Introduction

From its nineteenth-century emergence to its contemporary reinventions, the detective has become one of the most recognizable figures of popular narrative. Yet literary history insists that this figure is not born “fully central.” Early French crime narratives often distribute attention across families, victims, suspects, and social milieus, while the investigator can remain one function among others in a broader melodramatic or feuilletonesque machinery (Tourteau 1970). In the canonical account of the naissance du roman policier, the decisive novelty is precisely the progressive installation of an investigating intelligence as the narrative operator that organizes uncertainty, evidence, and resolution: the detective as a “reasoning machine” (Symons 1993) whose deductions become the true spectacle of the plot (Dubois 1985). Recent work in French genre theory has similarly emphasized the detective's gradual stabilization as an archetype rather than a fixed profession, acquiring coherence through recurrent textual cues and narrative functions rather than through institutional status alone (Lavergne 2009; Barré et al. 2025).

This article proposes a simple but scalable way to operationalize this historical claim. We build directly on Barré et al.'s computational “detective detector,” trained to identify detective characters in French fiction at scale, and applied to a corpus of approximately 3,000 novels spanning 150 years (Barré et al. 2025).Detector performance (reported here for context): the best-performing model (CamemBERT + SVM) reaches balanced accuracy = 0.923 on a labeled evaluation set (support: non-detective n = 419, detective n = 185), with precision/recall/F1 of 0.866/0.908/0.887 for the detective class. Instead of treating “the detective novel” as a pre-given genre label, we treat detectivehood as a character-level property that can be detected across generic contexts, then ask a structural question: when and how does the detective become central to the narrative system? To answer it, we reconstruct character interaction networks for each novel and track the evolution of the detective's network centrality over time. Our core hypothesis is that the “birth” of the genre can be read, in structural terms, as the consolidation of an archetype: an investigator who increasingly occupies a position of structural centrality in the narrative system.

Character Networks

We model each novel as a weighted character co-occurrence network derived from automatic character extraction and coreference resolution (for a comprehensive survey of character network methods, see Labatut and Bost 2019). Amalvy et al. (2025) show that NER errors strongly affect character detection while coreference errors disproportionately affect edge weights, with substantial variability across novels; restricting the graph to the top-N characters mitigates this by focusing on entities with sufficient mention frequency for stable coreference chains, and PageRank's reliance on the relational core (rather than absolute mention counts) further dampens local extraction noise. Starting from BookNLP-fr (Mélanie-Becquet et al. 2024; Bourgois and Poibeau 2025) outputs, we aggregate mentions into character entities (coreference chains), rank characters by mention frequency, and restrict the graph to the top-N characters (here N = 10) for speed, interpretability, and noise reduction in automatic extraction. Two characters are connected when they co-occur within a textual unit (sentence-window co-presence), and edge weights accumulate with repeated co-occurrence across the narrative, producing a compact representation of the story's interaction structure. This approach does not assume that every co-occurrence is a direct dialogue or “true” social tie; rather, it treats repeated co-presence as a robust proxy for narrative coupling at scale, compatible with noisy extraction and heterogeneous writing styles.

To quantify “centrality”, we compute PageRank on each weighted network (Page et al. 1999). PageRank is particularly suited to the detective figure because it rewards not only frequent proximity but also connective importance: a character becomes central when it is repeatedly positioned in relation to other important characters, functioning as an intermediary through which narrative attention circulates. Co-occurrence graphs are naturally undirected; in practice we use the standard PageRank formulation by treating the weighted adjacency as symmetric (equivalently, as a directed graph with reciprocal edges of identical weight). Let G = (V, E) be the character graph of a novel, with nonnegative edge weights wji and (weighted) out-strength sj = Σk wjk. PageRank defines a stationary distribution PR(·) over characters such that:

where d ∈ (0,1) is a damping factor (we use the standard d = 0.85). In narrative terms, this captures the detective's role as a privileged circulation point for information: the investigator becomes structurally important when the plot repeatedly routes key interactions through them.

Figure 1. Two automatically extracted character networks (top-N characters). Edge thickness reflects co-occurrence strength; the detective is highlighted (red). The detective's structural role evolves from relay to a more central position as the investigative schema consolidates.

Figure 1 contrasts two canonical cases and makes the intuition visible. In Gaboriau's Dossier 113 (1867), Verduret (Lecoq in disguise) is identifiable, but his position remains secondary: with a PageRank of 0.079, he appears more as a relay than as a center. The network is carried by characters who concentrate ordinary narrative interactions (family, money, intermediaries), while the investigator often intervenes at a distance—observing, cross-referencing, returning after the fact. The investigation organizes comprehension, but does not monopolize scenes of co-presence. Conversely, in Simenon's Le fou de Bergerac (1932), Maigret occupies a clearly dominant position: high PageRank (0.23), very high degree, and convergence of the thickest edges toward his node. He is the one who interrogates, cross-references, and circulates information.

Four Phases of Detective Centrality

Our main experiment measures, for each novel, two comparable quantities: (i) the PageRank of the main character, operationalized as the most central node in the network (i.e., argmaxv∈V PR(v)), and (ii) the PageRank of the main detective, defined as the detective-labeled character with the highest PageRank. We then aggregate these values by publication year and smooth them with a rolling window. Concretely, points in Figure 2 are yearly means (across novels published in the same year), and lines are 10-year rolling means computed over these yearly means. Figure 2 reveals four distinct phases in the evolution of detective centrality.

Figure 2. Detective centrality over time (1850–2020). Points represent yearly means; lines are 10-year rolling means over yearly means. The PageRank of the main detective (orange) shows four distinct phases: early decline, rise, plateau, and final convergence toward main-character centrality (blue).

Proto-detective phase (1850–1880). Data remains sparse, but the detective's PageRank appears relatively low and even declining. In novels of this period (including precursors like Gaboriau), the detective, when present, often occupies a secondary role in a narrative network still dominated by feuilletonesque structures: families, sentimental intrigues, revelations of genealogical secrets. The detective is present, but not yet the organizing center of the narrative.

Rise phase (1880–1930). We observe marked and continuous growth of the detective's PageRank, progressively approaching that of the main character. This period aligns with the crystallization of the detective genre in France: works by Leroux (Le Mystère de la chambre jaune, 1907), the Lupin cycles (Leblanc), and the creation of the first specialized collections (Le Masque, 1927). The detective becomes structurally central: no longer appearing episodically, but now connecting the different strata of the narrative.

Classical plateau (1930–1970). The two curves remain parallel but distinct, with the detective's centrality stabilizing below that of the main character. This plateau corresponds to the golden age of French classical detective fiction, dominated by Simenon's Maigret series. The detective has become a stable, recognizable figure—central enough to anchor the narrative, yet still sharing structural prominence with victims, suspects, and social milieus that define the genre's realist ambitions.

Final convergence (1970–2020). This phase unfolds in two sub-movements. From the late 1970s through the 1990s, the two curves briefly diverge: main-character centrality remains high while main-detective centrality lags behind. This 15–20 year gap corresponds to the rise of the néo-polar (Manchette, Daeninckx, Vautrin, Jonquet), which deliberately displaces the professional investigator from the narrative center in favor of marginals, activists, victims, or criminals—inquiry is still present, but it is no longer the schema that organizes the cast. From roughly the year 2000 onward, the curves converge decisively: a second rise brings the detective's PageRank into near-alignment with main-character centrality, a shift we tentatively attribute to the commercial consolidation of serial detective franchises (Vargas, Lemaitre, Thilliez) and to the broader institutionalization of the polar as a detective-centered genre. The widening dispersion of individual points reflects generic heterogeneity (roman noir, thriller, polar régional), but the central tendency is clear: the detective now is, in the vast majority of cases, the main character of the narrative. One methodological caveat is that this late-period pattern may partly reflect corpus composition (e.g., fewer novels per year in some decades, or uneven digitization/availability), which we will test by reporting yearly sample sizes and robustness checks in the full version.

Interpretation: Birth of an Archetype, Birth of a Genre

This quantitative evolution strikingly confirms Lavergne's qualitative hypothesis: the detective novel distinguishes itself from the judicial serial novel through a narrative recentering around the investigator (Lavergne 2009). What our measures add is a precise dating and structural characterization of this process: the transition is not abrupt (no clean break in 1866 with Gaboriau, nor in 1907 with Leroux), but progressive, spanning nearly forty years, and translates into a measurable transformation of the topology of character networks.

This structural approach to genre emergence resonates with Moretti's analysis of Conan Doyle's canonical success (Moretti 2000). Where Moretti identifies clues as the “specific device of exceptional visibility and appeal” that explains why readers selected Doyle over his rivals, our measure captures a complementary phenomenon: not the device that differentiates one author from another, but the structural transformation that characterizes the genre as a whole. Clues explain who wins the market; centrality explains what the genre becomes.

The growing dependence of narration on this figure suggests that mystery resolution increasingly relies on the detective's voice, gaze, and subjectivity. As Vareille formulates it, the detective novel is defined by “the interlocking of a temporal series leading to the discovery of the criminal (the investigation) with another, allowing one to go back in time and narrating the genesis of the crime” (Vareille 1989). This interlocking is only possible if one character—the detective—occupies the position of mediator between these two temporalities, collecting clues in the present to reconstruct the criminal past.

Foucault frames the nineteenth century as a moment in which crime becomes reorganized through an aesthetic rewriting that makes judicial control culturally “acceptable” while projecting delinquency into spectacular narrative worlds (Foucault 2011). The detective story performs a second displacement: the dramatic center shifts from punitive spectacle to investigative process. In parallel, Weber's diagnosis of modern rationalization provides a cultural background for the pleasure offered by detective fiction: the world becomes narratively treatable as a system of clues, causal chains, and methodical verification (Weber 2003). The rising PageRank of the detective is thus not simply a narratological curiosity: it is a measurable trace of a cultural reconfiguration in which the management of uncertainty—through procedures, evidence, and inference—becomes a privileged narrative technology.

Our contribution is to show that such claims can be modeled at scale with a deliberately simple articulation of tools: character-level archetype detection combined with network centrality measures. This yields a robust, inspectable bridge between literary theory and corpus-wide evidence: it produces a long-term structural signal while remaining anchored in interpretable local objects (individual character networks). To support reuse and scrutiny, we release code and derived data in a public repository: https://github.com/lattice-8094/DETECTIVES. More broadly, our results suggest that “genre emergence” can be operationalized not only as a shift in vocabulary or topics, but as the gradual stabilization of formal operators—whether at the level of narrative technique, like Moretti's clues, or at the level of character architecture, like the detective's centrality—whose measurable evolution reorganizes how stories distribute information, attention, and explanation.