The ranked union of anomalies, changepoints, emerging clusters and forecast breaches. Every explanation is generated from the evidence underneath it — no number appears in the sentence that is not in the evidence.

Terms are TF-IDF weighted against the whole filtered corpus, so a term that is common everywhere ranks below one that is distinctive here. Terms shorter than three characters and stopwords are excluded. Click a sample to read the full report text.

Rule-based extraction of UK postcodes, street names, organisations, dates and quantities from report text. Nothing leaves the API — no external NLP service.

Expected report volume for planning — a council can staff a crew, order a skip or schedule work from it. The shaded band is the 80% interval from residual variance; MAPE is the holdout error over the last horizon days. This is not a prediction about people or places.

Reports without a city name are grouped as “Unknown” and always sorted last. Click a row to filter the dashboard to that city.

Alerts compare the last 7 days with the per-7-day average of the baseline window, so the two sides are like for like. Only groups with at least the minimum number of recent reports and a ratio of 1.5× or more appear. The date range does not apply here — city and category filters do.

Recent counts are compared with the same area × category's own trailing baseline. The method column is the honest part: Poisson is the tail test for rare counts (baseline rate < 5 per window), robust z is median/absolute-deviation for normal counts, so a single blip cannot look like a surge at low volume.

The best single split (SSE minimisation) on each daily series — “when did this structurally change?”, not “is this a spike?”. Values are daily means either side of the split date.

Lift is how much more often a pair appears together than chance. External correlations are Pearson against the cached open-data series; a direction is only shown when it is positive or negative — a coefficient that rounds to zero is omitted by the API rather than called “none”.

EXIF note —

Method: min–max normalised per metric within the current filter, weighted, then scaled 0–100. Expand a row to see each component's raw value, its normalised value and its weighted contribution — and whether they add up to the score.

Each component is dropped in turn and the ranking recomputed. If the top area changes, the ranking depends on that component and is reported as decisive.

A metric missing for half the areas must look like that. Missing is never scored as zero; scores that lean on a partial metric are marked here.

Every task here was confirmed by a person and recorded in the audit trail. The agent can draft a work order, but it can never create one.

Pyodide and matplotlib are fetched on demand from the URL configured in config.js. If that fails — offline or blocked by CSP — the panel says so and the rest of the dashboard is unaffected. This is the correct sandbox for a hosted product: there is no server-side interpreter to escape from.

Curated open datasets, with the licence they are published under. Import runs the same ingestion pipeline as any other upload — it is not instant. Preview shows how the importer typed and interpreted each column. Enable/Disable keeps a dataset registered but excludes it from the agent's tool results; it never deletes the data.

Sources, in order: the curated catalogue, then data.gov.uk’s CKAN API (keyless). “Check before importing” posts a proposal and persists nothing; the import itself runs only when a person confirms it.

Columns

Sample rows

—

GeoJSON imports become drawable layers. “Draw on map” adds one to the Map view as a real Leaflet layer — it composes with the heat grid, hotspots and clusters, and the swatch beside each name is its map legend. Dense layers (building footprints) are capped when drawn so the browser stays responsive; the map says when that happens.

Generates synthetic UK reports containing discoverable patterns — a persistent hotspot, an emerging one, seasonality, weather linkage and a one-off burst. Demo rows are flagged is_demo in the database and never appear in the consumer app. Seeding is an administrator action; the generated patterns are listed when it finishes.

Observations are fetched on demand and cached locally, so queries never depend on a third party being up. Refresh fetches the cities in scope; concurrent refreshes are refused by an advisory lock.

Counts are live for the current filter. The graph is what makes “what is connected to what” answerable — reports to areas, categories, themes, mentions, media and external observations.

Small-cell suppression: any area × category × week cell with fewer than minCell reports is withheld and counted as suppressed, so an individual report cannot be identified from the published totals. A higher minCell publishes less but protects more.

—