FAI Data: Methodology & Transparency

A Fatal Accident Inquiry (FAI) is the Scottish equivalent of an inquest in England and Wales, held before a sheriff under the Inquiries into Fatal Accidents and Sudden Deaths etc. (Scotland) Act 2016. An FAI is mandatory for deaths in legal custody and deaths arising from an accident at work, and discretionary (at the instance of the Lord Advocate) for other deaths where it appears to be in the public interest.

The sheriff's determination sets out findings of fact under section 26(2) (including the time, place and cause of death) and may, under section 26(1)(b), make recommendations directed at preventing future deaths in similar circumstances. Recommendations are discretionary: many determinations make none.

FAI determinations have been subject to a statutory publication duty since 15 June 2017. Determinations from before that date were published at the court's discretion, so coverage of the pre-2017 period is inherently less complete than the period since.

Determinations are discovered via the SCTS public search index: the JSON API behind the determinations search widget on scotcourts.gov.uk (api.pa.web.scotcourts.gov.uk/web/search, indexType=FatalAccidentInquiryDeterminations). Each index entry carries the PDF link, court reference, sheriffdom, sheriff, and a response-status hint with a link to the case's landing page. An RSS feed (200-item rolling window) is kept as a cheap incremental cross-check between full index walks.

The local index currently carries 910 entries spanning 2006 to the present: both the pre-2017 discretionary-publication era and the full post-15-June-2017 mandatory-publication era. We do not currently separate the two eras in the index count, so this figure should not be read as “all determinations subject to the publication duty”. It is the full reachable index, older and newer material combined. This is distinct from reports of a "300+ outstanding" FAI backlog, which refers to inquiries not yet concluded, not published determinations.

We currently track 910 determinations, of which 245 are positively identified as carrying one or more recommendations. 665 are positively identified as carrying none, and 0 still need classification. Unknown records are not counted as recommendation-bearing.

Daily ingest. A scheduled job walks the search index daily for new determinations, downloads each new PDF, and runs the anchored extraction described below. A full re-walk of the index is run periodically to catch anything the incremental pass could have missed.

Weekly response sweep. A separate weekly job sweeps case landing pages for recommendation-bearing determinations whose response status is not yet known, looking for published responses or section 28(7) non-response notices.

Curated seed. A small curated CSV of determinations with recommendations (originally hand-compiled before the index-based ingest existed) continues to import on every deploy. It is idempotent and does not overwrite data derived from the live index or PDF extraction.

Determination PDFs are text-native (not scanned), and are converted to text with pdftotext -layout (falling back to pdfplumber where unavailable). Extraction is anchored on the statutory skeleton of section 26, not on keyword matching:

• Deceased name, sheriff and sheriffdom are read from the determination's header block.
• The publication date is read from the document's dateline (e.g. "Edinburgh, 9 August 2019"), never from the neutral citation year, which can differ from the year of publication by more than a year for some determinations.
• Cause of death and date/location of death are read from the section 26(2) findings.
• Recommendation text is extracted only from the block immediately following an explicit reference to section 26(1)(b), verbatim and with the original recommendation numbering preserved where the document numbers them.

Determinations that do not follow the standard statutory skeleton closely enough for anchored extraction to complete confidently are assigned the explicit needs_review state and stored with an extraction_flag for a manual or evidence-checked AI follow-up pass. They are never silently treated as either positive or negative, and are never filled in with fabricated text.

We track every determination the index surfaces, not only those with recommendations. Where a sheriff explicitly states that no recommendations are made under section 26(1)(b) (or where a statutorily-structured determination simply has no recommendations section), the record is stored with recommendation_classification = no_recommendations and recommendation_count = 0, clearly marked, rather than being omitted. 665 determinations are currently marked this way. These are hidden by default on the main determinations list behind a toggle, since most visitors are looking for recommendations, but they remain fully counted in the statistics on this site.

Section 28 of the 2016 Act requires each recipient of a recommendation to respond within 8 weeks, either setting out what action has been or will be taken, or explaining why no action is proposed. SCTS publishes the substance of what it receives on the case's landing page: either the response itself, or, under section 28(7), a formal notice that no response was received within the time limit.

We record each response (or non-response notice) as a separate row against the determination, verbatim, with the responding body, response date where stated, and a link to the source document or landing page. A determination's overall response status is derived from these rows, not asserted independently.

The 8-week window shown on each determination page is calculated from the determination's publication date, because the date a recommendation was actually received by each addressee is not published. This is shown as contextual information, not a compliance finding. A closed window without a published response may mean a response exists but has not yet been published, rather than no response having been sent.

We currently hold 59 published responses and 8 formal non-response notices.

Recommendation presence uses three explicit states: with_recommendations, no_recommendations, and needs_review. Deterministic document structure and explicit conclusion wording are used first. Only unresolved cases may be sent to Gemini 3.5 Flash-Lite, using a bounded excerpt rather than the whole corpus.

AI results are accepted automatically only above the configured confidence threshold and only when both the evidence quotation and every separately extracted recommendation occur verbatim in the supplied PDF excerpt. The recommendation count is derived from those grounded items rather than accepted as a model-generated number. Anything else remains needs_review. The method, model, confidence, classifier version, evidence and reason are stored with each decision.

Two separate metadata fields—the sector of the death (health, custody, workplace, or other) and the bodies a recommendation is addressed to—are classified only after recommendation text has been verified. This metadata pass never generates or alters recommendation text.

Determination PDFs are Crown copyright, published by the Scottish Courts and Tribunals Service under terms consistent with the Open Government Licence v3.0. Response documents are authored by the responding body (an NHS board, Police Scotland, a government department, an employer, and so on), not by SCTS. We host and quote them in the same public-interest capacity as Prevention of Future Deaths responses in England and Wales, without asserting a licence over content we did not create.

• 2017–2022 index coverage gap. The search index's incremental RSS cross-check reaches back reliably to around March 2022. Older determinations from the 2017–2022 window are currently reached only via the 52 case landing pages the sitemap lists, which by definition only cover determinations that reached the recommendation or response stage. Filling this gap fully would need either a browser-driven session against the determinations widget or a direct data request to SCTS; it is not yet done, and we are not overstating coverage for this period.
• Non-conforming documents. Some determinations need a manual or evidence-checked AI follow-up pass (see section 4) before their recommendation status is known. These records remain visible as awaiting classification and are excluded from the positive recommendation count until resolved.
• Response sweep coverage is still growing. The response sweep works through the backlog of recommendation-bearing determinations; a "response pending" or "unknown" status may simply mean the sweep has not reached that determination yet, not that no response exists.
• Sector and addressee coverage. These fields depend on the AI classification pass reaching a determination after its recommendation text has been extracted, so recently ingested determinations may show blank sector/addressee fields until the next classification run.

Data is refreshed daily (new determinations) and weekly (response sweep). If you identify an error (a misread name, an incorrectly extracted recommendation, a missed response), please contact us. We aim to investigate and correct verified errors promptly.

This page is updated when the methodology changes materially, including if the 2017–2022 coverage gap described above is closed.