The platform's spine is an ingestion pipeline. News from LexisNexis and Opoint flows through AWS (Lambda, SQS, S3) into one normalized model, indexed into OpenSearch across 49 languages. Every source carries a provider-neutral editorial rank — a 1–5 source-authority scale, supplied natively by one provider and mapped from the other's own ranking — so users filter by source quality without ever caring which provider an article came from.
On top of the pipeline sits the triage loop. Scheduled monitoring scores every newly ingested article and parliamentary record against each organization's Issues and Stakeholders; candidates above a per-domain threshold land as pending references in the AI Inbox, and below-threshold noise never reaches a human. Editors accept or decline each reference — an accepted reference is Selection membership, immediately scoping the watchlist and feeding the weekly reports, daily digests and board-ready briefings. One workflow, two domains: media and Swiss politics.
The political side is not hand-modelled. Stage chains for each affair type are derived from the Swiss Parliament's open-data corpus — 221,878 affairs, 57,574 labelled events — by measuring the event walks affairs actually take, stored versioned and re-derived monthly with a drift report. The first design used four hand-authored process templates; measuring the corpus replaced them.
The upstream API overwrites records in place, so “what changed” cannot be read from the current document. The platform snapshots each tracked affair and diffs on stable ids and content hashes, ranking every change by severity to order the watchlist email — and reports are denormalized at generation, so a past report never mutates. The newest capability, now rolling out, forecasts legislative timelines and outcomes from measured stage-transition durations and probabilities — including the measured discovery that median durations don't add up, so cumulative estimates are computed separately rather than summed.
