Incident Investigation Assurance

UI design artifacts for an AI assurance / QA layer over mining incident investigations — "was this incident investigated as thoroughly and correctly as it should have been?" The same controlled, auditable eval machine as the Graph Query Agent, re-pointed at the ICAM rubric and a Snowflake data plane.

Static mockups · no backend / no database required · open any screen below

ICAM rubric deterministic gates + isolated LLM judge per-dimension PASS/FAIL + rationale human-in-the-loop ISO/IEC 25059
The harness · per incident
incident + evidence→ deterministic gates→ grounding (RAG)→ isolated ICAM judge→ scorecard→ FAIL → HSE reviewer
Screens
Scorecard
Per-dimension PASS/FAIL + rationale on the worked example. Grounding + judge contract.
start
ICAM graph
Contributing-factor graph — missing layers show as holes in the Swiss cheese.
no-dep
Programme
Corpus dashboard — FAIL-reason distribution, KPI targets, judge↔expert κ.
Reviewer queue
HSE sign-off — assure / return for rework, with override logging.
Platform
Graph Query Agent
The base platform this assurance layer extends.
What it is. An assurance layer that catches deficient investigations before the regulator does. It evaluates two things it can do well — process rigour / completeness (was ICAM followed across all four contributing-factor layers?) and factual grounding (does every claim trace to real evidence?) — and flags, but does not decide, on true root-cause correctness. Cheap deterministic gates run first; an isolated, locked LLM judge then scores each dimension PASS/FAIL with a written rationale citing the ICAM layer and the evidence. Results are written to a durable, queryable store; any FAIL routes to a qualified HSE reviewer.

Governance. Human-in-the-loop is mandatory — safety-critical and regulated, the harness triages and explains; humans sign off, never autonomous. Designed to be demonstrably structured under ISO/IEC 25059, governed by ISO/IEC 42001, operated with the NIST AI RMF Measure function. Illustrative data only — the worked example (INC-2026-0412) is fictional.