1.1 The problem

An address tells a stranger where to go. In Gabon most homes have none that a stranger can use. The Universal Postal Union’s addressing sheet for the country shows a two-digit number before the town, but no list of postcodes has ever been published and none is used in practice [1]. Mail is delivered to post-office boxes (boîtes postales, BP), which identify a counter, not a home [2]. Libreville began naming its streets and numbering its doors between 2012 and 2014 and relaunched the work in 2021, yet in 2024 the press still described “the capital of streets with no name”, where residents give directions such as “behind the prison, to the left of the shop” [3], [4]. Elsewhere in the country, no programme of street addressing could be found.

The consequences are ordinary and costly. A delivery or a taxi ride ends with a telephone call; an ambulance reaches the quarter and then searches for the gate; a bank or an employer asks for an address that cannot be written down; a family abroad sends goods to a description. In the peri-urban quarters and along the national roads and rivers, compounds of several buildings share one gate, and villages extend for kilometres without a named street.

Conventional street addressing, as practised in many African cities since the 1990s, names streets, numbers doorways and maintains an address directory [5]. It is effective where it is completed and used, but it is slow and depends on signs that must be installed and maintained. Its average cost in a World Bank review of 21 cities was US$5.7 per addressed doorway at 2005 prices [5].

1.2 Contribution

This paper presents a national addressing model for Gabon and the registry built from it. Its contributions are:

1. A code that people can say and remember. A code is the name of a place people already use and a number of at most five digits, read in pairs like a telephone number: Louis 12 01.
2. National uniqueness without context. Each name owns one national range of numbers, cut into frozen blocks, one per area that carries the name, so that a code identifies one compound in the whole country with no town, province or device position.
3. Locality and growth. Numbers follow a Hilbert curve over each area, leave gaps for new homes and keep a reserve, so neighbours have neighbouring numbers and new compounds can be numbered between them.
4. Integrity for systems. A machine form carries a Damm check digit over the name, the département, the area and the number, which detects every single-digit error and every transposition of adjacent digits.
5. Complete, open coverage. All 665,089 compounds identified in the country received a code on 21 October 2026. The specification, the reference engine, the registry and a ready-to-deploy server are published under open licences [6].
6. Evaluation and impact. We measure the properties of the issued codes and estimate, under stated assumptions, their effect for citizens, logistics, emergency services and government.

1.3 Structure of the paper

Section 2 reviews existing postal, addressing and land systems in Gabon and international practice. Section 3 describes the data. Section 4 sets out the model, rule by rule, with a worked example on a real code for each formula. Section 5 describes the implementation and Section 6 the evaluation. Section 7 reports the designs we considered and rejected, with measurements. Sections 8 to 11 estimate the benefits for citizens, logistics and commerce, emergency services, and government. Section 12 discusses privacy, safety and governance, Section 13 the limitations, and Section 14 concludes.

2.1 Territory and population

Gabon has nine provinces, numbered GA-1 (Estuaire) to GA-9 (Woleu-Ntem) in ISO 3166-2. The United Kingdom’s Permanent Committee on Geographical Names counts fifty départements at the second level [7]; OpenStreetMap maps 49 of them, the set used in this work. French is the official language, and Fang, Punu, Nzebi, Myene and many other languages are spoken at home. The 2026 general census counted 3,518,621 inhabitants, a total approved by the Constitutional Court on 27 August 2026 and almost double the 1,811,079 of 2013; roughly 60 % of the population lives in the Estuaire, around Libreville, Owendo and Akanda [8], [9]. Port-Gentil, Franceville, Oyem and Moanda are the next largest towns.

2.2 Post and the post-office box

The UPU sheet gives two examples [1]. For home delivery, “05 TCHIBANGA 01” follows the delivery information; for box delivery, “BP 13210” is followed by “01 LIBREVILLE–GROS-BOUQUET 03”. The sheet labels the line “postcode and delivery office”. The leading digits coincide with the ISO numbers of the provinces (Nyanga is GA-5, Estuaire GA-1); the sheet does not state this, and it is our inference. The trailing digits identify the delivery office. No list of these codes has been published and we found no current use of them. A post-office box can serve as a company’s legal domicile, and since 2020 boxes can be requested online [2]. A box locates a counter, not a dwelling.

2.3 Street names and door numbers in Libreville

Libreville’s addressing programme started in 2012–2014. On 8 July 2021 the interim mayor relaunched it with 480 million FCFA from the International Association of Francophone Mayors (AIMF) and the European Union; of 6,000 street plates produced, 3,900 had yet to be installed, and of 60,000 door numbers required, 55,000 had yet to be placed [3], [10]. In May 2024 the programme, which aimed to name 3,650 roads, had not entered daily use [4]. OpenStreetMap holds 3,945 doors with a street and a number in Libreville, Owendo and Akanda, of which 911 are tagged as surveyed. We found no comparable programme in any other town.

2.4 Land, cadastre and the absence of an address register

The Agence nationale de l’urbanisme, des travaux topographiques et du cadastre (ANUTTC) manages land titles, re-marking and boundary corrections; it publishes no open parcel data [11]. The national projection for cadastre and survey is WGS 84 / Gabon TM 2011 [12]. A land title identifies a parcel and a right; it does not tell a courier where to stop, and most homes have none. We found no national address register, no addressing law and no national address geoportal.

2.5 International practice

The UPU’s S42 standard describes the components of postal addresses and a template per country [13], and ISO 19160-1 provides a conceptual model of address data, including aliases and life cycle [14]. Both describe addresses; neither creates them where none exist. The UPU’s 2012 white paper argues that an address is part of national infrastructure and of a person’s social and legal identity [15]. The World Bank’s review of street addressing documents 52 cities in 15 countries between 1989 and 2004: in the 21 cities that completed their operations, 1.2 million buildings were addressed at an average of US$0.6 per inhabitant, and the address directory produced a list of premises more complete than the tax rolls [5]. The same review shows that addressing decays when people do not use it, which motivates our choice to start from names people already use.

2.6 Grid codes and national digital addresses

A second family of systems derives a code from coordinates. Open Location Code (Plus Codes) encodes a point with a 20-symbol alphabet; the compound Louis 12 01 lies in the cell 6FGFCC5J+CH, about 14 m across [16]. In the region, Ghana’s National Digital Property Address System divides the country into 5 m squares and derives an alphanumeric code such as GA-543-0125 [17]. Grid codes cover every point at once and need no survey, but they address cells rather than homes, mix letters and digits that must be spelled, and carry no place name a listener can recognise. Section 7 compares these forms with ours. We retain a grid code for points that are not compounds (a stall, a field, a landing stage) but do not use it as the address of a home.

2.7 The rule adopted: official addresses first

Where the State or a commune gives an address, the model shows that address first, exactly as written, and the code beside it. In Libreville the named street within 30 m of a compound is shown first on 37,622 compounds; a door number from the commune’s plan replaces it once known (from the commune’s data, from surveyed OpenStreetMap doors, or confirmed by residents). Each département keeps a reserved field for a postcode, to be filled when one is published. No code changes when an official address appears.

4.1 Requirements

The model was designed against eleven requirements, in order of priority: (i) a code is easy to say, hear and remember, and obtaining it takes one action; (ii) it is unique nationally on its own; (iii) it identifies a compound and leads to its gate; (iv) official addresses come first where they exist; (v) it uses the language people use; (vi) it cannot be mistaken for a post-office box, a telephone number or a land title; (vii) it can be resolved offline; (viii) it is stable and never reused; (ix) it describes places, never people; (x) it is open and can be transferred to the State; and (xi) it is specific to Gabon and shares no code with any other country’s scheme.

4.2 The forms of a code

A code has several forms, listed in Table 2. People use the remembered form; systems use the machine form.

Let $P$ be an area’s plate name and $N$ its number of three, four or five digits. The remembered form is $P$ followed by $N$ grouped from the left in pairs (four digits: $AB\,CD$; five: $AB\,CDE$; three: $ABC$). The machine form is

where $Z$ is the département’s three-character key (Section 4.3), $I$ the three-digit index of the area in its département, and $c$ the check digit of Section 4.10. A plate name never contains a digit, so the number always begins at the first digit of the text.

4.3 Départements and code areas

Each département has a key of three characters, its province’s ISO number followed by two letters (1LV Libreville, 8PG Bendje with Port-Gentil, 9OY Woleu with Oyem), so that a postal clerk reading 1LV sees the Estuaire as on the 2002 UPU sheet. Each also has a frozen index $z \in \{0, \dots, 48\}$.

Each footprint $f$ belongs to the nearest place point of its own département. With $(\lambda, \varphi)$ for longitude and latitude in degrees and $\mathcal{A}_{z(f)}$ the areas of the footprint’s département,

ties going to the lower area key. Of the 4,922 areas, 2,932 contain at least one compound. The largest, Le Transfo in Port-Gentil, holds 13,031 compounds; Libreville has 114 areas and 191,073 compounds, and Port-Gentil 104 areas and 65,469 compounds. Figure 3 shows the distribution of area sizes. Each area is frozen at first issue: its key, its block of numbers and its frame, the bounding box of its footprints’ centres widened by 0.0005° on each side, over which its numbers are laid out.

4.4 Plate names

The name in a code is a plate name, short enough to paint on a gate and to say in one breath. With $\mathrm{NFC}(s)$ the Unicode-normalised name, $W(s)$ its words and $L$ the French linking words de, du, des, d, la, le, les, l, et, sur, en, au, aux, à, a plate name must satisfy

A name that fails Eq. (3) goes through fixed steps, the first that succeeds being kept: cleaning (brackets; the generic words Quartier, Carrefour, Cité, Campement, Village, so that “Cité Damas” becomes Damas); abbreviating “Saint” to “St”; and taking the most distinctive run of whole words, only when it is clearly ahead. A word is never cut. A list of ordinary French words (mission, catholique, carrefour, institut, sports, cimetière, poteau, …) counts as weak, so a short form keeps the proper name (“Mongo Milando” becomes Milando) and never the landmark word alone. Where every short form would contain a digit, consist only of ordinary words or equal another city’s name, the area carries its département’s chief town as a provisional name until the commune chooses (“Institut des Sports” is provisionally Libreville).

Three further rules fix the spelling. Spelling variants of one name (accents, case, hyphens) take one plate: the one that fits, then the one with its accents (“Akébé-Ville”, not “Akebe Ville”), then the one more people live under. Capitals in the sources are written as names (MIMBOTI becomes Mimboti) except the acronyms of companies and public bodies (SETRAG, SOMIMA). A documented spelling wins over a map’s: Mont-Bouët, as the city and the press write it, rather than “Mont Bouet”.

The 252 provisional names (Table 3) cover 87,684 compounds, which are numbered under a name their commune may change once. The official name always remains an alias (261 aliases are kept; 5 are dropped because two plate names claim them), so typing the full name finds the same compound.

4.5 The name key

Two spellings denote the same name when they have the same key:

Folding can only merge names, never split one (Table 4). The 4,922 areas carry 3,737 distinct plate names; 713 names are shared, by 1,898 areas. Mandji is carried by 18 areas, Aléné by 11, the provisional name Libreville by 10, and Bissobinam, Mbomo and Moukouagna by 9 each.

4.6 National name ranges

Every name owns one national range: 100–999, 1000–9999 or 10000–99999. The capacity rule is that an area’s block holds at least twice its compounds. For an area with $c$ compounds,

For a name whose areas have $c_1, \dots, c_k$ compounds, taken in the frozen order of their keys (département, then area index), the total and the number of digits are

so the limits are 810, 8,100 and 81,000 and the top tenth of every range is always spare. The blocks are laid end to end from $10^{d-1}$. Areas with no compound receive no block until their first compound.

Worked example: Louis. A single area of Libreville carries the name, with $c = 848$ compounds. Eq. (5) gives $B = 10 \lceil 1696/10 \rceil = 1700$, so the block is 1000–2699. Since $810 < T = 1700 \le 8100$, Eq. (6) gives $d = 4$, and 2700–9999 remain spare.

Worked example: Le Transfo, the largest area, $c = 13{,}031$: $B = 26{,}070$, block 10000–36069, $d = 5$, with 36070–99999 spare.

Worked example: Mandji, the most repeated name. Of its 18 areas, 10 contain compounds; their blocks total $T = 3{,}210 \le 8{,}100$, so the name has four digits and its blocks occupy 1000–4209 (Figure 5). Mandji 10 20 lies in Ogooué-et-Lacs and Mandji 39 50 in Offoué-Onoye: the number alone determines the area.

Table 5 gives the result for 2026: 2,505 areas have three digits, 368 four and 59 five. No name needs a qualifier and no area had to be split (Figure 4). Growth. A new area with an existing name, or an area whose block is full, takes the next block above every block of its name; blocks never move, shrink or get renumbered.

4.7 Numbering within an area

Compounds are numbered along a Hilbert curve [24] of order $k = 16$ laid over the area’s frame $(x_0, y_0, x_1, y_1)$, at each compound’s reference point (the centre of its largest building until a gate is confirmed). The point is first mapped to the grid

Its distance $h$ along the curve is then accumulated over $s = 2^{k-1}, 2^{k-2}, \dots, 1$, starting from $h = 0$:

where $\oplus$ is the bitwise exclusive or. The Hilbert curve preserves locality better than other common orders [25], so compounds close on the ground receive close numbers (Figure 6; Section 6.5 measures it).

Let the area’s block be $[F, F + B)$ and let its $n$ compounds be sorted by $(h, \text{stable id})$ and indexed $i = 0, \dots, n - 1$. The numbers are

and the tail reserve $[F + M, F + B)$ is left unused at first issue. Because $n \le B/2$ by Eq. (5), the last number $F + (n-1)\sigma$ always lies in the main part.

Worked example: Louis 12 01. The compound’s reference point is $\varphi = 0.4085002$, $\lambda = 9.4314043$, and the frame of Louis is $(9.429597,\ 0.406494,\ 9.437690,\ 0.414309)$. Eq. (7) gives

$$g_x = \left\lfloor \frac{0.0018073}{0.008093} \cdot 65\,535 \right\rfloor = 14\,635, \qquad g_y = \left\lfloor \frac{0.0020062}{0.007815} \cdot 65\,535 \right\rfloor = 16\,823,$$

and Eq. (8) gives $h = 896\,033\,648$, the 202nd of 848 positions along the curve ($i = 201$). With $B = 1700$, Eq. (9) gives $M = 1133$, $\sigma = \max(1, \lfloor 1133 / 1272 \rfloor) = 1$, and $N = 1000 + 201 = 1201$; the tail reserve is 2133–2699.

4.8 New compounds

A compound built after first issue is placed on the curve by Eq. (8). Let $a$ be the number of its predecessor on the curve (every issued compound counts, retired ones included; $a = F - 1$ if none) and $b$ that of its successor ($b = F + M$ if none), and let $\Phi$ be the free numbers $x$ with $a < x < b$ and $F \le x < F + M$. With $m = \lfloor (a + b)/2 \rfloor$, the new number is

that is $m$, then $m - 1$, $m + 1$, $m - 2$, … If $\Phi$ is empty, the lowest free number of the tail reserve is used, then the lowest free number of the area’s later blocks, and finally a new block of size $B(n + 1)$ from the name’s spare, above every existing block. A number is never given twice: retired numbers remain used. In Louis the step is 1, so a house built between two neighbours receives Louis 21 33, the first number of the tail reserve.

4.9 Decoding without context

A text is read as a name and a number of three to five digits, with or without spaces. With $K(a)$ the keys of an area’s plate name and aliases, $d(a)$ the digit count of its name and $\mathcal{B}(a)$ its blocks, the decoded set is

which by construction holds at most one area. A number in no block, in a name’s spare or with the wrong number of digits is reported as not issued, never guessed. Decoding uses no context: neither the user’s position, nor the device’s region, nor a town written after a comma.

4.10 The check digit

The check digit is computed with Damm’s algorithm [26], [27] over the concatenation

where $i_5$ is the name’s national index on five digits, $z_3$ the département’s index, $I_3$ the area index and $Q$ the table of a totally anti-symmetric quasigroup of order 10 (Table 6).

Worked example. Louis has national index 1801, Libreville’s index is 003 and the area’s index 070, so $\delta = 01801\,003\,070\,1201$. The successive values of $c_j$ are 0, 3, 2, 4, 1, 7, 8, 8, 9, 7, 8, 4, 2, 4, 1, so $c = 1$ and the machine form is GA-1LV-070-1201-1.

Every single-digit error and every transposition of two adjacent different digits changes the check [27], a property first obtained for decimal codes by Verhoeff with the dihedral group [28]. Because the name’s index, the département and the area enter $\delta$, a number moved to another area of the same name fails both the check and the block test of Eq. (11). The check appears only in the machine form; it is never shown to people or asked of them.

4.11 Spot codes

Any point has a spot code on a 10 m grid over its département’s frozen frame, whose origin $(\lambda_0, \varphi_0)$ is the south-west corner of the département’s bounding box floored to 0.01°, and whose reference latitude $\varphi_c$ is the box’s centre rounded to 0.01°. With $k_x$ and $k_y$ the metres per degree,

and the code is $\texttt{GA-}Z\ \texttt{E}X\ \texttt{N}Y$, with four digits per axis, or five where the département spans more than 99.99 km. Decoding returns the cell centre $\hat\lambda = \lambda_0 + 10 (X + \tfrac12)/k_x$, $\hat\varphi = \varphi_0 + 10 (Y + \tfrac12)/k_y$, so the error is bounded by half the cell’s diagonal:

Worked example. For Louis 12 01 in Libreville’s frame ($\lambda_0 = 9.36$, $\varphi_0 = 0.34$, $\varphi_c = 0.45$): $k_x = 111\,317$, $X = \lfloor 0.0714043 \cdot 111\,317 / 10 \rfloor = \lfloor 794.85 \rfloor = 794$ and $Y = \lfloor 0.0685002 \cdot 110\,574 / 10 \rfloor = \lfloor 757.43 \rfloor = 757$, giving GA-1LV E0794 N0757. The decoded centre lies 3.5 m from the reference point (Figure 7).

Points in no outline. The département outlines are generalised, so 80 issued compounds lie in no outline: in slivers between départements, on the shoreline and on islands, the farthest 18.4 km from an outline. Such a point takes its département from those whose outline lies within 20 km: among them, the one with the nearest outline whose frame holds the point, distances being measured to the nearest edge with longitude differences scaled by $\cos\varphi$ and ties going to the lower key; if no such frame holds it, the nearest cell of a frame at most 1,000 m away. Frames are never moved or widened, because either would change spot codes already published, and the decoder is unchanged, since a spot code names its département. Every issued compound therefore has a spot code. For example, a compound of Moka at $\varphi = 0.8104524$, $\lambda = 9.3792151$, on an island 18.4 km from the nearest outline, that of Komo-Mondah, falls in Komo-Mondah’s frame and receives GA-1KM E0881 N6971.

4.12 From footprints to compounds

A code belongs to a compound, not to a building. Footprints are grouped one area at a time, in a fixed order by stable id, before any number is given. With $\mathrm{dist}(a, b)$ the shortest distance between two outlines (zero when they touch or overlap) and $A(f)$ the area of footprint $f$, an annex $f$ with $A(f) < 25$ m² joins

The full rules are in Table 7. Only annexes join a house, and a hut cluster never joins two houses, so groups cannot chain.

The 783,855 footprints form 665,089 compounds, 1.18 footprints per compound: 566,928 single buildings, 92,635 houses with their annexes, 5,288 hut clusters and 238 mapped enclosures. By size, 566,985 compounds are one structure, 82,186 two, 12,644 three and 3,274 four or more. The median footprint is 70 m². Table 8 gives the figures by province.

The rules are deliberately conservative. Merging two codes later costs nothing, whereas splitting a wrong merge would give a household a new code. Residents’ answers in the application (“these buildings are one home”, “this is the gate”) complete the grouping where walls are not mapped.

4.13 Building profiles

Every structure carries a use class with a confidence and a source, its size, its number of floors where known, and a status. Classes are inferred in a fixed order: the OpenStreetMap building type (confidence 0.85), a shop, amenity or office tag on the footprint (0.75), an open place inside it (0.65), the footprint’s size (0.3–0.4), and otherwise residential by default (0.45). Table 9 shows the result.

OpenStreetMap records floors for 491 buildings. A home never carries a name; business and public buildings carry their public names only.

5.1 Reference engine

The specification (engine/SPEC.md) states every rule normatively. The reference engine (engine/reference/ga_engine.py) implements it in the Python standard library: name keys (Eq. 4), plate-name checks (Eq. 3), block sizes (Eqs. 5–6), the Hilbert index (Eqs. 7–8), numbering (Eqs. 9–10), decoding (Eq. 11), the check digit (Eq. 12), spot codes (Eqs. 13–15) and footprint grouping (Eq. 16). A second module derives plate names. The engine ships with the frozen register of areas, names, blocks and frames (engine/data/), about 3 MB, so that a device can resolve every code offline.

5.2 Registry

The registry (registry/) publishes every issued code: one compressed CSV file per département with one row per compound (machine, remembered and written forms, area, number, check, status, issue date, entrance point, spot code, official address, use class, structures and floor area), the frozen area table, and a manifest giving each file’s row count and SHA-256 digest. Every compound of every département is published; the rows contain nothing about the people who live there.

5.3 Server and API

A ready-made server (server/) lets a ministry, a postal operator or a developer serve the registry in one command, with Docker or plain Python 3.9 and no third-party packages. It imports the registry into SQLite with indexes on every form and an R*-tree spatial index [29], [30], verifying each file’s digest and the totals against the manifest and refusing any mismatch. It serves a JSON API under /v1: lookup of a code in any form, the code at a point, codes near a point, validation, spot codes, départements, areas and paged lists, all described in OpenAPI 3.1 [31] with recorded examples. It can export the registry back in the published format; after an import, the decompressed export is byte-identical to the published files.

The server can also issue codes for new compounds and new entrances by the rules of Sections 4.8 and 4.12, but only where it has been designated as the country’s issuer. Issuing must be done by one issuer per country: two issuers would see the same free number and give it to two homes. A deployment that has not been designated by the State or by N-WAYS LIMITED therefore runs read-only.

5.4 Conformance and parity

An implementation conforms when it reproduces the conformance vectors published with the engine (vectors.json): 724 name keys, 500 code forms, 2,394 decodings with and without check digits, 588 spot codes, 310 points in no département outline (the 80 issued compounds among them, points beyond a frame and refusals), 5 Hilbert indices, 6 grouping cases and 4 numbering cases on a real area with new homes, 4,531 in all. The engine’s 15 tests and the server’s 31 tests pass. The server’s issuing was checked against the issuer used in production on 62 issuance cases, 5 insertion cases and 300 random points, with identical results.

6.1 Uniqueness and ambiguity

All 665,089 machine forms are distinct, and so are the remembered forms, even after reducing each name to its key (Eq. 4). Of the 3,737 names, 373 are carried by more than one area that contains compounds; together these hold 259,144 compounds, 39 % of the total. Had each area been numbered from the start of its own range, as in the first design we considered (Section 7.2), 150,090 codes (22.6 %) would share their name and number with a code in another area, in 64,701 colliding pairs or groups. National ranges remove every such collision without adding a digit to any code that fits in five.

6.2 Length

The mean number length is 4.12 digits, and no number exceeds five (Table 5). Plate names of the 2,932 areas in use have 7.5 characters on average, and 7.8 when weighted by the codes that carry them. Seventy-five plate names in use exceed 12 characters; all are provisional names awaiting a commune’s choice, and they cover 17,122 compounds.

6.3 Decoding

We decoded the written forms of 50,000 codes drawn at random from the registry, and 20,000 of them typed in lower case with every accent removed, as a person might type on a telephone. All decoded to their own machine form, with no false match.

6.4 Error detection

For 2,000 random codes we formed every single-digit substitution and every adjacent transposition in the visible part of the machine form (area index, number and check). All 145,818 substitutions and 12,730 transpositions were detected, as the theory guarantees [27]. Errors outside that guarantee were also mostly caught: 89.7 % of 10,938 jump transpositions ($abc \to cba$) and 89.6 % of 13,248 twin errors ($aa \to bb$).

6.5 Locality of numbering

We measured the distance on the ground between compounds with consecutive numbers in every area of at least 50 compounds in six départements (Libreville, Bendje, Woleu, Mpassa, Ogooué-et-Lacs and Douya-Onoye: 350,235 pairs). Table 11 compares the Hilbert order with a Morton (Z-order) curve and a raster of horizontal bands over the same grid.

The Hilbert order shortens the mean step by 24 % against Morton and 37 % against a raster, and halves the share of long jumps against Morton, consistent with its known clustering properties [25].

6.6 Spot codes

Every one of the 665,089 compounds has a spot code. The decoded cell centre lies on average 3.82 m and at most 7.07 m from the reference point, the bound of Eq. (15); for the 80 compounds in no département outline, which take the département of the nearest outline (Section 4.11), it lies at most 6.65 m away. Re-encoding 50,000 points reproduced their published spot codes exactly.

6.7 Compound grouping

We re-ran the grouping rules with the 179 walls and fences mapped around buildings removed from the input. 4.5 % of these enclosures came out as exactly one compound, and a further 25.1 % missed only one building. Behind a wall the main house, the rooms and the kitchen are often separate buildings; without a mapped wall the rules keep them apart by design, and residents’ answers merge them.

6.8 Real codes

Table 12 lists the compound nearest well-known places, as issued.

Two cases illustrate the rules. A partner’s system that receives GA-1LV-070-1210-1, with two digits swapped, computes the check of 1210 as 4, not 1, and refuses the code instead of sending a courier to the wrong house. The texts mandji 1020 and mandji 3950 resolve to two different départements’ blocks of the single Mandji range, with nothing to choose from a list.

6.9 The server

On the full registry, import takes about 8 seconds and produces a 310 MB database; once decompressed, an export is byte-identical to the published files (the gzip containers themselves may differ in compression). A load test of 6,400 requests on 16 threads produced no error; typical latencies were about 1 ms for a lookup by code, 5 ms for the code at a point, and 100 ms for all codes within 200 m in central Libreville. For 150 random codes, all 24 fields returned by the server equal those of the operating N-Ways service. For the code at a point, the server agreed with the operating service on 153 of 200 random points: 36 differences arise because the published registry contains entrance points but not building outlines, so the server finds a compound only within 12 m of its entrance, while the operating service also matches anywhere inside the footprint; the remaining 11 are points north of 1.65° N at which the operating service returned no answer at the time of measurement because its country look-up did not resolve them; the published server, which serves Gabon alone, has no such step.

7.1 Code forms

The capacity rule is at least two numbers per compound. Table 13 compares the forms we considered for 665,089 compounds.

The limit of five digits follows the evidence on short-term memory. Miller’s classic estimate of about seven items [32] has since been lowered to about four chunks for material that cannot be rehearsed [33]. A number of four or five digits read in two groups, after a name that is already known, fits that limit; a sixth digit adds a third group.

7.2 Per-area numbering

The first design we considered numbered each area independently, with four digits and a check digit on the plate (for example “LOUIS 2471-8”). Section 6.1 shows that it would have left 150,090 codes ambiguous, so that a town would have had to follow the code wherever names repeat. Giving each name one national range removed the ambiguity at the cost of one digit for some areas: 30.9 % of compounds have five digits, mostly in towns that the map knows by a single point (Port-Gentil, Mouila, Koulamoutou), where one area covers the whole town until its quarters are named.

7.3 A check digit on the plate

A check character that detects every single error and every adjacent transposition requires either an eleventh symbol with a prime modulus or a quasigroup such as Damm’s. We chose Damm’s algorithm because it needs digits only, which read the same in every language. We removed the check from every form people see: a digit that carries no meaning for the speaker is the first to be dropped or garbled, and people’s errors are caught instead by the block test of Eq. (11), which refuses most numbers that were never issued. Systems, which copy codes rather than remember them, verify the check.

7.4 Curve

We compared the Hilbert order with a Morton curve and a banded raster (Table 11) and chose the Hilbert curve for its shorter steps and fewer long jumps.

7.5 Spelling variants

Grouping spelling variants with the name key of Eq. (4) would write a doubled letter once and fold “gui” into “gi”. Gabon’s place names were therefore grouped by accents, case and punctuation only before any number was given (“Mabangui”, not “Mabang I”; “Akomé II”, not “Akome I”); the key is used for decoding, where merging can only help.