2.1 Addressing in The Gambia

The national digital address. The National Digital Addressing Policy [2], prepared for the Ministry of Communications and Digital Economy by the e-Governance Academy of Estonia, records that the Government “adopted and implemented a digital addressing system using the Google Plus Codes addressing solution” and gives the Gambia Postal Services Corporation (GAMPOST) the integration of the national system into its procedures and its door-to-door delivery. The Digital Addressing Initiative publishes the programme [6]. Addresses are written as a short code with the town, for example F8HC+9G6 Bakau.

Council practice. Kanifing Municipal Council implemented a Plus Code addressing system for about 500,000 residents between July 2021 and February 2022, georeferencing 36,000 properties, and uses it for waste management and property tax collection [4]. Its officials report that rate collectors previously identified compounds by their proprietors’ names, that the codes are now linked to the rates database, that the council declined to replace them with simpler labels because those would lose accuracy, and that sequential street numbers are being introduced as a complement [1]. In January 2023 the Ministry reported 34,464 addresses digitised across Banjul and Kanifing in a 2022 pilot, and 48,770 addresses identified in Kombo North, almost twice the number of properties in the Brikama Area Council database [5].

Postcodes and door numbers. The Gambia has no postcode system in use, and GAMPOST publishes no postcode list; addresses are written with the town. Door numbers are rare. In OpenStreetMap, 19 compounds carry a street and a house number [7]; the model shows those first, exactly as written.

The census. The 2024 Population and Housing Census counted 2,422,712 people, 297,581 households, 201,827 compounds and 547,528 buildings or structures, with an annual intercensal growth rate of 2.5 % since 2013 [8]. Brikama local government area alone holds 47.5 % of the population. Enumeration areas are not open data; we use the census only as an external check on our counts (Section 6.7).

2.2 International practice

The Universal Postal Union’s S42 standard defines the components of postal addresses and country templates for rendering them [9]. It describes how an address is written, not how a code is chosen where streets have no names. The World Bank’s review of street-addressing programmes in 52 African cities between 1989 and 2004 found that, in a sample of 21 cities, only 8 % of streets had names before addressing, and that naming streets, numbering doorways and installing signs cost on average US$5.7 per addressed doorway [10]. Its central lesson for our work is that an address that people do not use decays: plates disappear and directions revert to landmarks.

National digital address systems in the region follow the grid approach. GhanaPostGPS, launched in 2017, divides the country into 5 m squares and gives each a code combining a regional prefix and a grid reference [11]. The Gambia’s choice of Plus Codes belongs to the same family: a code is computed from a position and needs no register.

2.3 Code systems and why none suffices alone

Table 1 compares four kinds of address against the uses an address must serve. A grid code such as a Plus Code is computable, offline and open, but its symbols are hard to say and remember. Random-word systems [12] name squares of a few metres with three dictionary words, which are easier to say, but they name squares rather than compounds, carry no sense of place, and depend on a proprietary word list. Street and door numbering is what people expect of an address, but it presupposes named streets that most of the country does not have [10]. A place name and a number is what Gambians already say (“in Bakau, near the…”); our task is to make it exact.

Short-term memory holds about four chunks [13]. Said aloud, a home code is a name and one or two groups of digits (“Bakau, twenty-three, eighty-nine”), within that span; a Plus Code is eight or more independent symbols.

4.1 Forms of a home code

A home code has four forms that map one to one (Table 3). Let $P$ be the place’s plate name, $N$ its number, $Z$ the zone’s name, $K$ the zone’s key, $I$ the place’s index in its zone and $c$ the check digit. Then

where $I$ is written with four digits and the grouping $g$ splits five-digit numbers for reading:

The order of display is fixed: a council’s street address where one exists, otherwise the Government’s Plus Code of the compound’s entrance with its town, and beside it the home code. The model issues no grid of its own and never presents a home code as the official address.

4.2 Zones and frames

A point $p = (\varphi, \lambda)$ lies in the zone whose outline contains it, tested by ray casting. On an overlap the lowest index wins; a point in no outline (a river bank, the shore) takes the zone with the nearest outline vertex within 2 km:

where $u$ is the projection to WGS 84 / UTM zone 28N (Section 4.12) and $v$ runs over the outline’s vertices.

Each zone carries a square frame used only to order compounds (Section 4.8). With $E$ and $N$ the projected coordinates of the zone’s outline, the frame’s south-west corner and side are

Kanifing’s frame starts at $E_0 = 313\,000$ m, $N_0 = 1\,483\,000$ m and has side $F = 16$ km.

4.3 Compounds

The unit of address is the compound: the plot that one gate serves. Footprints are grouped by four rules applied in order, with a union–find structure so that the result does not depend on the order of the input. A structure is urban when its 250 m tile holds at least 100 structures (1,600 per km²). The gap between two structures $a$ and $b$ with centres $\mathbf{c}_a, \mathbf{c}_b$ and areas $A_a, A_b$ is the distance between centres less the radii of circles of the same areas:

Two main structures in urban fabric are never merged by shape alone: in Serrekunda and Brikama, houses built wall to wall are separate compounds unless a wall is mapped or residents say otherwise. Of the 471,031 codes, 107,632 have more than one structure (70,813 two, 19,254 three, 17,565 four or more). A compound’s centre is the area-weighted centroid of its structures; its entrance is a mapped gate on its enclosure where one exists (17 today), otherwise its centre.

Each structure also carries a use class with a confidence and the evidence for it, following a ladder: OpenStreetMap tags, then N-Ways places inside the footprint, then shape [7]. Heights distributed with machine-learned footprints are rendering values rather than measurements and are not used. The main structure is a house in 370,331 compounds, a family compound in 44,230 and a hut in 33,223; 21,770 remain unclassified. Class identifiers such as res.maison are stable codes of the published data.

4.4 Places

The place word is a town, village, suburb or neighbourhood that people say. The inventory is built in five steps.

1. Landmarks out. A name containing point, island, roundabout, junction, traffic light, bridge, wharf, ferry, crossing, market or airport names no home; 41 were removed. Without this rule Dockyard Point would name much of Banjul, and Sukuta Traffic Light 4,780 homes in Sukuta.
2. Duplicates. A place mapped twice (a point and an area within 1 km with the same name key) keeps its first identifier (140 cases). A GeoNames locality is dropped when an OpenStreetMap place lies within 300 m, or within 1 km with the same name key (279 cases).
3. Brackets. A second name in brackets becomes an alias: Busura (Kunkujang) is Busura, and typing Kunkujang finds it.
4. Index. Within each zone, places are sorted by name key, then identifier, and numbered from 0. The index is frozen.
5. Point to place. A compound $c$ takes the nearest place of any kind within 2.5 km; else the nearest town or city within 10 km; else the nearest place of its zone, ties going to the lower index:

where $\mathcal{P}(z)$ is the set of places of zone $z$, $\mathcal{T}(z) \subseteq \mathcal{P}(z)$ its towns and cities, and $d$ the distance in UTM metres.

The result is 2,347 places: 1,187 localities, 1,072 villages, 38 suburbs, 27 towns, 18 neighbourhoods and 5 cities. The median place has 42 compounds; 419 places have none today, mostly hamlets whose buildings are not yet mapped, and keep their blocks for when they are (Figure 2).

4.5 Plate names

The name in the code must be short enough to say and to paint: at most 12 characters (letters, spaces and hyphens each count one; an accented letter counts one) and two words, linking words (and, of, the) not counted. A trailing number counts in the characters only, as in Brusubi 1:

where $L$ is the set of linking words. A name that fails is shortened in three steps: a trailing number is set aside; generic words (village, ward, town, settlement, area, layout, quarters, island, point, wharf, phase, roundabout) are removed; and if the name still does not fit, every run of one or two whole words that fits is a candidate, and the candidate whose words are rarest across all Gambian place names is chosen:

with $f(w)$ the number of Gambian place names that contain $w$ and the minimum taken in lexicographic order. Balanghar Kerjuri thus becomes Kerjuri: six villages are Balanghar something, only one is Kerjuri. In all, 1,772 names are used as they are, 3 are cleaned of generic words, and 572 (24.4 %) carry a provisional plate name chosen by Eq. (8), with up to three candidates, for their council to confirm. Their codes are issued now; if a council chooses another short name, the name may change once and the number never does, and the official name always stays an alias.

4.6 The name key

Two places whose names sound or look alike must never be distinguished by the listener alone. They share one name key $\kappa$, and every place with the same key shares one national number range (Section 4.7). The key is a fixed composition of foldings,

applied right to left, with the steps defined in Table 5.

So Serrekunda = Serre Kunda, Kerr Mbaye = Keur Mbaye, Sukuta = Sukkuta and Ndungu Kebbeh = Ndungu Kebbe, while Bakau ≠ Bakoteh. The folding errs on the generous side deliberately: joining two names that differ costs only a longer number range, whereas separating two names that sound alike would give two homes the same spoken code. Every official name and bracketed second name is an alias of its place; an alias whose key equals another place’s key, or that two places claim, is dropped (16 were), since otherwise Tukulorr, an old name of Porli Nyangen, would decode to the village of Tukulor.

The 2,347 places fold to 1,854 name keys. Of these, 298 keys are shared, by 791 places in all; the most repeated are Madina (15 places), Amdalai (10) and Kerewan, Misira and Nema (9 each).

4.7 One national range per name

A home code is said without its town, so all places that share a name key draw their numbers from one national range for that name, cut into one block per place. Because a number falls in exactly one block, a name and a number decode to exactly one place.

Block size. A place with $n$ compounds receives a block of three times its compounds, rounded up to the grain $g_d$ of a $d$-digit range and at least one grain:

At the national intercensal growth rate of 2.5 % a year [8] a place triples in $t = \ln 3 / \ln 1.025 \approx 44$ years; at 3.7 %, the pace of the fastest-growing towns in the region, in about $\ln 3 / \ln 1.037 \approx 30$ years.

Number length. The $k$ places of one name share one length $d$, the smallest of 3, 4 and 5 for which their blocks fill at most 90 % of the range, so that a tenth always stays free for new places of the same name:

A three-digit range thus has 900 numbers of which 810 can be allocated, a four-digit range 9,000 and 8,100, and a five-digit range 90,000 and 81,000. A place alone with its name has three digits up to 270 compounds, four up to 2,700 and five up to 27,000.

Blocks. Blocks are cut from the bottom of the range in the order (zone key, place index):

Worked example: Bakau. Bakau is alone with its name and had $n = 1455$ compounds at first issue. Three digits would need $3 \times 1455 = 4365 > 810$ numbers, so $d = 4$, and Eq. (10) gives $b = 100 \lceil 4365/100 \rceil = 4400$: the block is 1000–5399.

Worked example: Brikama. Five places fold to brikama (Table 6). Their three-digit blocks would need 7,460 numbers, more than 810; their four-digit blocks need 7,700, within 8,100. Every Brikama therefore has four digits, and the number alone says which one is meant: Brikama 1001 is in Central River and Brikama 7042 in West Coast (Figure 3).

Frozen. The block table is frozen at first issue. A block is never moved, renumbered or shrunk; a new place of an existing name takes a block from the top of the name’s free space; a full place takes a further block of one grain (Section 4.8). Of the 471,031 codes, 87,662 (18.6 %) have three digits, 251,933 (53.5 %) four and 131,436 (27.9 %) five. The largest place, Suware in West Coast, has 6,928 compounds and the block 10600–31399. No name needs more than five digits and none needs a qualifier. Five small names (Babilon, Dobang, Kenaba and Suruwa at three digits, Ganatown at four) have reached the 90 % limit; a further place of the same name would need the next length or a qualifier agreed with its council.

4.8 Numbering along a Hilbert curve

Within its block a place’s compounds are numbered along a Hilbert curve [18], so that neighbours on the ground receive neighbouring numbers, as house numbers do along a street. The Hilbert curve preserves locality better than row-by-row or Morton (Z-order) orderings [19]; Section 6.4 measures this on the registry.

Cells. The zone’s frame is divided into $2^{16} \times 2^{16}$ cells of side $s = F / 2^{16}$ (0.24 m in Kanifing). A compound’s centre $(E, N)$ falls in the cell

Index. The Hilbert index $H(x, y)$ is accumulated from the most significant level down. For $m = 2^{15}, 2^{14}, \dots, 1$, with $r_x = 1$ if the bit $m$ of $x$ is set and 0 otherwise (likewise $r_y$),

where $\oplus$ is the bitwise exclusive or. Consecutive indices are adjacent cells, so sorting by $H$ keeps neighbours together.

First issue. The $n$ compounds of a place with block $[\ell, h]$ are sorted by $(H, E, N)$, and the compound of rank $i$ (from 0) receives

Because a block is three times its compounds, consecutive numbers are about three apart, leaving two free numbers between neighbours.

Worked example: Bakau 2389. The compound’s centre projects to $E = 318\,287.3$ m, $N = 1\,490\,662.9$ m in Kanifing’s frame, with Hilbert index $H = 745\,833\,093$. It is the 460th of Bakau’s 1,455 first-issue compounds along the curve ($i = 459$), so by Eq. (15)

Its neighbours along the curve are 2386 and 2392. Figure 4 shows the whole of Bakau coloured by number.

After first issue. A new compound with index $H$ takes the free number midway between its Hilbert neighbours $a$ (the issued number just before $H$ in the place’s first block) and $b$ (just after):

A new block is taken from the name’s free space. If the name’s range has no free space left, no code is issued until the council names an extension of the place, which then receives its own block. No issued number ever changes, and a retired number is never given again. A house built between Bakau 2389 and Bakau 2392 receives $\lfloor (2389 + 2392)/2 \rfloor = 2390$.

4.9 Decoding without context

To decode a typed or spoken home code, the engine reads the trailing digits as the number $D$ and the words before them as the name $P$. It folds the name to its key and finds every name family the key belongs to, by plate name or by alias. For each family with length $d(q)$ it takes the last $d(q)$ digits of $D$ as the number and finds the one block that holds it:

Because the blocks of one name are disjoint, at most one place of a family holds a given number, so decoding is unique. A zone name after a comma, and a check digit that a system may send after another comma, are verified but never used to re-route a code. Section 6.1 tests Eq. (18) on every issued code.

4.10 The check digit

The machine form carries one check digit, for systems only. It is computed by Damm’s algorithm [20], [21] over a ten-digit payload, the name family’s identifier $\Phi$ in five digits followed by the number in five digits:

where $T$ is the table of a totally anti-symmetric quasigroup of order 10:

Because each row and column of $T$ is a permutation, every single wrong digit changes $c$; because $T$ is totally anti-symmetric, that is $T[T[c][x]][y] = T[T[c][y]][x]$ implies $x = y$, every swap of two adjacent different digits changes it too. The reference engine’s self-test checks both properties exhaustively on 3,000 random payloads.

Worked example. Bakau’s family identifier is $\Phi = 75$, so the payload of Bakau 2389 is $\delta = 0007502389$. The intermediate values $c_1, \dots, c_{10}$ are $0, 0, 0, 6, 2, 4, 2, 6, 3, 6$, and the check digit is $c = 6$: GM-KMC-0001-2389-6. The check digit binds the name and the number, and the place follows from the block table. It is not printed on plates or shown to people: its only purpose is to catch transcription errors between systems.

4.11 The official address: the Plus Code

The Plus Code of a point is computed, not issued [3]. Latitude and longitude in WGS 84 degrees are shifted to positive values, $\varphi' = \varphi + 90$ and $\lambda' = \lambda + 180$, and written in base 20 with the alphabet 23456789CFGHJMPQRVWX. For the five pairs $k = 1, \dots, 5$ with resolutions $\rho_k = 20^{2-k}$ degrees,

and the eleventh symbol splits the last cell into 5 rows and 4 columns:

The code is the ten pair symbols with a “+” after the eighth, followed by the eleventh. An eleven-symbol code names a cell of $\rho_5/5$ by $\rho_5/4$, about 2.8 m by 3.4 m at The Gambia’s latitude.

Worked example. The entrance of Bakau 2389 is at $\varphi = 13.47841$, $\lambda = -16.6787355$, so $\varphi' = 103.47841$ and $\lambda' = 163.3212645$. Eq. (21) gives the pairs $(5,8), (3,3), (9,6), (11,8), (7,10)$, written 7C, 55, F8, HC, 9G, and Eq. (22) gives $q = 4$, written 6: the full code is 7C55F8HC+9G6. The short form drops the first four symbols, which are shared across Greater Banjul, and is written with the town, F8HC+9G6 Bakau, as the Government publishes it (Figure 5). A short code is made whole again from the town’s position by the specification’s recovery rule. The engine reproduces the specification’s reference value (8FVC9G8F+6X for a point in Zürich) and its recovery examples.

4.12 The projection

Distances, Hilbert frames and grouping rules are computed in WGS 84 / UTM zone 28N (EPSG:32628). The Gambia lies close to the zone’s central meridian, $\lambda_0 = -15°$, and its eastern end lies 1.25° east of it, so the projection’s scale error inside the country stays below 0.1 %. The engine uses Krüger’s series to fourth order in the third flattening $n$ [22]:

with $a$ and $f$ the WGS 84 semi-major axis and flattening, $\alpha_1, \dots, \alpha_4$ Krüger’s coefficients in $n$, $k_0 = 0.9996$ and $E_{\mathrm{F}} = 500\,000$ m. Here $A \approx 6\,367\,449.146$ m and $n \approx 0.0016792$. The inverse uses the matching $\beta$ and $\delta$ series; the self-test checks that 2,000 random points in the country return to within $10^{-9}$ degrees.

5.1 Reference engine and conformance vectors

The normative specification is engine/SPEC.md; its executable reference is engine/reference/gm_engine.py, written for Python 3.9 or later with the standard library only. The engine implements every equation of Section 4. Its self-test covers the projection round trip, the two Damm properties, name-key equalities and inequalities, Plus Code encoding, decoding and recovery against the specification’s examples and 2,000 random round trips, the disjointness of every name’s blocks, and the unique decoding of a number in each of the 2,347 places. The file vectors.json holds conformance vectors (Plus Codes of public places, check digits, spoken numbers, name-key pairs, decodings of repeated names and plate names); any implementation must reproduce them value for value. The issuer’s production service implements the same specification; Section 5.4 reports the tests of parity between the two.

5.2 The registry

The registry in registry/ is the complete record of issued codes. Each zone has a file of codes, with one row per code: its forms, place, number and check digit, status, issue mode and time, entrance and its kind, Plus Code, use class and compound profile, official address where one exists, the compound’s centre from which numbers are ordered, and its grouping rule. A second file per zone holds every structure with its class, confidence, source, size and footprint geometry. The place table lieux.csv.gz lists every place with its blocks, name key, family, digits, aliases and position. MANIFEST.json gives the counts and the SHA-256 digest of every uncompressed file. Nothing is withheld: all 471,031 codes and all 672,836 footprints are published.

5.3 The registry server and API

A registry is useful to a ministry, a post office or a developer only if it can be served. The repository therefore includes server/, a deployable HTTP service written with the Python standard library that reuses the reference engine for every rule. It imports the registry into SQLite, verifying every digest and count and refusing a mismatch; serves a JSON API under /v1; and exports the registry back to the published format, byte for byte. Table 7 lists the endpoints; docs/api.md gives a real request and response for each, and server/openapi.yaml the formal contract. With Docker the whole deployment is docker compose up. On the full registry the import takes about 18 s and produces a 524 MB database, and an export is byte-identical to the published files.

Issuing follows the issuer’s rules exactly: a point within 3 m of an issued entrance, or an outline overlapping an issued footprint, already has a code; an outline must enclose between 12 m² and 2 ha; the place is the nearest by Eq. (6); the number follows Eq. (17), cutting new blocks when needed; a new code is provisional and issued on demand; and each key has a daily limit. A deployment issues codes only when the State or N-WAYS LIMITED has designated it as the country’s issuer; every other deployment runs read-only. Two independent issuers would sooner or later give the same number to two homes.

5.4 Parity with the issuer

We tested the server against the issuer’s own implementation. Five hundred issues in sequence across all seven zones, 139 of them after extension blocks had been cut, produced identical numbers, forms and blocks. Three hundred random points gave the same answer to the point query as the issuer’s live service, distances included. Two random samples of 51 codes matched the issuer’s code cards field by field, 1,020 fields per sample, with no difference. Seventy-two of these cases are kept as test fixtures and replayed by the server’s tests.

6.1 Uniqueness and decoding

The 471,031 codes have 471,031 distinct home forms (compared without regard to case), and therefore also distinct machine forms. Decoding each home form with Eq. (18), with no zone, town or position given, returned exactly one result for every code, and that result was the code itself. Without the national ranges, if each place numbered its compounds independently from the bottom of its own range, 17,183 codes would share their home form with a code in another place of the same name; 165,538 codes (35.1 %) lie in places whose name repeats somewhere else in the country.

6.2 Lengths

Plate names have between 2 and 12 characters, with a median of 8 (Figure 6a). Home forms have between 7 and 19 characters including spaces, with a median of 13 (Figure 6b). The longest arise from two-word plate names with five-digit numbers, such as Sinchu Alagi and Manjai Kunda. Of all codes, 72.1 % have a three- or four-digit number.

6.3 Coverage

Every compound formed from the 672,836 footprints holds a code, in all seven zones. Of the 2,347 places, 1,928 hold at least one code and 419 hold none yet. Nineteen codes carry a council street address as their official address and 17 lead to a mapped gate; the remainder lead to the compound’s centre until a gate is confirmed.

6.4 Ordering locality

We ordered the compounds of every place in three ways on the same $2^{16} \times 2^{16}$ grid and measured the distance between compounds with consecutive ranks, over 469,103 consecutive pairs (Table 8). The Hilbert ordering has the shortest steps at every quantile: its median step of 27.0 m is 12 % shorter than Morton order and seven times shorter than row order, and its 90th percentile is less than a tenth of row order’s. In Bakau the median step between consecutive numbers is 20.8 m, against a median distance of 14.2 m from each compound to its nearest neighbour.

6.5 Typing and listening errors

We applied every single-digit substitution and every transposition of two adjacent different digits to the number of every issued home code (17.35 million substitutions and 1.31 million transpositions) and decoded the result by Eq. (18). Figure 7 shows the outcomes. A substitution produces a number with a leading zero in 2.7 % of cases, falls in no block of the name in 18.0 %, and falls on a free number in 56.0 %; in all these cases the code is refused. It lands on another issued compound of the same place in 22.6 % and of another place of the same name in 0.7 %. The rate of 23.4 % is below the value of about one third that the block headroom alone would suggest, because numbers outside all blocks and leading zeros absorb part of the errors. Transpositions land on another issued code more often (34.1 %), because they keep the digits close to the original.

These residual errors are handled by two mechanisms. Before routing, applications show the place and zone of the decoded code, so that a slip landing in another place (Brikama, Central River instead of Brikama, West Coast) is visible to the caller. Between systems, the machine form’s check digit detects every single substitution and every adjacent transposition, as Section 4.10 guarantees; swapping the last two digits of GM-KMC-0001-2389-6, for example, gives GM-KMC-0001-2398-6, whose check digit should be 7, so the code is refused.

6.6 Plus Code accuracy

For every one of the 471,031 codes, encoding the published entrance point gives the published Plus Code, and recovering its short form from the position of the code’s place gives the full code again. The official address can therefore be reproduced from the registry and from the town written after it.

6.7 Comparison with the 2024 census

Table 9 compares the registry with the census counts by local government area [8]; the zones coincide with the areas, except that Central River combines Kuntaur and Janjanbureh. The registry holds 1.23 footprints per census structure but 2.33 codes per census compound. The ratio is lowest in Kanifing (1.62) and Banjul (1.95), and highest in Upper River (3.10). In Banjul and Kanifing the registry holds fewer codes than households, because several households share one building.

The census compound is a social unit, often a walled family plot with several houses tens of metres apart. The registry’s compound is a physical unit inferred from footprints: grouping rule R2 joins only small structures within 6 m, and rule R0 applies only where a wall is mapped, which is the case for 600 compounds. The registry is therefore finer than the census, and more so in rural zones, where family compounds are large and walls are rarely mapped. A finer unit does no harm to finding a door, because every code still leads to a building of the compound. It does mean that a family compound may hold several codes. Section 13 describes how these are reconciled.

6.8 Footprints and grouping by zone

Figure 8 shows, by zone, the source of the footprints and the rule that formed each compound. Rural concessions (rule R2) make up 15 % of compounds in Central River and 12 % in North Bank, and about 2 % in Banjul and Kanifing, where urban fabric prevents merging by shape. Machine-learned footprints form the majority everywhere except Banjul and Kanifing.

7.1 Forms

Only the adopted form is short, says where the home is, and is unique in the whole country without a qualifier.

7.2 Check digit

A check letter computed by ISO/IEC 7064 MOD 11-2 [23] needs an eleventh symbol, and spoken English letter names rhyme (B, C, D, E, G, P, T, V, Z). Verhoeff’s dihedral-group scheme [24] detects the same error classes as Damm’s but needs a permutation table in addition to the multiplication table. Damm’s algorithm uses one table, needs only the ten digits, and detects every single wrong digit and every adjacent transposition [21]. We also considered showing the check digit to people and rejected it: it lengthens the spoken code by a symbol that carries no meaning, and Section 6.5 shows that a displayed place and zone catch most slips before a route is taken.

7.3 Ordering

Section 6.4 compares three orderings on the full registry. Row order was rejected because consecutive numbers lie a median of 187 m apart, so a number says little about where in a place a compound lies. Morton order is simpler to compute than the Hilbert curve but jumps at quadrant boundaries, which lengthens its 90th-percentile step by 40 %.

7.4 Errors found during development

• Landmarks as places. The first inventory named homes after Dockyard Point and Sukuta Traffic Light (4,780 homes). Landmark words are now excluded (Section 4.4).
• A town named after the village beside it. Under an earlier rule that gave villages within 2.5 km priority over towns, central Farafenni took a neighbouring village’s name. The issue was withdrawn before publication; Eq. (6) now takes the nearest place of any kind within 2.5 km, so the town centre carries Farafenni.
• Colliding aliases. Tukulorr folded to the key of Tukulor. Aliases that collide are now dropped (16 were).
• A transcription error in Damm’s table. One row of $T$ was mistyped during development and was caught by the exhaustive self-test before any code was issued.