Digital Humanities methods enable the exploration and exploitation of digitized corpora at unprecedented scales. They also allow for refined research at several levels of granularity, from syntactic or hermeneutic perspectives, or through the identification of geographical named-entities, which allows us to observe the evolution of language and its territorial distribution. However, there are notable limitations in the performance of Named Entities Recognition tools for humanities research due to the variability of the input data (linguistic, diachronic, diatopic variability). Moreover, this lack of robustness to variation is particularly striking when dealing with literary corpora, even more so when it involves early modern texts. The correct recognition of named entities is correlated with the training of the language model implemented in the NER system. Language models are usually trained on so called “clean data” – assembled under optimal laboratory conditions – and for application to a specific corpus, which thus limits their generalizability to other corpora. Moreover, language models for early modern texts often require access to large corpora which have previously been transcribed using OCR. The quality of these transcriptions remains the subject of many current research projects[Baledent et al., 2020]. In essence, the malfunctioning of NER tools is attributed, on the one hand, to the level of quality of the transcriptions provided as input and, on the other hand, to the fact that the corpus being processed does not correspond to the corpus on which the language model was trained. To overcome the problem related to the quality of OCR transcripts, users implement a strategy that is costly both in terms of time and financially, consisting of cleaning of the transcribed text. Indeed, any number of errors can exist in OCR transcriptions[Stanislawek et al., 2019] and this search for perfection, though perhaps feasible on very small corpus, can be never-ending and represents a considerable expenditure of time at larger scales. Our project seeks to evaluate out-of-the-box NER tools, specifically Spacy, on minimally-corrected OCR transcriptions. This experiment should allow us to see the capacity of these tools to do their work outside of ideal laboratory conditions, aiming to get closer to a more everyday use of these tools, i.e. as a user who has neither the time, nor money for corrections, but nevertheless seeks actionable results. By way of this tension between ideality and reality, we have eschewed for the moment any ground-truth, which are costly to produce. Nevertheless, we use what we consider to be a reference text. The reference texts are extracted from ELTeC, a multilingual European Literary Text Collection in which entire novels are available in standardized version. The texts we use in hypothesis-testing consist of the OCR transcription of the same texts, downloaded in PDF format from the Gallica website. The first novel on which we focus is Marguerite Audoux’s Marie-Claire (1910), a novel of about 34,500 words. We carried out initial tests on short text extracts of about words and found that the pre-trained Spacy models are capable of recognising a number of terms even when roughly transcribed by the OCR tool. The ”fr core news sm” model finds 79% of entities present in both the reference and the hypothesis text, and 12.5% of entities which are incorrectly spelled in the hypothesis text.