How to interpret the Section 19 report map Introduction Section 19 reports are commissioned by councils in England and Wales after flood events to investigate the causes of flooding and provide recommendations on how to prevent future flooding. The Section 19 report map displays information on the impacts of flooding in individual places in England contained within Section 19 reports. The map has been derived by extracting information on flooding impacts from the reports using a Large Language Model (LLM). Relevant pages of each report were sent to the LLM along with instructions to return information on the impacts in the individual places mentioned, and when these things happened. Coordinates matching place names were then derived used geocoding. It is recommended that users familiarise themselves with the known issues and data quality flags described in this document before using the map. Particularly import points are emphasised in bold. The data underlying the map was produced by Helen Jackson of ClimateNode, and the map UI was produced by Kip Parker, Lead Developer at Icebreaker One, both working on a pro bono basis. We hope users find it helpful and informative in understanding both how flooding has affected communities in England and how it can be mitigated. Points of interest The map contains information on over 10,000 individual locations in England down to street and asset level. When the map first loads, the user will see high-level clusters of points of information. Click on these to zoom in to areas of interest and see pins indicating individual points of interest. Clicking on a pin reveals information on flood events in that location in a sidebar. Some individual pins contain numbers, which indicate the number of flood events mentioned in the sidebar. It is important to note that the coordinates of each place do not indicate the precise location of flood impacts and should not be interpreted as such. They are simply coordinates which indicate the general location of the place mentioned. Most coordinates have been derived from OpenStreetMap. Lots of places in England have the same name – for example, there are thousands of High Streets – and this can pose a challenge to identifying locations within text. Precise place name identification has been subject to rigorous human quality control and coordinates should in the vast majority of cases refer to the correct location. Sources of inaccuracy Most people who have used LLMs are aware that they can fabricate information ('hallucinate'). This possibility should have been reduced by giving the LLM clear instructions on extracting information from the specific textual sources provided to it, rather than using its own internal knowledge. However the possibility remains there may be some hallucinations presented within the information displayed on the map. It is strongly recommended that if users see information on the map which is important to them, for example information about flood events on their own street, or information which they are using for their own research, they do not assume accuracy but refer to the relevant passages in the underlying Section 19 report. Links to the original reports and page numbers are provided to do this. Section 19 reports themselves may contain inaccuracies. Typos around place names and dates are common, and have been corrected where spotted. The word sic is used to acknowledge that an inaccurate place name mentioned in a text differs to the one assumed to be the intended place name. Information extracted from images, tables and lists overlapping pages should be viewed with caution. It was noticed that images depicting flood flow routes in particular were liable to misinterpretation by the AI. Information contained in rows in large, complex tables can be misattributed to place names mentioned on other rows. Another potential source of inaccuracy stems from the way the LLM has been instructed to return information in standardised formats, which may lead to shortcuts and inaccuracies when passages contain complex or nuanced information about multiple events. In paragraphs about multiple flood events it has not always been possible to disentangle information about which impacts relate to which events. For that reason, the same information is sometimes presented repeatedly for all events mentioned in a paragraph, though not all observations will be relevant to all events. Users should treat information from passages about multiple events with caution accordingly. The following data quality flags are used to inform users that a given piece of information may be subject to one of more of these problems: IMG: information may have been extracted from an image TAB: information may have been extracted from a table LST: information may have been extracted from a list overlapping pages MUL: information has been extracted from a passage about multiple events PAG: information may have been extracted from text from a page other than the one indicated (see 'Page numbering' section below) Date information Most flood impacts described in a given Section 19 report will be those which occurred during the event(s) which is/are the subject of the report. However, many reports also provide detailed flood histories, which can go back to previous centuries, and/or frequent references to other recent flood events. The map sidebar presents information on any and all distinct events mentioned, listed under distinct date headings. Information on when flooding occurred contained in text can be very specific, i.e. a precise date, or very vague, for example "winter 2013-14", "Easter 1998", etc., or sometimes just a year or decade. Precise dates or date ranges are given where they are available, but otherwise verbal descriptions of dates are used. The term "ambiguous date" is used when information relates to: flooding in general rather then a precise event a precise event which the LLM has failed to identify flooding which has occurred repeatedly or at some point within a specific but very long time range, for example 2002-2020 Where different date descriptions of the same event have been used in different documents, attempts have been made to harmonise these to avoid repetition. For example, Report A describes flooding which happened during Storm Babet between the 18th and 23rd of October 2023 in Borsetshire. Page 10 of Report B describes the same event, but states it happened in 'October 2023'. Rather than list these as separate events, if there was no other notable flooding in Borsetshire in October 2023, these are assumed to be the same event and listed under the more precise date range '18-23 October 2023'. When users refer to p10 of Report B, they will find a reference to 'October 2023' and not the precise date range, so it is useful to bear in mind that this harmonisation has occurred. Some multiple descriptions of the same event may remain in the dataset. In a number of cases, the LLM returned date information which related to the dates of investigations or remedial actions following the flooding rather than the event itself. Attempts have been made to correct these so they refer to the actual event, but it's possible some may still remain. Page numbering All page numbers treat the cover page as page 1 and therefore stated page numbers may differ from any nominal page numbering within the document (for example, where the start of the body of the report after the cover, acknowledgements, table of contents, etc. is treated as page 1). Page numbers are not always accurate because a small amount of information from preceding and following pages is usually included when pages are sent to the LLM so as to not abruptly cut off sentences and paragraphs. Therefore information about a place or organisation may sometimes be on the preceding or following page to the one stated. If users are looking for a particular piece of information within a report following the link and it is not on the indicated page, they should also check the preceding and following pages. Page numbers are in particular less likely to be accurate where there is a list or table overlapping pages. Such lists and tables are treated as belonging to the page they begin on in order not to break them up. Coverage This map covers as many English Section 19 reports as we were able to collate in September 2025. There are gaps in coverage which future updates may address.