Contract Intelligence for the LIBOR Transition

When an interest-rate benchmark like LIBOR is discontinued, every contract that references it has to be found, read, and - where there is no robust fallback - amended before the benchmark ceases. A large institution can hold tens of thousands of such contracts spread across decades of paperwork: clean digital PDFs, Word documents, and scanned images of agreements signed long before anyone planned for a transition.

Reading them by hand is slow, expensive and inconsistent. The hard parts are locating the rate clause and any fallback language, judging how confident the read actually is - especially on a degraded scan - and prioritising remediation by exposure, all while keeping a defensible audit trail. Getting the confidence wrong in either direction is costly: a missed clause is a compliance risk, and a falsely-confident read on a noisy scan is worse than admitting uncertainty.

The Transition Desk ingests a contract in any of the three formats and reads its clauses - parsing digital text directly, or running a metered vision stage that judges how legibly a scan reads. It extracts the benchmark, the presence of fallback language, the relevant dates, the parties and an amendment indicator, and scores how confident it is. High-confidence reads are auto-accepted; medium ones go to a review queue where a human overrides the AI; low ones go to an exception queue for manual handling.

Everything validated rolls up into an amendment register, ranked by benchmark × materiality, ready to export to a remediation workflow. A deterministic clause-text benchmark and fallback detector, plus a confidence floor, are authoritative - the vision read is a legibility signal, not a substitute for the text analysis.

This is a faithful, downscaled reimplementation on a synthetic contract corpus - never the production system, and the client is not named. It proves a genuine dual-perception pipeline: a vision OCR legibility read on scanned pages and a clause text-analysis on every document, both metered, with a deterministic validation floor and a three-band triage on top. It deliberately simplifies real document ingestion, a production OCR engine, ticketing and amendment-export integrations, and model retraining - all labelled in Architecture → Out of scope. No real or copyrighted documents are used.