SENTINEL

Overview

SENTINEL was built by a Maynooth University team for the UN Global Hackathon on AI for Digital Trade Regulatory Analysis, where it placed in the Top 20. It takes an economy and a regulatory topic, finds the governing legislation on official government portals, and maps individual provisions to RDTII 2.1 indicators — each with an article-level citation and a verbatim quote.

The constraint that shaped the design

The output is legal evidence, so a plausible-sounding paraphrase is worse than no answer. That ruled out handing the whole task to a language model. The engine is a deterministic and probabilistic hybrid instead:

• Regex slices the statute into provisions.
• A keyword classifier routes each chunk to candidate indicators.
• The LLM extracts the feature and selects a quote.
• The quote is verified character-for-character against the source text.
• A deterministic Python scorer — not the model — assigns the indicator score.

Any mapping whose quote cannot be located verbatim in the source is rejected rather than displayed. Verification gates sit on top: a blind cross-model checker reviews positive scores before anything ships, and scope guards can only ever lower an out-of-scope score, never raise one.

Discovery

Evidence collection has to survive real government websites. The crawler works in layers to handle session-gated portals, retrieves scanned legislation through OCR, and localises its query terms per language — six of the seven Round 2 economies are non-English, so retrieval and query localisation, rather than fetching, are where accuracy is actually won or lost.

The audit view

A reviewer should not have to take a score on trust, so the workbench shows the whole derivation on one panel: the indicator, the law and its amendment date, the raw statutory text, the model's stated reason, the resulting score against its allowed range, and the individual scoring inputs that produced it. Token usage and cost sit at the top of the same card. A retrieval-backed chat runs alongside it for interrogating the underlying corpus.

Reproducibility

The graded artifact is not a folder of curated rows — it is a tool a reviewer runs to produce those rows. Runs are reproducible from an integrity-verified offline corpus guarded by checksums, character counts, and a statute-quality classifier, so a truncated or substituted law cannot quietly poison a submission. A full run emits a 1,246-row evidence submission covering 61 indicators across 12 pillars for 7 economies, with 96.2% of quotes located verbatim in their source. The key-free local-model path means a reviewer needs no API credentials to reproduce it — a whole submission costs about $13 on a hosted model, or nothing on the open-weight path.

Notes

• Honest negative results are recorded rather than buried; one escalation was measured, found not to help, and reverted.
• Coverage tiers are stated plainly — economies where live crawling is verified are distinguished from those verified against a local corpus.