Skip to content

Operational LLM guide

Read order

  1. methods/index.md and evidence/method.md for the product label, inclusion boundary and synthesis discipline.
  2. evidence/sources.md, icmm/screening_log.csv and icmm/included_cases.csv for provenance, selection and row-level report extraction.
  3. evidence-base/all-reports.md or the relevant generated evidence-base/reports/*.md page to locate a report and its primary Judiciary link.
  4. evidence/claims.csv (or its generated evidence/matrix.md view) when a statement requires a stable CLM- identifier and explicit source/report/code trace.
  5. findings/index.md and the relevant mechanism page for bounded cross-case synthesis.
  6. defence/transferability.md, then defence/requirements.csv and defence/assurance_tests.csv, for Defence application.
  7. evidence/validation-queue.md and evidence/research-log.md before presenting a claim or corpus as settled.
  8. Read outputs/ last: publication material is derivative and must not be used to override the evidence chain.

Authority chain

primary source → report → claim → mechanism → candidate requirement → assurance test

The chain distinguishes authority from translation. Not every included report needs a separate claim row, but a Defence requirement that lacks a direct claim trace must expose that limitation or remain a labelled hypothesis.

Guardrails

  • Preserve the difference between a coroner's finding or concern, a recipient's account, reviewer coding, cross-case synthesis, Defence translation and a hypothesis.
  • Do not generalise from one PFD to prevalence.
  • Do not treat recurrence in the 71-report current series as incidence, prevalence, causal weight, comparative product safety or evidence of a trend.
  • Describe the verified Derby case as Stephen Stringer, PFD 2024-0555; do not attribute the incident supplier because the official materials do not name it.
  • Treat 71 included reports from 87 retained candidates as the current-pass result: complete 2018 publication-cohort enumeration/extraction with 36 retained cohort decisions, targeted 2019–2026 discovery and a limited later-migration check—not an exhaustive interval-wide count or prevalence estimate. The other 343 cohort rows lack retained contextual dispositions and are not documented exclusions.
  • Remember that PFD 2025-0079 covers three deaths; report, case and death counts are not interchangeable.
  • Treat mechanism codes as non-exclusive reviewer coding. An em dash in the map means only that a code was not assigned to that row; it is not an evidence gap or proof that the mechanism was absent.
  • Treat workflow, setting and boundary groups as conservative browsing facets, not coronial classifications.
  • Do not infer that a recipient response was adequate, implemented or effective from its existence.
  • Describe every REQ- item as a candidate analytical translation until Defence validates it, and every AST- item as designed but unexecuted until evidence says otherwise.
  • Date-stamp product behaviour claims; a historical PFD does not establish current functionality.
  • Generated report pages, maps, LLM files and graphs route readers but never add evidence. The 2D/3D Graphify views describe repository structure, not cases or mechanisms.