Add a source without immediately scheduling it
Configure a feed, listing or snapshot, then test it before activation.
- Choose domains, polling interval and source-specific adapter settings.
- New sources start paused; run the discovery/retrieval/extraction canary before resuming.
- Pause, edit or run an individual source from the registry.
Follow a collection through its actual work
Operations shows the latest batch, produced observations and named follow-up actions.
- Inspect selected, completed and failed sources, recent runs and item backlog.
- Resume using saved item states and page cursors; a failed document need not erase earlier accepted work.
- Request a stop or rerun explicitly after investigating a failure.
Normalize different kinds of evidence
Articles, notices and PDFs become structured observations; periodic indicators keep a separate artifact pipeline.
- Retain source, subject, time, extracted measures and supporting excerpts.
- Retrieve web text or PDFs and use OCR when required.
- Keep normalized observations and indicators available for downstream exports.
Validate each observation independently
Model output is a proposal, checked locally before it becomes usable data.
- Check structure, taxonomy, dates, quantities and evidence grounding.
- Discard an unsupported observation while retaining independently valid observations from the same response.
- Keep extraction attempts and source provenance available for investigation.
Correct a finding without rewriting its extraction
Open the evidence, record a quality decision, and correct only fields supported by the source.
- Mark supported, unusable or needing correction; reopen a decision when necessary.
- Record notes, reviewer and append-only decision/correction history.
- Saving corrected fields does not automatically mark the observation supported.
Prepare inputs for forecasting research
The output is normalized, traceable data for demand and lead-time modelling.
- CLI exports expose stored observations and event summaries without scraping them again.
- Structured indicator preparation and persistence are implemented, but have no dedicated control-room table screen.
- How these inputs enter forecasting models remains undecided; no measured forecast improvement is claimed.
Who does what
Watch mode collects due sources after configuration. Users onboard and test sources, resolve failures and make quality decisions; forecasting integration remains research work.