıca has no Certain or Hypothetical segment within the 50 km reporting window. Absence within that window is censored, not infinite distance or proof that a Roman road did not exist.

The distinction matters because the atlas's brown dashed lines are **Conjectured**, while purple dotted lines are **Hypothetical**. The original local snapshot has 1,082 Certain, 13,309 Conjectured and 805 Hypothetical land-road segments. User-visible classes and historical network completeness are different variables. A point can look isolated when a nearby class is hidden. The pilot reports all-class and class-specific distances side by side; it does not presume what was switched on in the screenshots.

## Who gave the coordinates?

The [coordinate-lineage audit](alignment-lineage-v9.md) followed each of the 8,483 unique imported IDs back to its provincial raw TSV. All latitude/longitude strings match the merged 2022 source; all 7,810 mapped atlas numerical pairs remain unchanged. The extraction read EDCS map-link coordinates through Lat/Epig; Petra Heřmánková/SDAM compiled the aggregate. This establishes the transmission chain, not the identity of an individual georeferencer.

The inspected structured sources provide **no per-point coordinate creator, creation date, survey/georeferencing method or metric accuracy**. Those fields remain null. A dataset compiler, an epigraphic editor, a person who saw the stone, a village informant, a museum custodian and a surveyor are separate roles. Record dates cannot be reassigned as coordinate-observation dates. Decimal precision cannot supply an accuracy estimate.

Of 7,810 mapped rows, only 3,178 exact coordinate pairs are distinct; 5,861 rows share 1,229 anchors. Shared coordinates may represent genuine co-location, a locality anchor, duplicate cataloguing or another process. They cannot automatically be treated as 7,810 independent findspot observations, and coordinate sharing alone does not prove a centroid or an error.

The provider is EDCS, on the scope established by its extraction metadata; its underlying individual coordinate authority remains unknown. See the [2022 deposit](https://doi.org/10.5281/zenodo.7072337), exact local metadata/code locators and raw file identities in the lineage JSON. The atlas now displays this chain for catalogue records, with case-specific limitations for the three pilot objects.

## Publication checks on the three examples

**Lukavac/Lukavica.** [Ballif (1893)](https://www.cidom.org/wp-content/uploads/2015/12/Philipp-Ballif-Romische-strassen-in-Bosnien-und-der-Hercegovina-1893_opt.pdf#page=47), p. 39, says that the original findspot of the fragment then near Lukavica could not be determined. The table p. 51 no. 46 does not identify it as in situ; Patsch's p. 69 no. 46 supplies the reading from photograph and cast. [Popović (2019)](https://aseestant.ceon.rs/index.php/zrffp/article/download/23025/8360/), p. 202, places the relevant Lukavica context by the Studeni Jadar/Milići route and likewise leaves the exact spot unknown. The EDH snapshot has null coordinates; its responsible editor Gräf is not the demonstrated author of the imported EDCS point. No replacement pin is supplied. CIL III 13316 and XVII/4 442 remain concordances, not fresh direct page collations in this pass. Full limits and retained-source identities: [case audit](alignment-lukavac-v9.md).

**Karayakup.** [French, RRMAM 3.3 (2012)](https://doi.org/10.18866/BIAA/e-03), p. 53 no. 026, records the stone at Hüseyin Genç's house on 26 October 1995, reportedly removed from Kilise Mevkii approximately 8 km north. The assumed Amasia–Ancyra association is not a surveyed original position. No numerical milestone position is given in the entry. Neither the village anchor nor an invented point 8 km north can be used as a measured original setting. The raw village point remains context with comparison withheld.

**Bağlıca.** [French, RRMAM 3.4 (2013)](https://doi.org/10.18866/BIAA/e-04), pp. 86–87 no. 44(C), concerns Bağlıca near Boyabat and Sinop Museum inventory 13.53.70. Wilson and French's observations, and the headman's 1982 origin report, are recorded separately from coordinate authorship. The corpus map-sheet reference is not GPS. The village's Boyabat context is independently described by [Aydoğdu and Bakırcı (2020)](https://doi.org/10.26650/JGEOG2019-0035), p. 4; it does not locate the ancient milestone. No original-position or museum coordinate is invented. Both Turkish sources were directly collated from complete local/archived publisher PDFs; current publisher download failures are recorded in the [two-case audit](alignment-turkey-v9.md).

## A test that could answer the historical question

The claim must be specified before looking for a result: **Do independently located primary milestone settings fall outside a nominated reconstructed road corridor, beyond the positional uncertainty of both sources?** This is different from asking whether arbitrary catalogue points lie far from their nearest digitized line.

1. **Freeze the study and denominator.** Choose a documented region or road corridor, source versions, date treatment, road classes and an explicit collection rule. Record which roads are tested and why. Do not select only conspicuous gaps or pool Certain, Conjectured and Hypothetical lines into one class. Comparing classes across unmatched regions confounds certainty with network density and research coverage.
2. **Resolve physical objects and stations.** Group proven duplicate texts and repeated inscriptions on one physical object. Preserve unresolved identities. Use station/site groups rather than several rows sharing one anchor as independent observations; objects discovered together may also be correlated.
3. **Admit a usable spatial assertion.** Require a sourced original/primary setting, a defensible coordinate or source-derived area, a CRS/method and stated positional uncertainty. Repositories, secondary-use discoveries and village proxies remain separate cohorts. Unknown bounds stay unknown. T1 rock support and T2a primary context do not automatically supply precise geometry.
4. **Check the individual road pair.** Read its source bibliography and the milestone's publication history. [Itiner-e's documentation](https://itiner-e.org/documentation) and [2025 Methods](https://www.nature.com/articles/s41597-025-06140-z) explicitly describe using milestones and LIRE in road reconstruction. Shared evidence can make agreement dependent. A publication-volume overlap warrants review but does not alone prove that this particular stone determined the line.
5. **Establish temporal relevance.** Separate inscription, reuse, movement and discovery dates. Determine whether the line purports to represent road existence, course or a dated physical phase. Sparse route chronology cannot be made contemporary merely by choosing an inscription-date filter. Keep unresolved pairs in an indeterminate cohort.
6. **Measure without snapping.** Calculate distance to nominated candidate geometry and to the nearest eligible network as separate quantities. Use source-derived polygons/areas where appropriate. If genuine point and road displacement bounds u and v exist, a centre-line distance d yields a possible gap interval [max(0,d−u−v),d+u+v]. This is bound arithmetic, not a probability interval. Editorial search envelopes can support labelled sensitivity scenarios, not certified error claims.
7. **Define a decision and assess power.** Select a tolerance from the study's archaeological question and source resolution before examining results. Only a defensible positive lower gap bound beyond that tolerance establishes a spatial incompatibility for that pair; absence of a named road association or incomplete coverage still needs interpretation. Count compatible, incompatible and indeterminate separately. Any class comparison needs matched coverage; any resampling must respect station/corridor dependence. An unconstrained uniform-random point cloud is not an archaeological absence control. Estimate power from plausible effect size and independent sampling units, rather than adopting an arbitrary sample-size rule.

A large raw gap is a reason to investigate. A small gap is not independent corroboration when provenance, original position or source independence fails. The Lukavac example demonstrates how an unsupported locality point could produce a superficially impressive agreement as well as an apparent discrepancy under another class selection.

## Current feasibility boundary

The atlas has four reviewed T1/T2a objects. Three primary-setting representations have map centres, but their envelopes are editorial or approximate and none is currently admitted as a metre-bounded original-position point for comparison. Thirteen other curated findspot locators permit exploratory navigation/distance inspection with unknown accuracy; this does not pass the stronger validation gates. There are currently **zero admitted primary points with a justified metric error bound and zero independently tested pairs**. That is a limit of this admitted dataset, not a claim that no such evidence exists in publications or the field.

The feasible deliverable now is the source-linked discrepancy pilot and a prioritized coordinate audit. A defensible historical study should begin with a bounded corridor whose excavation/survey plans can be georeferenced with documented control, then audit dependence against the nominated Itiner-e segments. This release supplies the admission schema and reproducible measurements rather than a premature low-alignment verdict or p-value.

## Reproduction, atlas use and safeguards

Open **Alignment test & coordinate provenance** in the atlas. Choose one of the three cases and switch the road class to inspect the frozen result, contributor and bibliography. Unknown source errors remain blank. The optional calculator is explicitly a what-if arithmetic scenario; entering numbers does not promote a point into the historical test. Each case links its complete source audit and preserved catalogue record. The ordinary map additionally withholds the Lukavac and Bağlıca coordinates; Karayakup retains its village-context pin with comparison disabled. No source coordinate or route was moved to improve agreement.

Run `python3 tools/build_data.py`, `python3 tools/audit_alignment_lineage_v9.py`, `python3 tools/build_alignment_pilot.py`, `python3 tools/prepare_alignment_v9.py`, and `python3 tools/prepare_dossier.py` in that order. Rebuilding the metric requires Shapely and pyproj; recorded versions are 2.0.7 and 3.6.1. The metric uses per-point WGS84 azimuthal equidistant coordinates, geodesic edge subdivision at most 1 km, a conservative spatial-index candidate window and a 50 km reporting radius. Output distances are rounded after nearest-candidate selection. A separate full-snapshot geodesic candidate scan and 250 m refinement audit the numerical method. Source identities, raw nearest coordinates, software versions and eligibility counts are in `data/alignment-pilot.json`; publication PDFs remain outside the site. No viewport, chronology setting or clustering centroid enters this frozen pilot.