Resolution: Killed
Caveats: A genuine kill on the conjecture's own staked instrument, with the scope stated plainly. What was tested: on 75 network-routed segments across five provinces spanning the map's north (Germania Inferior), islands (Sicilia, Sardinia, Creta, Cyprus-adjacent), and Africa, converted units-first per the registration (so known sectional unit shifts cannot masquerade as seams), the segment-distance errors are statistically homogeneous - one error distribution describes the map better than five provincial ones. What survives for narrative, not verdict: Sicilia's errors run visibly tight (sigma 0.24 vs 0.63-0.72 elsewhere) and the exact k-means at k=2 isolates Sicilia (silhouette 0.39), but ANOVA p=0.157 - suggestive at best, nowhere near the p<0.01 bar the conjecture staked. Scope limits: the corpus is an honest complete-subset by contiguous region (7 regions, 94 primary segments), not a full census of the map's ~2,700+ figures - province-level seams elsewhere (the eastern segments, the Persian-parasang sections) remain untested; a fuller census could in principle resurrect a compilation-seam signal, and the registry row's ingestion note says exactly where to extend. Transmission corruption of numerals (disclosed at registration) inflates within-province variance and works against clustering; it cannot be separated from source-itinerary error habits with this design. The per-province means all sit near zero (-0.19 to +0.21 log units) - the map's figures are, on these routes, roughly honest; whatever the TP's sources were, their error HABITS do not partition by province here.
Registered before the segment-error corpus exists (triage: adjacent; the corpus is build B2 of GOAL_CONJECTURES_UNBUILT_BUILDS_20260716). Claim under test, per the conjecture's own registered prediction: the Tabula Peutingeriana's segment-distance errors are not homogeneous but cluster by province - a province-level error model beats a single-source model by dBIC > 10, and at least three distinct province clusters emerge whose between-cluster error variance exceeds within-cluster variance - the compilation seams of the map's lost source itineraries surviving as regional error signatures.
Resolution criteria — the registered fine print
Resolution criteria: POPULATION: Tabula Peutingeriana road segments with a legible distance figure in the PD spine transcription (Konrad Miller, Itineraria Romana, 1916; any modern dataset only if its licence is verified open at build time), both endpoints identified to placeable locations (Pleiades coordinates), and a computable real route length. EXCLUDED from primary: figures the transcription source marks illegible or insecurely emended; segments with unplaceable endpoints; open-water crossings; ambiguous-unit rows (all recorded with flags). UNITS fixed per the map's known sectional conventions and recorded per row: Roman mile = 1478.5 m default; Gallic sections in leugae = 2222 m; parasang sections per the transcription source's sectional notes. Because sectional unit boundaries are themselves known compilation evidence, unit conversion is applied BEFORE analysis, and any resulting cluster boundary that coincides exactly with a unit boundary is discounted in narrative (the claim is about error habits, not about units). TRUE ROUTE LENGTHS by fixed hierarchy with method recorded per row: (a) an open-licence digital Roman-road network dataset verified at build time; (b) documented path-tracing along mapped road corridors; (c) geodesic fallback - rows resolved only at (c) are flagged route_confidence=low and EXCLUDED from the primary analysis. PROVINCE ASSIGNMENT: by segment midpoint against ONE published ancient-provinces layer chosen and recorded in the build spec; provinces function as candidate source-regions. STATISTICS, computed only by the shepherd after corpus freeze: log_error = ln(TP_distance / true_route_length); analysis restricted to provinces with n >= 10 primary segments. M0 = single Normal(mu, sigma^2) over all included log-errors; M1 = per-province Normal(mu_p, sigma_p^2); both by MLE on the identical row set; dBIC = BIC(M0) - BIC(M1). Province clustering: k-means over standardized (mu_p, ln sigma_p) for k = 2..min(6, #provinces-1), k* selected by mean silhouette; one-way ANOVA of segment log-errors grouped by the k* clusters. CLAUSE PRECEDENCE, evaluated strictly in this order: (1) KILLED iff dBIC <= 10 (the province-level model fails to beat the single-source model by the staked margin). (2) SUPPORTED iff dBIC > 10 AND k* >= 3 AND the ANOVA gives p < 0.01 with F > 1. (3) Otherwise INCONCLUSIVE - explicitly including dBIC > 10 with k* = 2, and any case where fewer than 3 provinces reach n >= 10 (which also precludes k* >= 3). Narrative (non-binding): province-level (mu_p, sigma_p) table, cluster membership map, unit-boundary coincidence check, sensitivity including route_confidence=low rows.
Known-priors disclosure — what the registrant already knew
Known priors disclosure: Seen at registration: the shepherd triage (adjacent) records that multi-source compilation of the TP is established scholarship - the sectional unit shifts (miles vs leugae vs parasangs) are themselves cited as compilation evidence, which is exactly why units are converted out before the error analysis here - and that single-route distance comparisons across parallel itineraries exist; the systematic province-by-province error-variance clustering was not located. Noetic priors honestly held: expectation that some TP figures are corrupt in transmission (numeral copying errors), which inflates within-province variance and if anything works against the conjecture's clustering signal. No error statistic has been computed: the corpus does not exist, no segments have been extracted, and the build agent has not been launched at registration time.
Method and dataset — how it was measured
Exactly the registered criteria (packet e5c4a79e, ModelRun 26009): log_error = ln(TP_distance_m / true_route_m); analysis restricted to provinces with n>=10; M0 single Normal vs M1 per-province Normal, both MLE on the identical 75-row set, dBIC = BIC(M0)-BIC(M1); k-means over standardized (mu_p, ln sigma_p) for k=2..4 solved EXACTLY by exhaustive partition enumeration (5 province-points; deterministic, strictly optimal for the k-means objective), k* by mean silhouette; ANOVA of segment log-errors by the k* clusters. Script committed at docs/generated/instrument_builds/peutinger-segments/shepherd_analysis.py.
Dataset: The B2 Tabula Peutingeriana Segment-Error Corpus, built THIS DAY by the lane itself from the 'Not yet built' registry exhibit (inst-unbuilt-peutinger-segment-errors): Miller 1916 (PD) segment figures with Itiner-e (CC-BY-4.0) network-routed true route lengths and a DARE provinces layer - 108 candidate rows, 94 included_primary across 7 contiguous swept regions, 5 provinces clearing the registered n>=10 analysis floor (Germania Inferior 20, Sicilia 18, Creta et Cyrene 15, Sardinia et Corsica 12, Africa Proconsularis 10; 75 analysis rows), frozen at sha256-verified commit 488cbaa with the builder banned from computing any error aggregate.
computed 2026-07-16