2018 search audit

This audit records the focused historical backfill run on 15 July 2026. It is a method and disposition record, not source evidence. The publication-cohort manifest is the row-level authority for archive enumeration, report-attachment identity, extraction route and literal query-family hits. The screening log is the row-level authority for retained candidate decisions, and the included-report dataset is the authority for extracted report records.

Boundary and cohort

The inclusion date boundary was the date of report, from 1 January to 31 December 2018 inclusive. Discovery began with the Judiciary archive bounded by publication date, because that is the stable date exposed by the archive/API:

https://www.judiciary.uk/prevention-of-future-death-reports/?post_type=pfd&order=asc&after-day=1&after-month=1&after-year=2018&before-day=31&before-month=12&before-year=2018

The archive snapshot comprised 38 result pages and exactly 379 unique PFD publication pages after page-URL deduplication. The original structured JSON cohort index used for extraction had SHA-256 b5ca0aaa204254f9cc4376cfffdd990d31f25fb3187177fcd90d607f2166a508. The exact working index and downloaded report set were recovered locally without a new web crawl; their audit fields were normalised into the canonical 379-row CSV manifest, SHA-256 05df998c880e1f22148af80c9e7ffc8ef449dfb0ca68360fe94fc663e3f3a47a. The fingerprints are integrity aids, not substitutes for the official pages.

For each publication page, the first report attachment was separated from recipient-response attachments. The canonical manifest records its official URL, SHA-256 of the downloaded bytes, byte count, extraction route, success state and extracted-character count. Of the 379 report PDFs:

OCR success means that every scanned PDF produced machine-readable text for search and context screening. It does not mean that OCR is error-free; inclusion and exclusion decisions were checked against the rendered report and official page metadata.

Official search transcript

The official WordPress PFD API was queried with the same 2018 publication-date bounds. Every literal query below returned zero results, including several queries for which the downloaded report PDFs contained responsive text. This demonstrates that the 2018 attachments were not reliably full-text indexed by that search surface; zero is not evidence that no relevant report existed.

Common suffix in every call: after=2018-01-01T00%3A00%3A00&before=2019-01-01T00%3A00%3A00&per_page=1&_fields=id%2Clink%2Ctitle%2Cdate.

Search of extracted report text

Case-insensitive literal variants were searched across the 379 extracted/OCR report texts. Counts below are unique report documents with at least one family hit, not term occurrences and not inclusion counts. The manifest preserves the applicable Q18- family IDs for every publication page; expansion searches are reported in aggregate only.

High-frequency terms such as referral, handover, alert and screening were context-screened; a lexical hit alone did not establish an identifiable digital mechanism.

Reproducibility boundary

The manifest makes archive enumeration, report-attachment selection, extraction state and base-family hit mapping reproducible at row level. It contains 121 successful direct-text routes and 258 successful OCR routes; 110 reports have at least one recorded base-family hit and 269 do not. Thirty-six cohort rows link to retained Q-2018-AUDIT screening decisions (23 inclusions and 13 exclusions). The later-migrated 2018-0405 decision is in the screening log but, correctly, not in this publication-cohort manifest.

The working review did not retain a contextual disposition for every one of the other 343 publication pages. A blank screening linkage therefore means only that no retained candidate decision is present in the screening log; it must not be interpreted as a documented exclusion. The preserved material supports a complete enumeration and extraction claim for the 2018 publication cohort, but not a complete row-level screening-disposition audit. Closing that distinction requires a reasoned disposition for every manifest row.

Report-date limitation and later migrations

Publication date is not report date. Five mechanism-relevant pages in the 2018 publication cohort had reports dated in 2017 and were excluded from this focused backfill: Peter Cotter (2017-0388), Joseph Dune/Joseph Dunne (2017-0371), Brian Stannard (2017-0394), Sarah Kiff (2017-0407) and Mark Doyle (2017-0375). Their screening rows preserve the reason.

A supplementary API search for Ref: 2018 returned 393 pages representing 378 distinct reference strings; 118 pages were not present in the publication-year cohort. These are later-created or migrated pages and show that a publication-year archive cannot exhaust a report-date cohort. The supplementary set was not subjected to the same complete 682-page OCR pipeline. One clearly eligible later-migrated report, 2018-0405, was fully checked and included. Therefore this is a defensible backfill, not a claim that every report dated 2018 has been exhaustively re-screened across all later migrations.

Dispositions

The retained decision set comprises:

• 24 included reports dated in 2018;
• eight reports dated in 2018 excluded after full-report review because no eligible clinical information-system mechanism was established; and
• five digitally relevant reports dated in 2017 excluded solely because they fall outside this report-date backfill.

Karl Willis (2018-0256) is included because the coroner's concerns expressly concern an online prescribing questionnaire, unchecked self-certification and the patient's ability to prevent GP notification. Mavis Reeves (2018-0035) is included as a deliberately bounded telecare/health-IT design case: the report expressly contrasts analogue and digital Careline architecture, data transmission, line contention and emergency access. It is not coded as an EHR case, and the natural-causes conclusion does not establish that the telecare mechanism contributed to the death.

Four official pages show no PFD reference. They use transparent project identifiers, never represented as official references:

The Cuthbert Hingert page displays 2018-0280, while its report attachment filename displays 2018-0279. The dataset uses the page reference and retains the discrepancy as an open metadata recheck.

The screening log contains every retained decision and precise reason. The included-report dataset records the report-level mechanism, causal caveat, controlled facets and registered primary source. Recipient responses for these backfilled records were not assessed in this pass.