A computed route across a metropolitan graph

Data-layer infrastructure in the Elara-Cortex facility

Elara-Cortex Technical Report · TR-2026-05 · 19 July 2026

A like-for-like comparison of route outputs and operating cost between a self-hosted routing engine and consumer map APIs.

K. Lekola. Elara-Cortex Mathematical Infrastructure for Complex Systems · New Jersey and Johannesburg · route.elara-cortex.com

Abstract. Operators of fleets and logistics systems commonly buy routing as a metered consumer-map API. This report examines the alternative, a self-hosted routing engine, on the two questions that decide the substitution: are the outputs equivalent, and what does each architecture cost to operate. On captured side-by-side comparisons the Elara engine's distances agree with the consumer service to within 0.3–1.0% on the same routes on the same day. On deployment properties the architectures differ materially: the self-hosted engine runs offline and air-gapped, retains all location data in-house, is deterministic and auditable, imposes no caching or storage restrictions, and updates a route in place when a road changes (302× cheaper per change; 97% of changes require no re-solve). A worked cost model shows a single daily 500-stop distance matrix costing, at published list pricing (assumptions and re-verification in §4), on the order of $450 000 per year on metered pricing against one small self-hosted server; every assumption in the model is stated and the reader is directed to the current rate card to re-verify. Boundaries are stated in §6: live-traffic estimation and consumer-facing feature breadth are not claimed equivalent.

1 · Method

Route-output capture. June 2026. Each Elara figure is produced by the production engine; each consumer figure was read off Google Maps for the same origin–destination pair on the same day. Distances are compared like-for-like; consumer ETAs depend on live traffic models and are not compared (§6).
Deployment properties. Table 2 compares documented, externally checkable properties of the two architectures (hosting, data flow, determinism, caching terms, update behaviour).
Cost model. §4 builds the worked example exclusively from the provider's published list pricing at retrieval date, with the arithmetic shown.

2 · Route-output equivalence

Figure 1. Distance of the Elara route versus a consumer map API on the three captured routes (Table 1). Bars show the percentage difference in total distance; all three fall inside a ~1% band, i.e. the route outputs are equivalent to the consumer service on these routes.

Table 1. Captured side-by-side route comparison, June 2026. Delta is on distance.
Route	Elara distance	Elara ETA	Consumer-map distance	Delta
Johannesburg · Sandton → Soweto	37.6 km	55.9 min	37.5 km	0.3%
Midrand → Sandton	15.94 km	21.3 min	16.1 km	1.0%
Tokyo · Shibuya → Tokyo Station	7.44 km	7.7 min	7.4 km	0.5%

The engine returns live distances across the 1 500 hand-tested cities enumerated on the cities page (Johannesburg, Tokyo, Delhi, New York, London and more); distance parity was measured on the three captured routes in Table 1 (0.3–1.0%) and is not asserted worldwide. The formal solution-quality benchmarks against public instance sets, under community conventions, are in TR-2026-01 and TR-2026-02.

3 · Deployment-property comparison

Table 2. Architectural properties of self-hosted vs metered-API routing.
Property	Self-hosted Elara engine	Metered map API
Marginal cost at fleet or matrix scale	~$0 per route on own hardware	metered per call and per matrix element
Offline / air-gapped operation	yes, on own network	no; cloud only
Location-data residency	all queries remain in-house	every query leaves the network
Determinism and auditability	same route every time; inspectable	live service; not auditable by the client
Storage and reuse of results	unrestricted	restricted by terms of service
Custom cost layers (truck, drone, mine)	own cost models supported	limited or none
Behaviour when a road changes	route updated in place: 302× cheaper per change; 97% of changes need no re-solve (TR-2026-02 §4.1)	full re-request per change
Distance accuracy	parity on captured routes (0.3–1.0%)	parity

4 · Operating-cost model

Consider one common workload: a 500-stop distance matrix computed once per day. That is 500 × 500 = 250 000 elements per run, ≈7.5 million elements per month. At the provider's published list price of approximately $5 per 1 000 matrix elements, the metered cost is ≈$37 500 per month, ≈$450 000 per year, for a workload that the self-hosted engine serves from one small server at ~$0 marginal cost. When a road changes mid-operation, the update is transmitted as a delta up to 80% smaller than a full re-send, byte-exact and SHA-256-verified (TR-2026-02 §4).

Cost basis: Google Maps Platform Routes/Matrix list pricing (~$5 per 1 000 elements; 7.5 M elements/month ≈ $37 500/month ≈ $450 000/year), retrieved June 2026. List prices change and volume contracts differ; readers should re-verify against the current published rate card. The comparison is architectural (metered vs self-hosted), not an audit of any specific contract.

5 · Why the outputs match

A reasonable reviewer asks how a small team reaches distance parity with the largest map providers. The answer offered here is checkable rather than rhetorical. First, the routing mathematics is solved, not heuristic: where optimality is provable the engine attains the certified optimum, and on public benchmark sets it leads the strongest free solver (TR-2026-01). Second, consumer navigation distance is determined almost entirely by two things: the underlying road network and the shortest-path objective computed over it. Elara and the consumer services draw road geometry from the same public source (OpenStreetMap-derived data) and both return near-shortest paths under the same distance metric, so close agreement on total distance is the expected result rather than a coincidence. Third, the claim is self-hosted in the literal sense: this site and its live map run on the same engine offered to customers, so any reader can compute a route now and compare it against any map product they trust.

6 · Limitations

Table 1 is a three-route captured sample establishing distance parity, not a statistical survey; the systematic quality evidence is the certified benchmark series (TR-2026-01, TR-2026-02).
ETA and live-traffic estimation are not compared: consumer providers hold materially larger live traffic corpora (stated as their lead in TR-2026-04, Table 3).
Consumer-facing feature breadth (street imagery, indoor maps, place data) is outside scope; the comparison concerns routing as infrastructure.
The cost model uses list pricing at the stated retrieval date; negotiated pricing differs by customer and volume.

References

Luxen, D., & Vetter, C. (2011). Real-time routing with OpenStreetMap data. In Proc. ACM SIGSPATIAL GIS, 513–516.
OpenStreetMap contributors (2024). OpenStreetMap. https://www.openstreetmap.org
Perron, L., & Furnon, V. (2024). OR-Tools (v9). Google. https://developers.google.com/optimization
Uchoa, E., Pecin, D., Pessoa, A., Poggi, M., Vidal, T., & Subramanian, A. (2017). New benchmark instances for the capacitated vehicle routing problem. European Journal of Operational Research, 257(3), 845–858.
Google (2024). Google Maps Platform pricing. https://mapsplatform.google.com/pricing/ (retrieved June 2026).

Data, receipts & series

Series: TR-2026-01 (benchmark paper) · TR-2026-02 (measured performance) · TR-2026-03 (correctness under adversarial review) · TR-2026-04 (local computation in navigation) · EL-2026-01 (evidence ledger)
Verification paths: live map · free API key (run your own like-for-like) · provider rate card (current, for §4 re-verification)
© 2026 Elara-Cortex Mathematical Infrastructure for Complex Systems · New Jersey · Johannesburg · The Elara mathematical framework is a trade secret and is not disclosed (patent applications in progress); every claim here is verifiable without it, and source escrow is available under commercial terms; every result is measured, never simulated.