Architectural Scale, Visual Communication, and Wayfinding Experience in Turkish City Hospitals

Abstract

Announced bed capacity is often read as if it recorded how a city hospital is used on foot. The measured variables are that capacity and two author-adjudicated language shares in unsolicited public texts. This study tests whether announced beds rank with the hospital-level share of retrieved texts describing navigation-related difficulties or indoor walking and mobility burdens. It does not estimate a causal effect of hospital size on wayfinding performance, and it does not audit physical signage or trace routes. Across the 26 hospitals named on the ministry directory retrieved in July 2026, the complete Yandex corpus comprised 566 texts; 670 Şikayetvar complaints were kept separate and were not pooled. In that complete corpus, author-adjudicated labels marked 35 navigation-related texts and 27 walking-or-mobility texts. Primary tests used 15 hospitals with at least 15 eligible Yandex texts (486 texts; 32 navigation-related and 26 walking-or-mobility). Neither share was statistically supported after correction for the two primary tests (navigation-related: Spearman ρ = 0.512; permutation p = 0.0534; Holm-adjusted p = 0.1040; walking-or-mobility: ρ = 0.513; p = 0.0520; Holm-adjusted p = 0.1040). Removing Bilkent and Etlik weakened the walking association, through both leverage and reduced power at k = 13. OpenStreetMap building footprint ranked with announced beds (ρ = 0.876 on 14 hospitals) and, once beds were partialled out, was not independently associated with either share. The footprint is the horizontal union of mapped building polygons. It is not an estimate of foot traffic inside a multi-storey building. The longest mapped block axis runs from 205 m to 656 m, so the campuses are horizontally extended rather than towers. Visual wayfinding communication is an interpretive layer linking spatial organisation to reported navigation experience; the interior-architecture contribution is this directory-wide language test and a codebook for later audits, not a measured evaluation of spatial configurations or signage systems. The coefficients are not route lengths, signage scores or patient-level prevalences.

1. Introduction

Turkish city hospitals concentrate outpatient clinics, diagnostics and inpatient care in large, often multi-block complexes. On the mapped plans, the longest block axis runs from 205 m to 656 m, so the campuses are horizontally extended rather than towers (Section 2.2.2). That concentration can raise service capacity at the urban scale. It also lengthens indoor journeys and multiplies the points at which a visitor must choose a corridor, a floor or a block. Announced bed capacity records how large the complex is said to be. It does not record whether a writer reports getting lost or walking a long way. This paper tests whether announced bed capacity ranks with navigation-related language and with walking-or-mobility language in public texts about the 26 hospitals named on the ministry directory retrieved in July 2026. The test is a hospital-level rank association. It is not a causal estimate of size on wayfinding performance, and it is not an audit of signs or indoor routes.

A large building can still be intelligible when a visitor can follow the circulation. Architectural legibility is the environmental support for that task. Weisman, O’Neill and later interior-wayfinding reviews separate the plan, the signs and what the person already knows [1,2,3,4]. Section 2.3 sets out that evidence, including percentages from university buildings that do not transfer directly to a hospital.

Hospital studies have already joined layout to observed routes and to space syntax, including the ground floors of Konya and Etlik city hospitals. Section 2.4 reviews those single-building results [5,6,7,8,9,10,11,12,13]. The absence of hospital wayfinding research is therefore not the gap. The comparison specified here is narrower: whether announced bed capacity, across this directory, ranks with navigation-related language and with walking-or-mobility language in retrieved public texts.

Online reviews can carry that second term, and they can also mislead it. Free-text hospital comments have been classified at institution level [14]. A sentiment score is not the outcome required here. A visitor can praise the building and still report being unable to find cardiology, and a sentence that the hospital is large is not, by itself, a wayfinding failure. The text measure in this paper is a spatial-experience classification, not a star rating and not a general polarity.

Four layers that are frequently collapsed in the evaluation of large hospitals are kept apart, and they are ordered. Architectural scale leads to spatial organisation and to visual wayfinding communication; those two layers shape what a visitor experiences; unsolicited language is the trace of that experience that can be retrieved. Spatial organisation, functional medical adjacency and visual wayfinding communication were not measured here. Announced capacity, the scale used in the tests, is a bed count with a written source. Spatial organisation is the physical arrangement of departments: blocks, floors, routes and choices. A clinic name in visitor language usually names a departmental outpatient unit at that address. It is not a hierarchy separate from the departments. Visual wayfinding communication is the information layer that tells a visitor what that arrangement is: signs, names, arrows, maps, symbols, type and colour. The user outcome splits again, because getting lost is not the same event as reading a poor sign, asking for help or walking a long way. The estimated tests use the navigation-related share defined in Section 4.3. That share pools being lost, a block or floor problem, a sign complaint and asking for help. A sign complaint is a visual-communication complaint. The navigation-related union is not a visual-communication score. Announced capacity and the language shares were measured first. Physical scale was then taken from OpenStreetMap building footprints, as an extension added after the text analysis. Indoor routes, departmental adjacencies and the visual wayfinding system itself were not measured. Figure 1 draws measured associations with solid arrows and unmeasured mechanisms with dashed arrows. Figure 2 records the sampling flow from the 26-hospital directory to the 15-hospital primary tests.

Figure 1. Measured variables and unmeasured mechanisms. The measured layer (solid arrows) is announced bed capacity, the collinear OpenStreetMap building-footprint check, and two author-adjudicated language shares (navigation-related; walking-or-mobility). Those shares are not lexicon scores. Solid arrows are hospital-level rank associations; they are not causal paths. Dashed arrows are untested mechanisms: functional medical adjacency, spatial organisation (departments placed in blocks and on floors, routes and decision points), and the visual wayfinding system (hierarchy, naming, type, pictograms, maps, colour). A sign complaint in a retrieved text is not a signage audit.

Figure 2. Sampling and coding flow. The directory frame is 26 hospitals. Eligible Yandex texts number 566, of which 35 were navigation-related and 27 walking-or-mobility in the full corpus. Primary rank tests use 15 hospitals and 486 texts (32 and 26 marked). Şikayetvar (670 texts) is a contrast corpus and is not pooled. Holm adjustment of the two primary tests leaves both at p = 0.1040; those tests are not statistically supported. The unit of the rank tests is the hospital.

The outcome is a hospital share: marked texts divided by retrieved texts of at least 40 characters. Announced bed capacity is the ministry profile value used in the original analysis. The Yandex and Şikayetvar texts were retrieved in July 2026; they were not harvested again on 22 September 2026. The file Supplementary S24 records 22 September 2026 as the date those ministry profile URLs were entered in the identity table, not as a second collection of reviews. A language share is that marked count divided by that hospital’s eligible Yandex texts. The unit of the rank tests is the hospital, not the review. Repeating one hospital’s share across its review rows would not create additional hospitals. Figure 1 is that separation: the tested links are announced capacity, footprint and the two language shares; indoor routes and visual wayfinding communication stay outside the test. Section 4 states the codebook, the 15-text rule, the permutation test and the later OpenStreetMap layer. Section 5.2 reports the hospital comparison.

2. Background: Scale, Legibility and What Users Write

The comparison sits at the junction of six lines of work: the spatial consequences of large hospitals, architectural legibility, healthcare wayfinding, walking effort, information support and unsolicited accounts of patient experience. They are reviewed in the order that produces the model. Capacity is not physical size, and physical size is not the number of choices on a route. Distance and decision-making are different burdens. Signs can change how a configuration is used. They do not replace the configuration. What visitors write afterwards is a record of experience, and the sampling frame of that record has to be stated.

2.1. Development and Organisational Context of City Hospitals in Türkiye

City hospitals in Türkiye are a national replacement of the hospital stock, financed in part through public–private partnership under Law No. 6428 [15,16]. A July 2026 statement described 27 city hospitals in service and about 40,000 beds [17,18,19]. The sampling frame is not that operating count. It is the 26 hospitals named on the directory retrieved in July 2026 [20]. Samsun and Ordu were reported in service in July 2026 and were not on that page; neither was added after the harvest. Leaving the newest campuses out could bias the comparison if they differ in layout or in how visitors write. They had no directory profile and no review corpus in this harvest. Photographs, as-built plans, staff counts and healing-space inventories are not part of this file. None of these statements describes the walk from an entrance to a clinic; Figure 1 keeps that route outside the estimated test.

2.2. Hospital Scale and the Spatial Consequences of Large Complexes

Announced beds, floor area, block count and the public route from entrance to clinic are different descriptions of a large hospital. Two hospitals with the same bed count can still differ in footprint and in circulation. Capacity scale, physical scale and circulation scale are therefore kept apart. A bed count can enter the model as capacity. It cannot stand for the route.

2.2.1. Functional Medical Adjacency and Accessibility Constraints

The clinical hierarchy of a hospital is its departments. Blocks, floors and clinics are not a substitute for that hierarchy. A department is reached at a physical address, commonly a block and a floor on these campuses, and a clinic name in visitor language usually names the departmental outpatient unit at that address. The block-or-floor code records a writer who cannot identify that address. It does not reorganise the hospital by blocks. Medical programme and building technology are what keep the complex from being an ordinary large building. Imaging rooms, operating and intensive-care adjacencies, controlled emergency access, service cores and the public information system fix where a department sits and how a visitor is told to move. Those features are constitutive of the hospital. They are not measured predictors. A similar number of beds does not imply an equivalent clinical programme or circulation network. A hospital is arranged around the procedures it is built to deliver. Outpatient clinics, imaging, laboratories and pharmacies form public-access chains that a visitor must follow; emergency, operating and intensive-care units follow a different adjacency and isolation logic, with controlled access that is not the same as the public route to a clinic. Departmental proximity requirements, diagnostic pathways and the distribution of outpatient units can alter both the physical effort and the navigational complexity of hospital visits. Two hospitals with the same announced bed count can therefore differ in block layout, in public choice points and in walking load even when footprint is similar. Inter-department connection and isolation are circulation-scale quantities. Hospital-specific department networks were not available and were therefore not included as measured predictors. No medical-profile variable, no adjacency matrix and no isolation index enters the rank tests. These characteristics remain plausible explanatory factors rather than observed predictors, and no causal pathway from bed capacity to navigation experience can be established [21,22,23].

Large medical centres have been associated with stress, wayfinding difficulty and spatial disorientation, and circulation has received less research attention than other parts of the building [24]. A review of 102 layout studies grouped the challenges as overcrowding, waiting, visibility, wayfinding and walkability, among others [21]. The same programme later specified a configuration model and a design-support architecture in which patient walking distance and wayfinding difficulty are separate criteria [22,23]. That separation is the one this paper uses between access burden and spatial complexity. Bed counts describe the complex. They do not describe the journey to a clinic.

2.2.2. Plan Extent of the Campuses

Among the 14 primary hospitals with a mapped footprint, the layout is horizontally extended rather than a vertical tower. The longest mapped block axis runs from 205 m at Manisa to 656 m at Etlik, and mapped block counts run from 1 at Kocaeli to 18 at Kartal. Thirteen of those fourteen campuses have more than one mapped block. A vertical point-tower would be read from a short footprint together with a high storey count, and that combination is not present in these plans. Figure 3 draws three of them at one scale: Etlik, a spread campus of 15 mapped blocks; Bilkent, four mapped blocks on a long footprint; and Göztepe, a smaller campus of six mapped blocks. The drawing is the OpenStreetMap ground projection retrieved on 24 September 2026. It records that horizontal extent. Each panel is titled Graphic scale and Solar orientation. Graphic scale is the 200 m bar. Solar orientation is the north arrow, which is grid north of the projected plan. It is not an as-built floor plan.

Figure 3. OpenStreetMap plan outlines of three primary-set campuses, drawn to the same scale from the 24 September 2026 extract. The plans are horizontally extended rather than towers: the longest block axes are 656 m, 597 m and 277 m. Grey polygons are mapped building footprints. The black line is the campus outline. (a) Etlik: 4050 announced beds, 15 mapped blocks, longest block axis 656 m, footprint 211,732 m². (b) Bilkent: 4050 beds, four mapped blocks, longest block axis 597 m, footprint 147,739 m²; area-weighted mean 6.95 storeys, the only primary campus whose storey tags cover at least 80% of footprint area. (c) Göztepe: 815 beds, six mapped blocks, longest block axis 277 m, footprint 19,460 m². Each panel is titled in English. Graphic scale marks the 200 m bar. Solar orientation marks the north arrow. The three panels use one scale, so the bars are equal. The arrow is grid north of the projected plan. It is not a sun-path or shadow study. About seven storeys on a plan of this length is a low- to mid-rise campus. These are ground projections. They are not architectural floor plans, sections, three-dimensional models, interior department maps or pedestrian routes.

Where height can be read, it agrees with the plan. Gross floor area is stored only where building: levels tags cover at least 80% of footprint area (Supplementary S23). In the primary set, that condition holds only at Bilkent (tagged share 0.997; estimated gross floor area 1,027,196 m²; area-weighted mean 6.95 storeys on a footprint of 147,739 m²). Outside the primary set, it also holds at Balıkesir (mean 7.00) and Kütahya (mean 6.00). A mean near seven storeys on a footprint hundreds of metres long is a low- to mid-rise campus, not a tower. The untagged campuses are not given a separate height rank. No section or three-dimensional model was drawn, because a storey tag is not a measured section.

2.3. Spatial Legibility, Complexity, and Wayfinding

Wayfinding is the sequence by which a person orients, chooses a route, acts at decision points and reaches a destination. Architectural legibility is the environmental support for that sequence [1,2]. In ten university buildings, judged plan simplicity accounted for 56% of the variance in self-reported disorientation, and familiarity accounted for 9% [1]. The setting was not a hospital, and the outcome was self-report. The plan and the person’s knowledge are separate terms. O’Neill separates configuration, its cognitive representation and observed wayfinding, and treats floor-plan configuration and signage as factors that can be varied independently [2,3]. Signs do not cancel a difficult plan. Whether any campus in this directory repeats similar towers and undifferentiated corridors is not known here, and no entrance-to-clinic length was measured.

An integrative review places floor-plan configuration, landmarks, signs and maps among the environmental factors, and more complex plans are associated with longer travel, wrong turns and backtracking [4]. In hospitals, route choice follows the configuration people can see, not only the shortest line on a drawing [25], and topological knowledge of a layout develops with use, so a first visit and a familiar visit are not the same task [26]. Space syntax describes the building as a network of permeability [27]. A PRISMA review of 38 studies found that the field relates spatial properties to experience by description, correlation and regression [28]. Integration, choice and depth belong to that family. They are not a counted decision point on a stated route.

2.4. Healthcare Wayfinding and Spatial Configuration

Hospital wayfinding studies usually stay inside one building. A video simulation tested drivers’ turn decisions at a parking entrance rather than pedestrians inside a clinic corridor [29]. Observed routes have been set beside visibility graphs [5]. Circulation complexity in the Milad 1000-bed hospital was related to wayfinding behaviour [9]. Space syntax has been used to plan a wayfinding system in Taiwan [10], and a 2026 study of one medical centre pairs spatial complexity with the distribution of signs [11]. A bibliometric review covers the literature from 1974 to 2020 [12]. Plan, virtual reality and observed behaviour have been compared in healthcare buildings [13]. In Türkiye, the ground floor of Konya City Hospital has been analysed with space syntax, and a doctoral thesis examined wayfinding in that circulation [6,7]. The ground floor of Etlik City Hospital has been analysed with space syntax and visibility graphs [8]. Each study answers a question about one plan. None places the city-hospital directory on one route indicator.

Ge and colleagues combined space-syntax indicators with an on-site task in three tertiary hospitals: 50 participants and 600 observations across four outpatient zones. Hesitation time and hesitation count were higher in visually complex layouts than in regular ones, and age and education changed the size of those associations [30]. That is evidence that outpatient configuration and wayfinding behaviour can be measured together. It is not a census of this directory, and this paper does not repeat the task.

The campus outside the front door is a different segment. At one large hospital, space syntax together with vehicle shadowing and pedestrian mapping in the arrival zone covered 219 vehicles and 2096 pedestrian behaviours. Connectivity and integration were correlated with both volumes [31]. A campus that is hard to reach by bus is not a clinic that is hard to find after the entrance. Campus-transport language is therefore kept separate from indoor wayfinding language.

2.5. Walking Distance, Access Burden, and Patient Effort

A long straight corridor is physically heavy and cognitively simple. A short route with many junctions and a change of block or floor is cognitively heavy and may still be short. Hospital design-support models already list walking distance and wayfinding difficulty as separate outputs [22,23].

2.6. Visual Wayfinding Communication

Wayfinding is produced neither by the plan alone nor by the sign alone. Arthur and Passini treat architecture, graphic information and spoken help as one process [32]. Calori and Vanden-Eynden treat the information system as a question of where a message is offered as well as what it says [33]. As used here, hospital wayfinding is the correspondence between spatial organisation and the visual information that communicates it. Plan and signs can be varied independently [3]. A difficult plan is not repaired by declaring that signs exist.

A review of 34 publications grouped signing recommendations into text, hierarchy, symbols, colour, standardisation, the user, legibility, language and placement, and noted that an expanded hospital often inherits a system that no longer matches the plan [34]. In three Portuguese hospitals, 1287 patients and visitors often rated the signs positively and still ignored them and asked staff [35]. A favourable opinion of the signs is not evidence that the signs carried the route. Symbol comprehension varies with the user and the symbol [36]. Asking is a coping behaviour. It is not proof that a sign failed, and it is not the same event as being lost.

A virtual-hospital experiment compared an existing sign system with a redesigned one in 64 participants [37]. That is a test of an information system, not of bed capacity, and it is not a measurement of a city hospital in Türkiye. The medical-centre study cited above locates 194 signs on axial and visibility maps and separates spatial complexity from informational support [11]. It supplies no sign map for this directory.

The codes only partly respect that distinction. Getting lost, missing the block or the floor, a missing or contradictory sign, and having to ask are entered together in the navigation-related share. The first two are navigation outcomes. The sign mark is closer to a visual-communication cause. Asking is a coping behaviour. Pooling them answers whether capacity ranks with a broad navigation complaint. It does not answer whether the visual system failed. What was measured is that four-part language union, the walking-or-mobility union, announced beds, and, later, mapped footprint. What was not measured is the sign system itself: hierarchy, naming, directional continuity, typography, contrast, pictograms, colour, maps and landmarks. Those elements supply the frame for reading a sign complaint, a missed block or floor, and a request for directions in the retrieved texts. An audit of those elements would be a second study. It is not a result of this one, and the navigation-related union is not that audit. A statement that the signs were sufficient is kept out of the difficulty share. Walking-or-mobility language stays outside it, because a long indoor walk is a physical load, not a failure to read a sign. Section 5.3 now reports the four navigation marks one by one against capacity, so that a sign complaint is not treated as if it were the same event as asking.

2.7. Post-Occupancy Evaluation and User Experience

Ulrich and colleagues placed wayfinding among healthcare-environment topics with a usable evidence base [38]. Post-occupancy evaluation returns to a building after use, as in the arrival study that checked space syntax against observed vehicles and pedestrians [31]. The present study is of that kind only in a limited sense. The buildings are in use, and the measured quantities are announced bed capacity and author-adjudicated marks in retrieved texts. There is no random sample of patients and no observed route beside each sentence.

2.8. Online Reviews as Unsolicited Patient-Experience Data

Greaves and colleagues showed that sentiment scores from online comments can be compared across institutions [14]. A sentiment score is not the outcome required here: a visitor can praise the staff and still report being unable to find a clinic. Online complaints treat the care environment as a category of its own [39], and public reviews overlap with, but are not identical to, a formal patient survey [40]. An online corpus does not replace a survey. The writers are people who posted, the platform’s default order is not a random draw, and a complaint site and a map-review page do not share a sampling frame. The codes are built for missed blocks, missing signs, long indoor walks and difficult campus access. They are not a prevalence of getting lost among all patients.

Existing studies usually take one hospital, or one link in that chain. A comparison of announced bed capacity with unsolicited navigation-related and walking language across the Turkish city-hospital directory has not been the object of a published study. The linguistic work in this paper is not a substitute for a layout solution. A marked sentence is a retrieved trace of a reported experience. It is not a measured route, not a sign audit, and not a prescription for how a functional hospital should be arranged. The contribution to interior architecture is therefore the separation of capacity-scale from reported navigation and walking language, and a codebook that later fieldwork can pair with unmeasured spatial configurations and visual information systems. Figure 1 separates tested associations from the route, adjacency and sign-system quantities that were not measured. Figure 2 records the sample that those associations describe, distinguishing the complete 566-text corpus from the 15-hospital, 486-text primary tests.

3. Conceptual Framework and Hypotheses

3.1. Three Meanings of Scale

Scale is three sets of quantities, not one judgement, and a fourth set is named so that it is not silently absorbed into the first three. Capacity scale covers announced beds, outpatient volume and clinic count. Physical scale covers gross floor area, footprint, number of blocks and number of floors. Circulation scale covers the public network and the distribution of entrance-to-clinic routes. Programme scale covers medical profile, department adjacency, and isolation versus rapid-access links. Two hospitals with the same bed count can still differ on the second, third and fourth sets. In the present comparison, the quantity available with a written source for all 26 hospitals is announced bed capacity, a capacity-scale variable [20]. Physical scale was added later from OpenStreetMap. The declared primary quantity is the union of mapped building footprints inside the campus polygon. Block count, the long axis of the largest block and site area are exploratory. Gross floor area is left blank unless building:levels tags cover at least 80% of footprint area; an empty cell is not entered as zero. Programme scale was not operationalised. Financing, including public–private partnership, is context. It is not a spatial score. A long route with no choice and a short route with many choices are different loads. Those route quantities were not calculated. The supplementary note states unused index definitions. It does not report a metre or a decision point.

3.2. The Rank Association

The outcome for hospital j is a share, not a count and not an ordinary-least-squares score. The navigation-related share is the number of retrieved texts carrying that union, divided by the number of eligible texts. The walking-or-mobility share is defined the same way. With few hospitals, a non-normal share and many zeros, Spearman’s rank correlation is used as a non-parametric measure of monotonic association [41]. It does not model the extra uncertainty of a small denominator. The minimum-text threshold, the threshold repeats and the Wilson intervals display that uncertainty. They do not put the size of the denominator into the correlation.

Share_j = positive texts_j/eligible texts_j.

(1)

Table 1 summarizes the measured constructs and conceptual variables used in this study.

Table 1. Construct–measure mapping. Status distinguishes quantities that enter the rank tests from mechanisms that remain conceptual.

Construct	Definition	Data Source	Unit	Status in This Study
Capacity scale	Announced number of beds	Ministry profile, July 2026	Hospital	Measured; primary explanatory descriptor
Physical scale	Mapped building-footprint union	OpenStreetMap/Overpass	Hospital (m2)	Partly measured; collinear with beds (VIF = 9.85)
Functional organisation	Department hierarchy; clinic–imaging links; emergency isolation; public-access chains	—	—	Not measured; plausible confounder
Spatial organisation	Blocks, floors, routes, decision points	—	—	Not measured
Visual information system	Hierarchy, naming, type, pictograms, maps, colour, placement	—	—	Not audited on site; only signage-failure language is observed
Navigation-related language	Union: lost, block/floor, signage failure, asked	Author-adjudicated Yandex texts	Share of eligible texts at hospital j	Measured; H1 dependent share
Walking/mobility language	Union: long indoor walk, mobility	Author-adjudicated Yandex texts	Share of eligible texts at hospital j	Measured; H2 dependent share
Size mention (exploratory)	Hospital described as large/huge with no navigation or walking mark	Lexicon screen only	Share of eligible texts	Exploratory; not an author-adjudicated outcome

The navigation-related union is a language indicator, not a measure of the sign system, of plan organisation, or of functional medical adjacency. Fourteen signage-failure texts are not a quantitative audit of physical signs.

3.3. Research Questions and Hypotheses, in Operational Terms

RQ1–RQ4 and H1–H2 are marks used in this paper. They are not borrowed from a cited article. Each is defined by a numerator, a denominator, a hospital set and a test. RQ1 is descriptive: for each directory hospital with eligible Yandex texts, the navigation-related share and the walking-or-mobility share are the marked counts divided by the eligible count. RQ2 is H1: among hospitals with at least 15 eligible Yandex texts, Spearman’s rank correlation of announced beds with the navigation-related share, with a two-sided permutation p value. RQ3 is H2: the same test for the walking-or-mobility share. RQ4 is a composition count: how the navigation-related texts split among lost, block or floor, signage failure and asked, without entering those four marks into the primary Holm pair. Holm adjustment is applied only to H1 and H2. Component correlations, size mentions, OpenStreetMap partials and threshold repeats are exploratory. The 15-text rule is an analytic choice: below that count, one marked text moves a share by more than six percentage points.

4. Materials and Methods

This section states the sampling frame, the retrieval rules, the codebook, the rank tests and the later OpenStreetMap layer. Section 4.6 reports the ethics of this public-text retrieval, with reference to the AOIR guidelines [42]. Figure 4 and Table 2 set out the eight-step sequence before the detailed subsections.

Figure 4. Eight-step analysis workflow, placed here so the sequence is visible before Section 4.1, Section 4.2, Section 4.3, Section 4.4, Section 4.5 and Section 4.6. Table 2 is the same sequence as a table. The unit of the rank tests is the hospital. Step 8 states the interpretation limit: the drawings are OpenStreetMap ground plans; indoor routes and signage systems were not measured.

Table 2. Eight-step analysis workflow. The unit of the rank tests is the hospital, not the text.

Step	Operation
1. Sampling frame	Ministry directory of 26 city hospitals, retrieved in July 2026. Not a claim that 26 was the complete national operating count.
2. Retrieval	Yandex visitor texts and Şikayetvar complaints kept as separate sources. Eligible Yandex texts: ≥40 characters. Dropped short strings were not archived.
3. Analysis unit	Codes are assigned to texts. Primary correlations use hospital-level shares: share_j = y_j/n_j.
4. Labelling	Four navigation marks (lost, block/floor, signage failure, asked) and two walking marks (long walking, mobility). Final Yandex labels are author-adjudicated.
5. Primary tests	H1 and H2: Spearman rank correlation of announced beds with the two union shares among hospitals with n ≥ 15 (k = 15). Holm adjustment on this pair only.
6. Secondary analyses	Component shares, size-mention screen, alternative labels, nested bootstrap, label-error simulation, nested thresholds.
7. Spatial description	OSM footprint as collinear physical-scale check. Block count and longest-block axis are exploratory. Missing footprints are left blank.
8. Interpretation limit	Plans, indoor routes, signage systems and hospital-specific clinical adjacencies were not measured and are not inferred from the texts.

Table 2 is a reading guide for the same sequence drawn in Figure 4. It does not replace Section 4.1, Section 4.2, Section 4.3, Section 4.4, Section 4.5 and Section 4.6.

4.1. Hospitals

The set is the 26 city hospitals named on the Ministry of Health directory retrieved in July 2026 [20]. That page is the sampling frame. It is not a claim that 26 was, in that month, the complete national count of every campus in service. A ministry statement reported in July 2026 described 27 city hospitals in service and about 40,000 beds, with 13 further hospitals under construction [17]. The presidential address at the Samsun opening stated the same operating count and bed total, and described eight additional hospitals still at project stage, 21 further campuses in all [18]. Ordu was reported to have begun accepting patients in July 2026 [19]. Samsun and Ordu were not on the retrieved directory page, and neither was added after the harvest. Campuses still under construction or still at project stage are outside the comparison because they had no directory profile in that retrieval and no review corpus in this dataset. Being named on the directory is an inventory fact. It is not a traced route, and it is not the 15-text rule used in the primary correlations. Procurement was not one method. The Ministry’s account distinguishes classical tender and the general budget from public–private partnership, and states that ten city hospitals were built by the partnership method while many other hospitals were completed from the general budget [15]. This file does not record the contract type of each of the 26 hospitals. Law No. 6428 authorises the partnership route [16]. It does not identify which campuses used it, and a financing statute is not a plan and not a floor area. Announced bed capacity was read from each hospital’s profile on the ministry site in July 2026. Table 3 states the code definitions. Table 4 gives the page. The file Supplementary S24 records 22 September 2026 as the date those profile URLs and bed counts were entered in the identity table around the original submission files. No second harvest of the 566 Yandex texts was performed on that date. No capacity figure in the analytic file was updated after the identity table was recorded. Kahramanmaraş’s ministry profile did not state a count. The hospital’s “about” page states 1040 beds, and that current profile value is the one used. A history page on the same site, dated January 2022, states 1000. The two official figures conflict, and the note is in the supplementary hospital-level table. Kahramanmaraş is not among the hospitals with at least 15 Yandex texts, so the choice does not enter the primary correlations. The figure is announced capacity. It is not occupied beds, and it is not floor area.

Table 3. Code definitions and exclusion rules. The middle column is a definition, not a quotation from a review.

Code	Counts When the Text States It	Does Not Count
Lost	The writer got lost or could not orient	“The hospital is large” alone; a doctor who “disappeared”
Block or floor	The wrong block or floor, or no way to tell which	A block name used only as an address
Signage failure	A sign was missing, contradictory, or insufficient	“The signs were sufficient”
Asked	The writer asked because the route or the signs did not carry the search	Thanks to staff, or asking with no failed search
Long walking	A long indoor walk	Distance from the city to the campus
Mobility	A wheelchair or walking limit is tied to the indoor route	Wheelchair supply, escort, or a clinical limit with no route burden
Transport	Campus access by bus, minibus, or location outside the city was a difficulty	“Access was easy”

Navigation-related language is Lost, Block or floor, Signage failure, or Asked. H2 uses long walking or mobility. The two shares are also reported separately.

Table 4. Ministry page used for announced bed capacity, accessed in July 2026. Supplementary S24 later records 22 September 2026 as the identity-table stamp; that stamp is not a second review harvest.

Hospital	Beds	Page	Accessed
Adana	1680	TR-93923	July 2026
Bilkent	4050	TR-92802	July 2026
Etlik	4050	TR-94168	July 2026
Antalya	1500	TR-102822	July 2026
Aydın	1300	TR-116401	July 2026
Balıkesir	1100	TR-94297	July 2026
Başakşehir	2755	TR-92809	July 2026
Bursa	1435	TR-92806	July 2026
Elazığ	1085	TR-92801	July 2026
Erzurum	1670	TR-95753	July 2026
Eskişehir	1150	TR-92803	July 2026
Gaziantep	1875	TR-100515	July 2026
Göztepe	815	TR-94298	July 2026
Isparta	845	TR-92799	July 2026
Cemil Taşcıoğlu	1005	TR-94299	July 2026
İzmir	2250	TR-100630	July 2026
Kahramanmaraş	1040	TR-29643	July 2026
Kartal	1140	TR-94301	July 2026
Kayseri	1607	TR-92800	July 2026
Kocaeli	1220	TR-95587	July 2026
Konya	1295	TR-92807	July 2026
Kütahya	610	TR-103216	July 2026
Manisa	558	TR-92804	July 2026
Mersin	1440	TR-92797	July 2026
Tekirdağ	605	TR-92808	July 2026
Yozgat	475	TR-92798	July 2026

Pages are on khgm.saglik.gov.tr except Kahramanmaraş, which is the hospital’s “about” page (1040, the current profile value). A history page dated January 2022 states 1000. The full URL and that conflict are in the Supplementary S14. Profile pages were accessed in July 2026, as in the original submission. Supplementary S24 records 22 September 2026 as the date those URLs were entered in the identity table. The 566 Yandex texts were not collected again on 22 September 2026. No capacity figure in the analytic file was updated after that identity table was recorded.

4.2. Review Texts

Three public sources were requested in July 2026. They were not requested again on 22 September 2026. Each Yandex page was the business-reviews URL listed in the Supplementary S24. Texts were split on the reviewer marker “seviye şehir uzmanı Abone ol”, then cut at reply, photo or like markers. On the saved pages, a note stored with the generative-coding file recorded 43 retrieved cells that still contained more than one review after that split. The note did not keep review identifiers. Re-parsing the same pages finds 42 eligible cells that still contain the string “seviye şehir uzmanı”; 20 of those 42 also contain a second full dated stamp, and those 20 are the exclusion file in Supplementary S12. Unique Yandex review identifiers were not retained after chrome was stripped, so the mixed cells cannot be separated into newly labelled reviewer-level texts. They remain retrieved units in the denominator. Section 5.1 and Supplementary S13 state why the exclusion file has 20 records rather than 43, and what stays unidentified. The saved page is the platform’s default order. It is not a latest-N list. A zero parsed count, as at Kayseri, is an empty retrieval in this harvest. It is not evidence that the platform holds no reviews. Şikayetvar was a contrast corpus only: two listing pages, then at most 40 complaint pages per hospital, with the title and the complaint body kept and the platform timestamp stripped. Google Maps returned an HTML shell with no review sentence, and Google text is not in the rates. A dated calendar day exists for 392 of 566 Yandex texts; 174 texts remain undated or lack a year; hospitals do not share one observation window, and the dated reviews precede the July 2026 capacity figures (Section 5.1). Detailed logs remain in Supplementary S15 and S16. No route was traced, so decision points, route length and walking time have no hospital values.

Platforms are not pooled. A star rating without text is outside the denominator. Institutional replies are not visitor reviews. Texts shorter than 40 characters, counted with spaces and punctuation, were dropped at retrieval and were not retained. Their number, their hospital distribution and their code rate were not archived, so those quantities cannot be recovered. A table cell of 0 would imply that a drop count had been observed; no such count was observed, and none is reported. Among the 566 saved Yandex texts, none is shorter than 40 characters; the median length is 129 characters. The 40-character rule was a length screen, not a threshold chosen to maximise a p value. Short texts such as a bare “no signs” could have been informative and are absent from this corpus. The printed count on the saved Yandex page (Page n on the saved Yandex page) is not that drop count. Printed and parsed counts differ because a saved page can combine more than one review in one cell, and because replies and chrome are stripped. Reviews were not collected by searching for the words kaybol or tabela. Names, diagnoses and plate numbers are not fields in the analytic table. The manuscript does not quote writers.

4.3. Spatial-Experience Codes

The counting unit is the text. One text may carry several codes and still counts once in the denominator. Navigation-related language, as used in H1, is the union of four codes: getting lost or being unable to orient; a block or floor that could not be identified; signage described as missing, contradictory or insufficient; and having to ask in order to proceed. The union carries that name because the four marks are not one mechanism. Being lost is an outcome. A sign failure is closer to a visual-communication complaint. Asking is a coping behaviour and is counted only when the text links the question to a failed search. The union is not a measure of the sign system. Section 5.1 reports the four marks one by one. Section 5.3 reports each mark’s rank correlation with announced beds. Those component tests are exploratory and are not Holm-adjusted with H1 and H2. Walking-or-mobility language is a long indoor walk, or a mobility limit that the text ties to that indoor route. A wheelchair word alone is not enough. Campus transport—buses, the campus being outside the city, a missing service—is a third share and is not added to navigation-related language. A sentence that only says the hospital is large or huge is counted as a size mention and is not navigation-related. That size screen is a lexicon rule (çok büyük, kocaman, devasa and close variants, with no lost, block, sign or walking mark). It is not an author-adjudicated outcome. It marked 45 of 566 Yandex texts (41 of 486 in the primary set). Its hospital-level rank correlation with announced beds was ρ = 0.196 (permutation p = 0.4792) and is not used as a third primary test. Positive signage, including a statement that the signs were sufficient, is kept separate and does not enter the navigation-related share (6 Yandex texts). The same rules are listed in Table 3.

The screen is a Turkish lexicon with two guards. A lost-word match is dropped when the preceding words name a doctor, a nurse, a result, an appointment or a belonging, so that a missing physician is not read as a missing corridor. A sign or a transport match is read with the following words, so that “the signs were sufficient” and “access was easy” are not failures. Negation immediately around “kaybolmadı” is not a failure. This screen is a rule list, not the transformer pretraining model described by Devlin and colleagues [43]. Turkish morphology, spelling and irony will move some texts across the boundary. Section 4.3.1 compares two NLI models that applied the same hypotheses. Navigation-related marks are rare, so a missed paraphrase can still move a hospital rank. Exact duplicate Yandex texts were checked by normalised string; none were found. A screen for Turkish-specific letters or a short list of function words marked 564 of the 566 Yandex texts. The other two are still Turkish sentences the short list missed. No text was dropped for language, and no language-identification model was used. Precision on an enriched sample would not, in any case, be performance on the natural balance [44]. The lexicon is retained as a screening method. It is not the primary outcome.

4.3.1. Author Labels and Method Error

All 566 retrieved Yandex texts were coded against Table 3. The two authors did not first code those texts independently of one another and then reconcile a disagreement table. They applied the codebook to the full retrieved set with hospital names visible, and therefore without blinding to announced capacity, and they resolved disagreements with a previously blinded Grok 4.7 reading (made without the lexicon flags) into a third, final label file. Pre-adjudication author–author agreement is therefore not available. Those adjudicated labels remain the primary Yandex outcomes in Section 5. Computational assistance organised the sheets and listed candidate texts; the authors take responsibility for every final label. Because the Grok reading entered that adjudication, agreement between the Grok file and the final labels is not independent validation. Independent human agreement, recorded before any consensus substitution, is the hospital-masked 140-text panel in Section 4.3.4: site codes replaced hospital names, the hypothesis and the author labels were withheld, and the two workbooks were independently shuffled. The lexicon, NLI and generative comparisons in Section 4.3.1, Section 4.3.2 and Section 4.3.3 remain method checks; they are not a substitute for that panel. Şikayetvar contrast labels remain lexicon marks; that corpus was not recoded by hand.

Table 5 compares the lexicon, the Grok reading and two local NLI models with the final labels. The lexicon recovered 17 of 35 navigation-related texts (precision 0.810; recall 0.486) and 14 of 27 walking-or-mobility texts. That aggregate hides variation across the four navigation marks. Against the same final labels, the lexicon recovered 10 of 24 lost texts (precision 1.000; recall 0.417), 2 of 5 block-or-floor texts (precision 0.667; recall 0.400), 7 of 14 signage-failure texts (precision 0.778; recall 0.500), and 0 of 10 asked-for-directions texts (precision 0.000; recall 0.000) [44]. Asking is the component the lexicon missed completely. NLI A marked 477 texts as navigation-related, with 444 false positives (specificity 0.164). NLI B showed the same over-marking pattern. Relative to the final labels, the task-specific lexicon had high precision and incomplete recall; the generic zero-shot NLI classifiers substantially over-classified spatial-experience language. Pass A is MoritzLaurer/mDeBERTa-v3-base-mnli-xnli. Pass B is MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli. Both ran locally in July 2026, multi-label, threshold 0.70. The score is a pipeline entailment score, not a calibrated probability [43]. Gwet’s AC1 is defined in [45]; Hill and colleagues report it beside a human panel [46]. The two NLI passes remain a method-sensitivity check. They do not replace the final labels. Section 4.3.2 reports a second sensitivity check, in which two generative models applied the same codebook.

Table 5. Final author-adjudicated labels versus the lexicon, the Grok reading, two NLI models and two generative models.

Method	Outcome	Gold n	FP	FN	Prec.	Rec.	Spec.	F1
Lexicon	Nav.	35	4	18	0.810	0.486	0.992	0.607
Lexicon	Walk/mob.	27	3	13	0.824	0.519	0.994	0.636
Grok reading	Nav.	35	2	0	0.946	1.000	0.996	0.972
Grok reading	Walk/mob.	27	5	1	0.839	0.963	0.991	0.897
NLI A	Nav.	35	444	2	0.069	0.943	0.164	0.129
NLI A	Walk/mob.	27	289	4	0.074	0.852	0.464	0.136
NLI B	Nav.	35	276	9	0.086	0.743	0.480	0.154
NLI B	Walk/mob.	27	122	10	0.122	0.630	0.774	0.205
Qwen2.5-7B	Nav.	35	19	7	0.596	0.800	0.964	0.683
Qwen2.5-7B	Walk/mob.	27	15	3	0.615	0.889	0.972	0.727
Turkish-Llama-8B	Nav.	34	5	13	0.808	0.618	0.991	0.700
Turkish-Llama-8B	Walk/mob.	27	12	7	0.625	0.741	0.978	0.678

Gold n is the count in the final labels. FP and FN are against those labels. Prec. is precision; Rec. is recall; Spec. is specificity. NLI A is mDeBERTa-v3-base-mnli-xnli; NLI B is multilingual-MiniLMv2-L6-mnli-xnli. Qwen2.5-7B is Qwen/Qwen2.5-7B-Instruct, scored on all 566 texts. Turkish-Llama-8B is ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1, scored on 565 texts: one navigation-related text did not parse, so its navigation gold n is 34. Nav. is the navigation-related union. Walk/mob. is long walking or mobility. Gwet’s AC1 and observed agreement are in Supplementary S17. The Grok reading is not a second person. The generative rows are not a hired two-coder panel.

4.3.2. Generative Double Coding

In July 2026, two instruction-tuned models applied the Section 4.3 codebook to the same 566 texts, independently of each other and of the author labels. Model A is Qwen/Qwen2.5-7B-Instruct. Model B is ytu-ce-cosmos/Turkish-Llama-8b-Instruct-v0.1. Both ran in 4-bit NF4 weights with greedy decoding, a repetition penalty of 1.0 and seed 42. Hospital names were withheld. The prompt carried the seven outcomes and one extra flag, uncertain, for a sentence that cannot be coded 0 or 1 without guessing. When that flag is 1, the seven outcomes stay 0. No parsed text broke that rule. Separate model runs of one corpus have been treated as coders, and a predefined scheme can be scored with Cohen’s κ [47,48]. Generative coding is not a substitute for a human check when the decision is interpretive [49]. A guide to qualitative coding with large language models places that check with the analyst, not with the model [50]. Gwet’s AC1 is reported beside κ because several codes are rare [45,46]. This pass is model–model agreement. It is not a hired two-coder panel.

Model B did not parse one text. That text is one of the 35 navigation-related author labels, so Model B is scored on 565 texts. Two further Model B rows parsed on a second attempt. On the 565 paired texts the seven outcomes matched exactly in 499 texts (88.3%). For the navigation-related union, Cohen’s κ was 0.646 and Gwet’s AC1 was 0.952. For walking or mobility, κ was 0.685 and AC1 was 0.958. A text with at least one of the seven outcomes had κ = 0.782 and AC1 = 0.914 (Model A 107 texts; Model B 86; both 79). On the Landis–Koch band, those κ values sit in the substantial range [51]. They are not high enough to replace a human panel, and they are not high enough to substitute the generative pair for the author labels. Rare positives pull κ down while AC1 stays high; that is why both coefficients are reported. The uncertain flag did not agree: Model A marked 75 texts and Model B marked 5 (κ = 0.059). If that flag is counted, exact agreement falls to 428 of 565 texts and 137 texts disagree. Those 137 texts and the unparsed text were listed for adjudication (138 rows). The adjudication columns were left empty. They were not written into the label file used in Section 5. Disagreements with the author labels are not spread evenly. Of 26 Qwen–author disagreements on the navigation union, 6 are at Etlik and 3 each at Bilkent, Başakşehir and Kocaeli. Of 19 Llama–author navigation disagreements, 5 are at Etlik and 4 at Kocaeli. High-capacity campuses contribute a large share of the hard sentences, which is one reason the Ankara-drop is reported in Section 5.3.

Table 5 places both models against the author labels. Qwen2.5-7B recovered 28 of 35 navigation-related texts (precision 0.596; recall 0.800; κ = 0.659) and 24 of 27 walking-or-mobility texts (precision 0.615; recall 0.889; κ = 0.711). Turkish-Llama-8B recovered 21 of 34 navigation-related texts in its scored set (precision 0.808; recall 0.618; κ = 0.683) and 20 of 27 walking-or-mobility texts (precision 0.625; recall 0.741; κ = 0.660). Both sit between the lexicon and the NLI passes: higher recall than the lexicon on these unions, and far fewer false positives than either NLI model. The author labels remain the primary Yandex outcomes.

Campus transport is the weak code in this pass. Qwen2.5-7B marked 55 texts against 25 author labels (precision 0.255). Turkish-Llama-8B marked 40 texts against 24 labels in the scored set (precision 0.300). On the 565 paired texts, a string screen for otopark, park yeri or park sorunu matched 42 of 54 Qwen transport marks and 29 of 40 Turkish-Llama transport marks. A screen for otobüs, minibüs or dolmuş matched 2 and 1 of those marks. The codebook counts bus, minibus or a site outside the city. It does not count a parking complaint by itself. The generative transport marks are not an alternative transport share.

Returned evidence spans were checked as substrings of the text. Qwen2.5-7B had 18 spans that were not in the text (90 spans were verbatim). Turkish-Llama-8B left the evidence field empty on 63 texts, returned 19 verbatim spans and had 4 spans that were not in the text. Item-level codes, without review wording, are in Supplementary S18 and S19. Model–model agreement is in Supplementary S20. Mixed retrieved cells are treated in Section 4.2 and Section 5.1, not as a generative-coding result.

4.3.3. Classification Uncertainty and Alternative Labels

The Wilson intervals in the supplementary table describe sampling uncertainty of a share only if the retrieved texts are treated as a sample. They do not absorb misclassification. Two further checks were added in this revision. A nested bootstrap resamples the 15 primary hospitals with replacement and then resamples texts within each drawn hospital (10,000 draws, seed 22). That interval is wider than the hospital-only bootstrap because it puts the size of each denominator back into the share. For the navigation-related share the nested 95% interval was −0.165 to 0.749 (median ρ = 0.353). For walking or mobility, it was 0.008 to 0.831 (median ρ = 0.527). A label-error simulation then flips each text with the false-positive and false-negative rates observed against the author labels (5000 draws). Under the Grok disagreement rates, the navigation coefficient stayed near 0.50 (median ρ = 0.503; 95% interval 0.297 to 0.566). Under the Qwen rates, which include a 0.20 false-negative rate on navigation-related texts, the median fell to 0.299 (interval −0.120 to 0.633). Recomputing H1 and H2 on the Grok labels, the Qwen labels and the Llama labels, and after dropping texts where the author and Qwen disagree, is reported in Section 5.6. A separate human-disagreement simulation, using the A–B mismatch rates from Section 4.3.4, is reported in Section 5.7. The Grok/Qwen/Llama checks are not an independent human panel.

4.3.4. Independent Human Coding Panel

Two independent coders marked 140 Yandex texts (seed 22): all 55 texts that carry a navigation-related or walking-or-mobility mark in the author file, together with 85 texts that carry neither mark. Hospital names in the sheets were replaced by site codes and stripped from the wording. The hypothesis, the author labels, the lexicon flags and the model files were withheld. The two workbooks were independently shuffled. Each row asked for the four navigation marks, the two walking marks, campus transport and an uncertain flag. If uncertain = 1, the other seven codes stay 0 for the union scores. Agreement is Cohen’s κ and Gwet’s AC1 for each mark and for the two unions, with coder-versus-coder precision and recall and the hospital distribution of disagreements (Supplementary S10). On the navigation-related union, κ = 0.913 and Gwet’s AC1 = 0.957 (97.1% agreement). On the walking-or-mobility union, κ = 0.921 and AC1 = 0.983 (98.6% agreement). Lost was 0.837; long walking 0.921. Block-or-floor, signage failure, asking and mobility were 1.000, on very small positive counts (2, 13, 7 and 1). Coder A’s navigation-related positives were a subset of coder B’s (27 versus 31; four texts B-only). The same nesting held for walking (13 versus 15). The workbooks were not identical. Uncertain was the most discordant extra flag (κ = 0.763). Union disagreements were not concentrated on one campus (Kocaeli 2; Etlik, Konya, Aydın and Bilkent 1 each). The workbooks contain review wording and are not deposited. The English codebook is Supplementary S8b.

4.4. Association

The two primary tests are H1 and H2 on the Yandex corpus. Campus transport, the Şikayetvar contrast corpus, the split of walking into long-walk and mobility language, the four navigation component shares, the size-mention screen, the threshold repeats, mixed-cell exclusion, and calendar-window subsets are exploratory. Hospitals with at least 15 coded texts enter the primary rank correlation. Fifteen is an analytic choice, not a theoretical constant: below that count, one positive text moves a share by more than six percentage points. The same correlations are repeated at 10, 20 and 30 texts. Those repeats are nested subsets of the same texts, not extra samples. Hospitals below 15 texts remain in Section 5.2. The p value is a two-sided permutation of the hospital shares, 100,000 shuffles, seed 22, with one added to the numerator and the denominator. A hospital-level bootstrap, 10,000 paired draws, seed 22, resamples hospitals once and recomputes both correlations and their difference on the same draw. Holm adjustment is applied to the two primary p values only. Fisher’s z approximation gives a minimum detectable Spearman coefficient of 0.669 at k = 15 for 80% power and two-sided α = 0.05. That figure is an approximation only; it is not the power of this zero-heavy permutation design. A permutation Monte Carlo that preserves the observed share marginal, including zeros, and tests each replicate with the same Spearman–permutation procedure (1500 simulations; 3000 inner permutations) gave estimated power of 0.53 for navigation and 0.46 for walking if the true ρ were 0.51 at k = 15. At k = 13, after dropping Bilkent and Etlik, the same walking simulation gave power of 0.37. Non-significant Holm tests at the primary threshold are therefore compatible both with a weak association and with a moderate association that this sample cannot resolve. Section 5.4 reports k, text totals and the number of hospitals with a zero share at each nested threshold. Wilson intervals in the supplementary table describe sampling uncertainty of a share only if the retrieved texts are treated as a sample. They do not absorb misclassification. Kendall’s tau-b and leave-one-hospital-out repeats are robustness checks on the walking association. A further check that uses the denominator is a hospital-level binomial generalised linear model. For hospital j, the marked count Y_j is treated as Binomial(n_j, p_j), where n_j is the number of eligible retrieved Yandex texts and logit(p_j) = β0 + β1 log(beds_j). Estimation is iteratively reweighted least squares (40 iterations; starting values β0 = −4, β1 = 0). Inferential p values are not Wald statistics. They are two-sided permutations that shuffle the pairs (Y_j, n_j) jointly against the bed vector (10,000 shuffles, seed 22), with p = (extreme + 1)/(nperm + 1). Pearson dispersion is Pearson χ² divided by k − 2. That model is exploratory. It does not replace H1 or H2.

4.5. OpenStreetMap Campus Metrics

Physical scale was measured after the text analysis, from OpenStreetMap, via the Overpass API [52,53]. Supplementary S23 records Overpass timestamps of 24 September 2026. The declared primary OSM quantity is footprint_m2: the union of mapped building polygons whose representative point falls inside the campus polygon. OpenStreetMap was not used to estimate foot traffic, indoor routes or movement between storeys. The footprint contains no pedestrian count and no path through a multi-storey interior. Walking-or-mobility language comes only from the retrieved texts. Matching hospital polygons within 300 m were merged, which is intended to join separately mapped parts of one campus. Campus identifiers were checked against a 26-panel plot. İzmir was restricted to the Bayraklı city-hospital polygon after a private hospital with a similar name entered the name query. Erzurum uses the polygon still tagged as the regional training hospital because that is the amenity polygon at the city-hospital helipad. Kartal uses the polygon tagged as the former training hospital. Elazığ and Tekirdağ have campus outlines but almost no indoor building parts; their footprint cells are left blank. A blank is not a zero. Gross floor area is reported only where building:levels tags cover at least 80% of footprint area. Block count, the long axis of the largest block and site area are exploratory. This OSM layer was not locked in a dated protocol before the review labels were seen. Among hospitals that also meet the 15-text rule, the reportable OSM tests are the Spearman correlation of footprint with each language share and the rank partial correlation of footprint with that share given announced beds, 100,000 permutations, seed 22. Collinearity was quantified as the two-predictor variance inflation factor VIF = 1/(1 − r²), where r is the Pearson correlation of announced beds with footprint on the k = 14 hospitals that have a mapped footprint. On the raw scale, r = 0.948 and VIF = 9.85. After a natural-log transformation of both variables, r = 0.915 and VIF = 6.15. The rank coefficient used elsewhere in the paper is Spearman ρ = 0.876 and is not the r inside the VIF formula. Conventional linear-model guidance treats VIF above 5 as concerning and above 10 as severe. Both the raw and the log values are high enough that beds and footprint are not entered together as competing predictors. Announced beds remain the single primary scale variable. Footprint is retained only as a collinear physical-scale check and as rank partial correlations given beds (Section 5.5). Section 5.5 also reports the partial correlations after dropping the three manually overridden campuses, after dropping campuses that still carry former or training-hospital names, after removing Bilkent and Etlik, and after filling the two blank footprints with site area as a documented sensitivity that is not recommended as a primary rule. The k = 11 override-drop row and the k = 12 former-or-training-name row are exploratory subset sensitivities: they depend on which campuses are excluded, they coincide with still-severe collinearity, they were not the pre-specified OSM test, and they do not replace the listwise k = 14 result. Section 5.5 lists the OSM descriptors and the recorded opening year for the 15 hospitals that enter the primary tests. Those rows are not photographs, not as-built plans, and not a staff or healing-space inventory. Section 2.2.2 and Figure 3 describe plan extent. The mapped axes, from 205 m to 656 m, are horizontally extended campuses rather than towers. A full height ranking is not made: the 80% storey-tag rule holds for only one primary campus, Bilkent, where the area-weighted mean is 6.95 storeys (Section 2.2.2).

4.6. Ethics and Software

The retrieval used already public web pages: Yandex Maps business-review pages and Şikayetvar complaint pages. Those texts were posted by their writers before this study. No person was recruited and no writer was contacted. No survey, interview, observation or experiment was administered. Health details and names that may sit in a public review were not extracted as variables, and writers are not quoted. Analyses are hospital-level shares. Identifiable comments are not reproduced and raw wording is not deposited. Ethics-committee approval is not required for this design. The materials are unsolicited public documents, not a recruited human-subjects sample. A numbered committee waiver was not issued because committee review is not applicable, not because an application was refused. The legal basis recorded here is Turkish Law No. 6698 on the Protection of Personal Data (KVKK), Article 5(2)(d) (processing of personal data that the data subject has made public) and Article 28(1)(ç) (scientific research with data that do not identify the person). The rank correlations, permutations and bootstrap intervals were computed in Python 3.12.3 with the project script, using the standard-library random generator, seed 22. OpenStreetMap geometries were handled with GeoPandas 1.1.4 (https://geopandas.org). Figures were drawn with Matplotlib 3.11.0. No separate statistics package produced the coefficients.

5. Results

5.1. Retrieved Corpus

Of the 26 hospitals in the directory frame, 25 contributed eligible Yandex texts. The complete retrieved Yandex corpus contained 566 texts. Author-adjudicated labels marked 35 navigation-related texts and 27 walking-or-mobility texts in that complete set. Fifteen hospitals, contributing 486 texts, met the minimum of 15 texts for the primary analyses. Within this primary subset, 32 texts carried a navigation-related mark and 26 carried a walking-or-mobility mark. Fifteen of the 35 navigation-related texts in the complete corpus carry more than one of the four component marks, so component counts are not added to make the union (Supplementary S1). Şikayetvar complaint articles were saved for 26 hospitals, 670 texts, of which 8 (1.2%) carried navigation-related language. That contrast did not use the same retrieval rule, so it is not an independent confirmation. Google contributed no text. Route length and signage performance were not measured.

The note of 43 mixed cells was stored without review identifiers. Re-parsing the saved pages with the same splitter identifies 42 eligible cells that still contain the string “seviye şehir uzmanı” after the cut on “N. seviye şehir uzmanı Abone ol”. All 42 match a saved text (Supplementary S13). The forty-third cell in the stored note is not reproduced by this parse and cannot be named. The exclusion already reported used a narrower rule inside those 42: a second full dated stamp, with day, month and year. That rule marks 20 texts, and those 20 are exactly the records in Supplementary S12. The other 22 contain the unsplit reviewer marker but not a second full date, which is why the exclusion file has 20 records rather than 43. Of the 42, 32 lie in the original 15-hospital set (6 navigation-related and 1 walking-or-mobility). Across all 566 texts, the 42 carry 7 navigation-related marks and 1 walking-or-mobility mark. Excluding those 20 leaves 546 texts and still 15 hospitals at the 15-text rule (472 primary texts; 30 navigation-related and 26 walking-or-mobility marks). On that reduced set, navigation-related ρ = 0.464 (permutation p = 0.0837; Holm p = 0.1000) and walking-or-mobility ρ = 0.520 (p = 0.0500; Holm p = 0.1000). Those coefficients do not replace H1 or H2. Excluding all 42 recoverable cells, rather than splitting them into new labelled reviews, leaves 524 texts. Konya falls from 16 texts to 14 and Manisa from 16 to 13, so both leave the 15-text set. The rank test is then on 13 hospitals and 427 texts (22 navigation-related and 23 walking-or-mobility marks): navigation-related ρ = 0.679 (permutation p = 0.0140; Holm p = 0.0280) and walking-or-mobility ρ = 0.649 (p = 0.0196; Holm p = 0.0280). That change of hospital set is exploratory. Removing Bilkent and Etlik from those 13 leaves navigation-related ρ = 0.552 (p = 0.0807) and walking-or-mobility ρ = 0.390 (p = 0.2388). Holding the original 15 hospitals fixed, including the two denominators that fall below 15, gives 454 texts (26 navigation-related and 25 walking-or-mobility marks), navigation-related ρ = 0.379 (p = 0.1628; Holm p = 0.1628) and walking-or-mobility ρ = 0.591 (p = 0.0233; Holm p = 0.0465). Neither row replaces the published H1 or H2 pair (ρ = 0.512 and 0.513; Holm p = 0.1040). The same permutation procedure as Supplementary S12 was used (100,000 shuffles, seed 22). The remaining limitation is that the 42 cells were dropped, not separated into reviewer-level texts, and that the unreproduced forty-third cell, if it was a further mixed cell, is still unnamed inside the published denominator.

A common observation period cannot be imposed on the primary 15-hospital set. A dated calendar day exists for 392 of 566 texts; 174 remain undated or lack a year. Dated texts fall mainly in 2024–2025 and precede the July 2026 capacity figures. Restricting to day-dated texts in calendar 2025 leaves 324 texts and only nine hospitals with at least 15 such texts (227 texts; 13 navigation-related and 8 walking-or-mobility marks in that primary subset). That subset is selected for hospitals that currently show enough dated 2025 reviews on the default-order page, still includes Bilkent and Etlik, and is not a shared window for the original 15. Navigation-related ρ = 0.958 (p = 0.0003) on those nine hospitals is exploratory, sample-selection sensitive, and is not a new primary test. Walking-or-mobility ρ = 0.438 (p = 0.2352). A time series of announced beds is not in the analytic file, except the Kahramanmaraş 1000 versus 1040 conflict already reported in Table 4; capacity consistency during 2024–2025 cannot be verified from this file. Unequal windows can favour campuses whose saved pages currently display more recent dated posts.

Table 6 splits the navigation-related outcome into the four marks that the H1 union pools. Across 566 Yandex texts, 14 name a sign as missing, contradictory or insufficient, in 10 hospitals. That count is the visual-communication sentence in this corpus. It is not an audit of the signs. The navigation-related union is 35 texts. Fifteen of those texts carry more than one of the four marks, so the component counts are not added to make the union. In the primary 486 texts, the same four marks were 21 lost, 4 block or floor, 13 signage failure and 9 asked. The table is descriptive. Section 5.3 reports each mark, and the lexicon size-mention screen, as a rank correlation with bed capacity.

Table 6. Composition of the navigation-related outcome in the retrieved Yandex texts.

Component	Texts	% of Yandex Corpus	Hospitals
Lost or unable to orient	24	4.2	14
Block or floor	5	0.9	5
Signage failure	14	2.5	10
Asked for directions	10	1.8	8

Each row is a final-label mark. A text may carry more than one mark and still counts once in the navigation-related union used for H1. Percentages use all retrieved Yandex texts of at least 40 characters, including hospitals below the 15-text threshold. Hospitals are those with at least one text carrying the mark. The table does not enter the rank tests.

5.2. The Hospital Comparison

Table 7 places announced capacity beside the Yandex counts and the OpenStreetMap footprint. A hospital with a small retrieved set can show a large percentage. That percentage is reported, and it is kept out of the rank correlation when the text count is below 15. A zero Yandex denominator, as at Kayseri, means this retrieval returned no eligible text. It does not mean the hospital has no wayfinding problem, and the absence of a traced plan from this paper does not mean that no plan exists. Among the hospitals in the primary set, the navigation-related share ran from 0.0% at Isparta (n = 17) to 25.0% at Konya (n = 16).

Table 7. Announced bed capacity, author-adjudicated Yandex language counts, the printed page count and OpenStreetMap campus footprint.

Hospital	Beds	n	Primary	Nav.	Walk/Mob.	Fp.	Bl.	Page n
Adana	1680	14	No	1 (7.1%)	0 (0.0%)	125,645	26	9
Bilkent	4050	42	Yes	3 (7.1%)	5 (11.9%)	147,739	4	13
Etlik	4050	43	Yes	5 (11.6%)	7 (16.3%)	211,732	15	5
Antalya	1500	33	Yes	1 (3.0%)	3 (9.1%)	43,115	4	7
Aydın	1300	3	No	1 (33.3%)	0 (0.0%)	41,581	1	4
Balıkesir	1100	9	No	0 (0.0%)	1 (11.1%)	26,967	1	17
Başakşehir	2755	48	Yes	3 (6.2%)	0 (0.0%)	145,124	15	47
Bursa	1435	18	Yes	2 (11.1%)	0 (0.0%)	60,437	9	10
Elazığ	1085	1	No	0 (0.0%)	0 (0.0%)	—	—	3
Erzurum	1670	6	No	0 (0.0%)	0 (0.0%)	51,024	5	1
Eskişehir	1150	11	No	0 (0.0%)	0 (0.0%)	58,568	2	3
Gaziantep	1875	13	No	1 (7.7%)	0 (0.0%)	95,623	8	9
Göztepe	815	42	Yes	1 (2.4%)	1 (2.4%)	19,460	6	8
Isparta	845	17	Yes	0 (0.0%)	0 (0.0%)	28,611	2	1
Cemil Taşcıoğlu	1005	27	Yes	0 (0.0%)	0 (0.0%)	24,709	4	9
İzmir	2250	44	Yes	3 (6.8%)	3 (6.8%)	104,520	11	12
Kahramanmaraş	1040	4	No	0 (0.0%)	0 (0.0%)	23,543	5	4
Kartal	1140	48	Yes	2 (4.2%)	1 (2.1%)	63,679	18	6
Kayseri	1607	0	No	–	–	65,644	5	1
Kocaeli	1220	44	Yes	5 (11.4%)	4 (9.1%)	47,565	1	5
Konya	1295	16	Yes	4 (25.0%)	2 (12.5%)	50,018	5	13
Kütahya	610	10	No	0 (0.0%)	0 (0.0%)	21,920	2	2
Manisa	558	16	Yes	1 (6.2%)	0 (0.0%)	31,894	2	2
Mersin	1440	18	Yes	1 (5.6%)	0 (0.0%)	80,424	11	1
Tekirdağ	605	30	Yes	1 (3.3%)	0 (0.0%)	—	—	3
Yozgat	475	9	No	0 (0.0%)	0 (0.0%)	14,077	1	9

Beds are the announced capacities in Table 4. Each language cell is the count of marked texts and that count as a percentage of Yandex texts of at least 40 characters. They are not visitor prevalences. A dash in a language cell is a zero denominator. Fp. is the OSM building-footprint union, m². Bl. is the OSM block count. A dash in Fp. or Bl. means indoor building parts were not mapped. A blank is not a zero. Primary marks hospitals with at least 15 Yandex texts. Page n is the count printed on the saved Yandex page. It is not the number of texts dropped by the 40-character rule. Those discarded strings were not retained and were not archived. A 0 in that sense is not available and is not written. Printed and parsed counts differ because a saved page can combine reviews and because chrome is stripped.

5.3. Capacity and the Two Language Shares

Across the 15 hospitals with at least 15 retrieved texts, the Spearman rank correlation between announced bed capacity and the navigation-related share was 0.512, bootstrap 95% CI 0.040 to 0.800 (two-sided permutation p = 0.0534, 100,000 shuffles). The correlation with the walking-or-mobility share was 0.513, bootstrap 95% CI -0.015 to 0.853 (two-sided permutation p = 0.0520, 100,000 shuffles). Those two results are H1 and H2. Neither Holm-adjusted p value is below 0.05. Figure 5 plots bed capacity against the navigation-related share for hospitals at or above the text threshold. Figure 6 does the same for walking-or-mobility language. Long walking, mobility, campus transport and the Şikayetvar contrast are the remaining rows of Table 8. H2 remains the walking-or-mobility composite. Şikayetvar is not plotted, and Şikayetvar shares remain lexicon marks.

Figure 5. Announced bed capacity and the share of retrieved Yandex texts with navigation-related language. Point area is proportional to the number of retrieved texts; the legend shows n = 15 and n = 40. Vertical bars are Wilson intervals for the retrieved texts. They are not intervals for a patient population.

Figure 6. Announced bed capacity and the share of retrieved Yandex texts with walking-or-mobility language. Point area is proportional to the number of retrieved texts; the legend shows n = 15 and n = 40. Vertical bars are Wilson intervals for the retrieved texts. They are not intervals for a patient population.

Table 8. Primary tests on author-adjudicated language shares, OSM footprint tests, exploratory splits and the Şikayetvar contrast, for hospitals with at least 15 texts.

Outcome	Role	k	ρ	95% CI	p	Holm p
Nav. lang.	Primary H1	15	0.512	0.040 to 0.800	0.0534	0.1040
Walk/mob.	Primary H2	15	0.513	−0.015 to 0.853	0.0520	0.1040
Long walk	Split	15	0.533	−0.028 to 0.862	0.0423	—
Mobility	Split	15	0.339	−0.134 to 0.721	0.2175	—
Campus access	Exploratory	15	0.399	−0.189 to 0.790	0.1411	—
OSM footprint–walk/mob.	OSM primary	14	0.355	−0.248 to 0.803	0.2113	—
OSM footprint–nav.	OSM primary	14	0.604	0.040 to 0.897	0.0242	—
SV nav.	Contrast	26	0.106	−0.301 to 0.495	0.6046	—
SV walk/mob.	Contrast	26	−0.196	−0.553 to 0.190	0.3348	—

Yandex rows use the final author-adjudicated labels. Şikayetvar rows remain lexicon marks. Confidence intervals are hospital-level bootstrap percentiles. Holm adjustment is applied only to H1 and H2. OSM footprint rows use 14 hospitals: Tekirdağ has an OSM campus outline but no mapped building parts and is left blank. Partial Spearman coefficients given announced beds are in Section 5.5. Split, transport, OSM and Şikayetvar rows are not Holm-adjusted.

Table 8 collects the primary tests and the exploratory splits. Holm adjustment of the two primary p values leaves both at 0.1040. Those two tests are not statistically supported. Kendall’s tau-b is 0.329 for navigation-related language and 0.409 for walking-or-mobility language. The difference, walking rho minus navigation-related rho, is 0.001 (bootstrap 95% CI −0.522 to 0.468). The interval includes zero, so the two associations are not shown to differ. Near-equal point estimates are not a tested dissociation. Leave-one-hospital-out walking rho ranged from 0.397 to 0.637. Removing both Ankara hospitals, Bilkent and Etlik, leaves rho = 0.269 on the remaining 13 hospitals (permutation p = 0.3674; bootstrap 95% CI −0.290 to 0.733). The interval includes zero, and the p value is above 0.05. Two readings of that drop should be kept together. Twelve of the twenty-seven walking-or-mobility texts in the whole Yandex corpus are from Bilkent and Etlik, so the point estimate is partly carried by those two high-capacity campuses. At the same time, permutation Monte Carlo power at k = 13, if the true walking ρ were still 0.51, is about 0.37. Fisher’s z 80% MDE at k = 13 is about 0.71 and is an approximation only. The wider interval is therefore both a leverage result and a power result. It does not, by itself, show that the remaining hospitals have no association. The four navigation marks are reported separately in Section 5.4. Asking for directions ranked with capacity (ρ = 0.701; permutation p = 0.0053; 9 marked texts in the primary set; 8 of 15 hospitals have a zero share). Lost language was weaker (ρ = 0.365; p = 0.1796). Signage-failure language was near zero (ρ = 0.069; p = 0.8026). Block-or-floor language was also weak (ρ = 0.106; p = 0.7104; 5 texts in the full corpus). Those component tests are exploratory, are not Holm-adjusted, and do not rescue H1. An unadjusted p below 0.05 on nine asking texts is not a confirmed mechanism and is not a signage result. A one-factor summary of the four hospital-level component shares is equally limited. Cronbach’s α was 0.405. The first principal component explained 46% of the standardised variance, loaded mainly on lost and asked, and ranked with beds at ρ = 0.608 (p = 0.0189). Pairwise hospital-level Spearman coefficients among the four shares were low except for lost with asked (ρ = 0.513). Several indicators are zero-heavy at hospital level (Supplementary S2). A confirmatory latent-variable model is not justified at k = 15 with these event counts; we therefore report disaggregated components rather than impose an empirically unsupported common factor. Replacing beds with their logarithm does not change the walking Spearman coefficient (ρ = 0.513), because the logarithm is strictly increasing and the tied bed counts stay tied. A Pearson correlation of the walking share with log beds is 0.627 (permutation p = 0.0127). A binomial model of walking counts on log beds, using the same 15 hospitals and 26 marked texts, has a positive coefficient (1.146). Positive counts and their denominators were permuted together against bed counts (permutation p = 0.0404; 10,000 shuffles). Pearson dispersion is 1.434. That value does not show strong extra-binomial scatter in this fit, but with 15 hospitals it is only a limited check. Doubling announced beds multiplies the model odds by about 2.21, which is a transformation of this coefficient, not a causal effect. The same model for navigation-related counts does not show that association (coefficient 0.456; permutation p = 0.1876). Announced capacity and the number of retrieved Yandex texts, across all 26 hospitals, had rho = 0.256 (permutation p = 0.2076). Median beds were 1295 among the 15 hospitals with at least 15 texts and 1150 among the 11 below that threshold. Of 566 Yandex texts, 7 carried both codes, 28 carried navigation-related language only and 20 carried walking-or-mobility language only (Jaccard 0.127; phi 0.18). The binomial check specified in Section 4.4, on the same 15 hospitals, gives β1 = 1.146 for walking-or-mobility counts (permutation p = 0.0404; 26 marked texts; intercept −11.512; Pearson dispersion 1.434) and β1 = 0.456 for navigation-related counts (permutation p = 0.1876; 32 marked texts; intercept −6.037; dispersion 1.426). Doubling announced beds multiplies the fitted walking odds by about 2.21; that transformation is not a causal effect. Wide bootstrap intervals, the influence of Bilkent and Etlik, and the exploratory status of the binomial, mixed-cell and calendar-window rows are retained in the interpretation.

5.4. Threshold Sensitivity

Table 9 repeats the rank tests at minimum text counts of 10, 15, 20 and 30. Each row is a nested subset of the same texts, not an independent replication. A hospital that falls below a threshold leaves that column. Where fewer than eight hospitals remain, the cell is a count rather than a coefficient. Both primary coefficients stay positive across the nested thresholds; unadjusted p values move with the subset. Table 10 reports the four component correlations. Table 11 gives the hospital count, the text total and the number of hospitals with a zero navigation or walking share. At the primary threshold, two hospitals have a zero navigation share and seven have a zero walking share. Fisher’s z 80% MDE at k = 15 is 0.67; permutation Monte Carlo power if the true ρ were 0.51 is 0.53 (navigation) and 0.46 (walking). The observed coefficients of about 0.51 therefore sit in a low-power region for this design. The larger coefficients at 20 and 30 texts are estimated on 10 and 9 hospitals; those rows have still less power, not more independent information. The Yandex walking-or-mobility coefficient stayed positive at every nested threshold: 0.412 at 10 texts (p = 0.0798, 19 hospitals); 0.513 at 15 texts (p = 0.0520, 15 hospitals); 0.648 at 20 texts (p = 0.0499, 10 hospitals); 0.624 at 30 texts (p = 0.0766, 9 hospitals). Campus-transport coefficients weaken once the smallest retrieved pages are left out. Şikayetvar navigation-related language does not rank with bed capacity at the thresholds that still include the full set of hospitals.

Table 9. Threshold sensitivity of the Spearman correlations. Nested subsets, not independent replications.

Min. n	Yandex Nav.	Yandex Walk/Mob.	Yandex Transp.	SV Nav.
10	0.578/0.0108/19	0.412/0.0798/19	0.309/0.1962/19	0.106/0.6046/26
15	0.512/0.0534/15	0.513/0.0520/15	0.399/0.1411/15	0.106/0.6046/26
20	0.723/0.0227/10	0.648/0.0499/10	0.418/0.2276/10	0.106/0.6046/26
30	0.686/0.0466/9	0.624/0.0766/9	0.370/0.3223/9	k = 3; not estimated (fewer than 8 hospitals)

Each cell is rho, the two-sided permutation p value and the number of hospitals. “not estimated” means fewer than eight hospitals remained. Yandex walking is the walking-or-mobility union. Nested thresholds are not independent replications.

Table 10. Component-level rank correlations with announced bed capacity among hospitals with at least 15 Yandex texts.

Outcome	Texts in 566	Texts in 486	ρ (k = 15)	Permutation p
Lost or unable to orient	24	21	0.365	0.1796
Block or floor	5	4	0.106	0.7104
Signage failure	14	13	0.069	0.8026
Asked for directions	10	9	0.701	0.0053
Navigation-related union (H1)	35	32	0.512	0.0534
Long walking	22	21	0.533	0.0423
Mobility	7	7	0.339	0.2175
Walking-or-mobility union (H2)	27	26	0.513	0.0520
Size mention only (lexicon)	45	41	0.196	0.4792

H1 and H2 p values are the original hospital-level permutations (100,000 shuffles, seed 22). Component and size-mention rows are exploratory and are not Holm-adjusted. Size mention is a lexicon screen, not an author-adjudicated outcome. Asking rests on nine primary-set texts.

Table 11. Nested threshold, hospital count, text total and zero-share hospitals.

Min. n	k Hospitals	Texts	Nav. Zeros	Walk Zeros	Nav. ρ (p)	Walk ρ (p)
10	19	534	4	11	0.578 (0.0108)	0.412 (0.0798)
15	15	486	2	7	0.512 (0.0534)	0.513 (0.0520)
20	10	401	1	3	0.723 (0.0227)	0.648 (0.0499)
30	9	374	0	2	0.686 (0.0466)	0.624 (0.0766)

Nested thresholds are not independent replications. Nav. Zeros and Walk Zeros count hospitals with a zero share on that outcome. Fisher’s z 80% MDE at k = 15 is 0.669 and is an approximation only. Permutation Monte Carlo power if the true ρ were 0.51, preserving the observed share marginal including zeros, is 0.53 for navigation and 0.46 for walking at k = 15 (1500 simulations; 3000 inner permutations). The observed primary coefficients of about 0.51 sit below the 80% Fisher MDE and in a low-power region of the permutation design.

VIF is 1/(1 − r²) from the Pearson correlation of announced beds with footprint on k = 14 hospitals with a mapped footprint. On the raw scale, r = 0.948 and VIF = 9.85; on the log–log scale, r = 0.915 and VIF = 6.15. Both remain high. Partial coefficients are rank partial Spearman given beds, 100,000 permutations, seed 22. Filling a blank footprint with site area substitutes campus outline for indoor building parts and is a sensitivity only.

Within the retrieved Yandex frame, both primary rank tests were positive in sign and both Holm-adjusted p values were 0.104. H1 is the navigation-related association. H2 is the walking-or-mobility association. Absence of support after Holm adjustment is not evidence that either association is zero. Both statements are limited to the default-order pages in Section 4.2. Yandex labels for those tests are the author-adjudicated file in Section 4.3.1. The generative pair in Section 4.3.2 was not substituted for that file. Independent human coding of a 140-text sample is in Section 4.3.4 and Section 5.7; it does not replace H1 or H2.

5.5. OpenStreetMap Physical Scale

Figure 3 is the plan drawing for three of the campuses whose footprints are tested in this section. The plans are horizontally extended ground projections, not towers. The drawing is not a footfall surface, not an indoor route and not a section. Building footprints were recovered for 24 of the 26 directory hospitals. Elazığ and Tekirdağ retain campus outlines but no usable indoor building parts; those footprint cells are blank. Among the 14 hospitals that also met the 15-text rule, footprint ranked strongly with announced beds (Spearman ρ = 0.876; Pearson r = 0.948; raw-scale VIF = 9.85; log-scale VIF = 6.15). The unadjusted association of footprint with walking-or-mobility language was 0.355 (permutation p = 0.2113; bootstrap 95% CI −0.248 to 0.803). The association with navigation-related language was 0.604 (p = 0.0242; CI 0.040 to 0.897). Because footprint and beds are collinear, the reportable OSM result is the rank partial correlation given announced beds. Partial ρ was −0.128 for walking-or-mobility language (p = 0.6691) and 0.399 for navigation-related language (p = 0.1597). Neither partial test supports an independent footprint association once capacity is held in the ranks. Table 12 holds the matching sensitivities. Filling Tekirdağ’s blank footprint with site area, which is not indoor building area, leaves the walking partial near zero (partial ρ = −0.144; p = 0.6139). Dropping the three manually overridden campuses raises the navigation partial (partial ρ = 0.703; p = 0.0156) on only 11 hospitals and a VIF of 23.01. Dropping campuses that still carry former or training-hospital names similarly raises the navigation partial (partial ρ = 0.652; p = 0.0155) on 12 hospitals and a VIF of 10.43. Those two p values are exploratory and sample-selection sensitive. They were not pre-specified confirmatory tests, they appear only after hospitals are dropped, and they do not overturn the listwise k = 14 partial of 0.399 (p = 0.1597), which remains the reportable OSM result. Neither subset row is a new primary finding that footprint is independently associated with navigation-related language once beds are held in the ranks. Block count did not rank with either language share. The long axis of the largest block, an exploratory circulation-scale proxy, ranked with walking-or-mobility language (ρ = 0.678; p = 0.0100; partial ρ given beds = 0.663; p = 0.0087). That axis test was not the declared OSM primary and is not Holm-adjusted against the other OSM metrics. Supplementary S21 stores the four metrics and both outcomes. Supplementary S22 is the 26-campus check plot. Table 13 lists the campus descriptors for the 15 hospitals in the primary set.

Table 12. OpenStreetMap footprint collinearity and campus-matching sensitivity.

Scenario	k	VIF	Partial ρ Nav. (p)	Partial ρ Walk (p)
Listwise blank footprint (primary)	14	9.85	0.399 (0.1597)	−0.128 (0.6691)
Drop three override campuses	11	23.01	0.703 (0.0156)	−0.135 (0.6938)
Drop former or training names	12	10.43	0.652 (0.0155)	−0.098 (0.7748)
Drop Bilkent and Etlik	12	7.90	0.426 (0.1704)	−0.265 (0.4098)
Fill blanks with site area (not recommended)	15	3.70	0.306 (0.2725)	−0.144 (0.6139)

Table 13. Campus descriptors available for the 15 hospitals with at least 15 Yandex texts.

Hospital	Beds	Opened	Site ha	Fp. m2	Blocks	Axis m
Bilkent	4050	2019	30.5	147,739	4	597
Etlik	4050	2022	102.2	211,732	15	656
Başakşehir	2755	2020	49.5	145,124	15	513
İzmir	2250	2024	54.5	104,520	11	458
Antalya	1500	2024	22.0	43,115	4	362
Mersin	1440	2017	23.5	80,424	11	272
Bursa	1435	2019	59.4	60,437	9	281
Konya	1295	2020	33.4	50,018	5	282
Kocaeli	1220	2023	4.8	47,565	1	358
Kartal	1140	2020	23.2	63,679	18	256
Cemil Taşcıoğlu	1005	2020	5.7	24,709	4	254
Isparta	845	2017	9.2	28,611	2	245
Göztepe	815	2020	12.3	19,460	6	277
Tekirdağ	605	2020	11.4	—	—	—
Manisa	558	2018	9.3	31,894	2	205

Tekirdağ retains a campus outline without usable indoor building parts. Supplementary S9 stores the same rows as a csv.

Opened is the year in Supplementary S25, not a construction-start date and not a phased-block history. Site and footprint are OpenStreetMap areas. Blocks and the long axis of the largest mapped block describe horizontally extended campuses. Figure 3 draws three of these plans at one scale. Each panel is titled Graphic scale (200 m) and Solar orientation (north). The columns are not a full height ranking. A dash is an unmapped indoor footprint, not a zero. Photographs, plans, staff counts and healing spaces are not in this table.

5.6. Alternative Labels and Label-Error Simulation

Replacing the author navigation labels with the Grok reading left a positive rank correlation with beds (ρ = 0.566; p = 0.0304). The Qwen labels gave ρ = 0.604 (p = 0.0193). The Llama labels gave ρ = 0.260 (p = 0.3454). For walking or mobility, the same substitutions gave ρ = 0.661 (Grok), 0.622 (Qwen) and 0.532 (Llama). Dropping the 26 texts on which the author and Qwen navigation labels disagree left ρ = 0.391 (p = 0.1487) on the remaining texts, still grouped by the same 15 hospitals. The walking coefficient after dropping Qwen walking disagreements was ρ = 0.504 (p = 0.0577). These rows are method checks. They are not the independent human panel. The Grok substitution is particularly not independent, because the Grok reading was one of the files from which the authors formed the final labels. Human-coder substitutions are in Section 5.7. The nested hospital-then-text bootstrap in Section 4.3.3 is the interval that puts text-level sampling back into the share. It is wider than the hospital-only interval used in Table 8, and the navigation nested interval includes zero. The Grok-rate error simulation did not move the navigation coefficient through zero. The Qwen-rate simulation, which assumes a much higher false-negative rate, did. A correlation of about 0.51 in this corpus is therefore sensitive to how missed navigation paraphrases are treated, and it is not sensitive to the small Grok disagreement rate.

5.7. Independent Human Coding and Label Sensitivity

The primary H1 and H2 tests remain the author-adjudicated labels (ρ = 0.512 and 0.513; published permutation p = 0.0534 and 0.0520; Holm p = 0.1040). They are not replaced by the independent panel. Where both coders agreed, those labels were written onto the 140-text sample and the unsampled author-negative texts were left unchanged. Under that consensus, primary-set navigation-related texts fell from 32 to 28 (ρ = 0.430, p = 0.1098) and walking-or-mobility texts from 26 to 15 (ρ = 0.620, p = 0.0165; Holm p = 0.0330). Using coder A or coder B alone on the sample gave the same pattern: navigation weaker, walking stronger. Dropping the texts on which both coders agreed with each other and disagreed with the author left 471 primary texts (navigation ρ = 0.368, p = 0.1756; walking ρ = 0.609, p = 0.0191). Both coders agreed with each other but not with the author on four navigation-related texts and on 12 walking-or-mobility texts. The independent raters were stricter than the authors on walking language. That split is a sensitivity result. It is not a new primary test and does not rescue H1. A nested hospital-then-text bootstrap that also flips labels at the observed A–B disagreement rates (2.9% navigation, 1.4% walking; 5000 draws) gave 95% intervals of −0.221 to 0.700 for navigation and −0.046 to 0.814 for walking. Using author-versus-coder-A mismatch rates instead (5.7% and 10.0%) gave −0.267 to 0.677 and −0.211 to 0.739. Those intervals include zero. They describe clustered sampling plus plausible human label error; they are not a reason to treat the A–B kappas as low. Table 14 reports the two-coder agreement on the 140-text sample.

Table 14. Independent two-coder agreement on the 140-text blinded sample (seed 22). Unions are scored after applying the uncertain rule. Workbooks are not deposited.

Code	n	κ	Gwet AC1	Agree	A+	B+
Lost	140	0.837	0.954	0.964	15	20
Block or floor	140	1.000	1.000	1.000	2	2
Signage failure	140	1.000	1.000	1.000	13	13
Asked directions	140	1.000	1.000	1.000	7	7
Navigation-related union	140	0.913	0.957	0.971	27	31
Long walking	140	0.921	0.983	0.986	13	15
Mobility	140	1.000	1.000	1.000	1	1
Walking-or-mobility union	140	0.921	0.983	0.986	13	15
Transport (not a primary outcome)	140	1.000	1.000	1.000	9	9
Uncertain flag	140	0.763	0.948	0.957	17	11

n = 140 Yandex texts. κ is Cohen’s kappa; AC1 is Gwet’s AC1. Agree is raw percent agreement. A+ and B+ are positive counts. Landis–Koch: 0.81–1.00 almost perfect; 0.61–0.80 substantial. Coder A’s positives on the two unions were nested in coder B’s. Block-or-floor and mobility have two and one positives; perfect agreement there does not validate those rare codes by itself. Primary H1/H2 remain the author-adjudicated tests in Table 8.

6. Discussion

On the retrieved Yandex frame, announced bed capacity ranked similarly with navigation-related language and with walking-or-mobility language. Both Spearman coefficients were about 0.51, and neither Holm-adjusted p value was below 0.05. The primary pair is therefore not statistically supported. The nested hospital-then-text interval for navigation included zero (−0.165 to 0.749). At k = 15, permutation Monte Carlo power if the true ρ were 0.51 is about one half. Once the navigation union is split, asking is the only component with an unadjusted p below 0.05 (ρ = 0.701; nine texts); signage-failure language is near zero (ρ = 0.069). That split is exploratory and does not rescue H1.

The labelled pattern is a reason to assess physical access and navigation separately. It does not identify the mechanism of either outcome. In the retrieved texts, 14 Yandex reviews, spread across 10 hospitals, describe a sign as missing, contradictory or insufficient. That count is a user complaint about signs. It is not an audit of typography, continuity or symbol comprehension. That sign complaint is the visual-communication mark inside the navigation-related union. The union is not a visual-communication score.

The distinction can be shown without quoting a writer. A navigation illustration is a sentence in which the clinic cannot be found and the block cannot be named. A visual-communication illustration is a sentence in which a sign is missing or contradicts another sign. An assistance illustration is a sentence in which the writer has to ask in order to go on. A walking illustration is a sentence in which the way is clear and the indoor walk is still described as long. These four sentences mark the code boundary. They are not quotations, they are not tied to a named hospital, and they do not report a measured minute or a measured metre.

Four mechanisms could produce a capacity–walking association, and none was identified here. The route from entrance to clinic may be longer. Clinics may sit far from laboratories and imaging, including where a service core and clinic wings lengthen the public path even if they shorten staff movement. Patient volume and the tendency to post may differ with announced capacity. A walk described in a review may also begin in the car park rather than at the clinic door. Campus edges were not mapped, so that last reading is not a fact about these 26 hospitals. A fifth, still unmeasured, mechanism is the medical programme together with the building technology that places it, as Section 2.2.1 sets out. Those features make the complex a hospital rather than an ordinary large building, and they were not entered as predictors. Two campuses with similar bed counts can differ in which departments must be adjacent and which must be isolated, and that difference can change walking and choice load without changing announced capacity [21,22,23]. The long-walking component, rather than the mobility component, carries more of the capacity association. That is still a labelled-text result. OpenStreetMap footprint, added after the labels, ranks with announced beds (raw VIF = 9.85; log VIF = 6.15). Once beds are partialled out, footprint does not account for walking language. The long axis of the largest block, an exploratory quantity, still ranked with walking language after that partial (ρ = 0.678; partial ρ given beds = 0.663). It was not the declared OSM primary, it was selected among several OSM metrics, and it is not a traced corridor.

The present estimates sit inside hospital wayfinding. They also sit inside a wider question about how people move through, and write about, large public complexes in cities. Multi-agent simulations of urban safe liveability treat resident decisions, environmental conditions and policy interventions as interacting over time [54]. That work is not a model of a clinic corridor, and this paper is not a simulation of a district. It is a reminder that a hospital campus is one of the places where urban liveability is used on foot, and that an unsolicited sentence about being lost or walking a long way is a resident-side trace rather than a complete environmental score. A later study could join the text shares used here with a simulation or with observed indoor agents; that combination was not attempted in this file. The calendar mismatch in the retrieved pages is one reason such a joining would need a common time window that this harvest does not supply.

Implications for Design and Wayfinding Audits

The text categories may help identify questions for subsequent post-occupancy and signage audits, but they are not themselves measures of plan connectivity, corridor length, signage legibility or observed navigation performance. No hospital-level design diagnosis is inferred solely from the retrieved texts. A later audit could treat lost language as a question about entrance decisions and landmarks; block-or-floor language as a question about the visibility of storey and pavilion identity; signage-failure language as a question about information continuity at decision nodes; asking as a question about where staff intervention is required; long walking as a question about entrance–clinic and clinic–imaging distances; and mobility language as a question about vertical circulation and accessible routes. Those pairings are a hypothesis list for fieldwork, not a completed audit of any of the 26 campuses. A future research direction is to integrate text-derived spatial-experience indicators with measured circulation networks and agent-based simulation of visitor movements [54]. Such integration could test potential intervention scenarios. No multi-agent simulation was performed in the present study, and urban safe-livability models are not identical with clinical indoor wayfinding. A further description, not used here, stores a building as named parameters. The Industry Foundation Classes and IndoorGML record storeys, rooms, doors and circulation as entities that can be queried. Navigation-related language could in principle be joined to those parameters. No such model was available for these 26 campuses, and none was built. The codes in this paper remain labels on public texts.

Şikayetvar is a contrast corpus, not a second draw from the same frame, and its labels were not author-adjudicated. It did not show the same positive walking association, so the portability of the Yandex pattern remains unresolved. A near-zero or reversed association is part of the result. A complaint site is already selected for dissatisfaction.

The writers are people who posted, and silent visitors are absent. Each Yandex page was the platform’s default order; hospitals do not share a calendar window; 174 texts are undated; and dated reviews precede the July 2026 capacity figures. Two independent coders marked a masked 140-text sample with almost-perfect union agreement (Table 14); the primary tests still use the author-adjudicated file, and walking language is the mark on which those authors were more liberal than the panel. No plan was traced, so the shares cannot be read as a route length or a signage score. OpenStreetMap footprint is collinear with announced beds. The codebook and the 15-text rule were not locked in a dated preregistered protocol. Limits are collected in Section 7.

7. Conclusions

This cross-hospital comparison provides a descriptive assessment of how announced hospital capacity relates to navigation-related language and walking-or-mobility language in retrieved public texts. The complete Yandex corpus comprised 566 texts, kept apart from 670 Şikayetvar complaints; author-adjudicated labels marked 35 navigation-related texts and 27 walking-or-mobility texts. The primary tests used 15 hospitals and 486 of those texts (32 and 26 marked). The two associations were positive (ρ ≈ 0.51) but not statistically supported after correction for the primary tests (Holm p = 0.1040). The walking coefficient weakened when Bilkent and Etlik were removed, through both leverage and reduced power, and the two correlations were not shown to differ. The interior-architecture contribution is this directory-wide language test and a codebook that can later be paired with unmeasured layout and signage audits; reported experience is not treated as the performance of spatial configurations or of signage systems. Split into parts, asking rather than signage-failure language sits closer to the navigation union; that split is exploratory, rests on nine texts, and does not rescue H1. Two independent coders marked 140 masked texts with κ = 0.913 (navigation-related union) and κ = 0.921 (walking-or-mobility union). Substituting their consensus does not rescue H1 and is not used as a new primary test. The results should be used to motivate targeted spatial and visual communication audits, not to rank hospitals or infer causal effects. Limits of this research: This design is a directory-wide public-text comparison, not a case-study post-occupancy evaluation of selected buildings. Author coding of the 566 texts was not independent dual coding and was not blinded to hospital identity or capacity; independent agreement is the 140-text panel. A stored note named 43 mixed cells without identifiers. Re-parsing identifies 42, of which the dated-stamp exclusion removes 20; dropping all 42 changes the hospital set and does not replace H1 or H2 (Section 5.1; Supplementary S13). The cells were not split into new labelled reviews. Hospitals do not share a review window; 174 texts are undated; dated reviews precede the July 2026 capacity figures; a common 2025 window leaves only nine hospitals at the 15-text rule and is not a new primary test. Photographs, as-built plans, space-syntax graphs and block-to-block connection diagrams were not collected. Staff headcount, occupied-bed census and inpatient/outpatient mix were not in the analytic file; announced beds remain the scale variable. Design and construction phasing beyond the opening year in Supplementary S25 was not reconstructed. Indoor and outdoor healing spaces were not inventoried. OpenStreetMap supplies campus footprint, block count, site area and the long axis of the largest mapped block (Table 13). Those axes run from 205 m to 656 m, so the readable layout is a horizontally extended campus rather than a tower. Figure 3 shows three of those plans at one scale. They are ground projections, not as-built floor plans, not sections and not footfall. A usable storey tag in the primary set exists only at Bilkent (area-weighted mean 6.95 storeys on a footprint of 147,739 m²), which is low- to mid-rise on a long plan. The other primary campuses are not ranked by height. The Yandex pages are a default-order retrieval, not a probability sample of recent visitors. Ethics-committee approval was not required, because the corpus is already public unsolicited text, no person was recruited or contacted, and results are hospital-level aggregates without identifiable quotations (KVKK Law No. 6698, Art. 5(2)(d) and Art. 28(1)(ç)). Later comparative work would need route-based spatial measures together with an audit of signs, maps, naming, typography and directional continuity, and it could be joined to simulation of how people use large urban complexes [54].