TRIAGE

The problem

Everything the operational risk function computes stands on the loss register, and the register is manufactured at the worst possible moment: by a busy first-line analyst, weeks after the event, choosing a Basel category from a dropdown they half remember. Misclassification propagates silently into capital models, scenario calibration, indicator backtesting, and every board trend chart. The most consequential dataset in the discipline gets the least engineering.

The fix has to live at the point of capture. Annual cleanup projects treat symptoms and decay immediately; classification done well at intake compounds through every downstream consumer.

What it does

TRIAGE is a classification workbench built around the event narrative:

Intake. Narrative first: paste what happened and only the narrative is required. Dates, amounts, and reporting unit enrich the record but never block it. Analysts stop fighting the form and start describing the event.

Two classification engines. A deterministic rules engine that works fully offline, for institutions that need reproducibility above all, and an AI mode that reads the narrative and proposes event type, cause, and business line with a confidence score and a written rationale. Either way the human decides; the engines make the decision reviewable.

Register. Classified events land in an auditable register with their rationale attached, so a sampled review or an examiner request is a query rather than an archaeology project.

Governance and batch. Second-line review workflow for the boundary cases, and batch reclassification for the part every bank quietly dreads: the existing backlog, rescored consistently with the same engines and the same audit trail.

The prototype runs against a synthetic mid-tier bank tenant with sample events, and states plainly on screen that all data is synthetic.

Why it matters

Classification quality is invisible right up until something downstream depends on it: a capital model sampling the wrong severity distribution, a scenario anchored on a category that is quietly polluted, a KRI graded against loss months that are mislabeled. The Basel Committee’s Principles for the Sound Management of Operational Risk expect comprehensive, honest capture of operational risk exposure. A register with rationale-carrying classifications is the difference between asserting that and demonstrating it.

Honest framing

TRIAGE is a working prototype and a blueprint, not a product. The deliberate design choices are the point: narrative before form, rationale on every machine judgment, human decision on every record, and an audit trail as the exhaust of normal work. It demonstrates that evidence-grade loss data is a capture-time property, buildable now.