# NUMIDIA EYE — capability catalog > Wildfire intelligence for Algeria, served as an open web surface over the > NUMIDIA API: live NASA FIRMS VIIRS detections on an interactive wilaya map, a > live-v1 structured-data verifier, rule-based incident priority (priority-v1) > and incident reports (report-v1). This file is kept in step with the code. Every capability below is one the application actually implements. ## Key facts - Coverage: detections are clipped to the Algeria bounding box by the backend, not by the front end. The bundled boundary layer carries 48 wilaya polygons. - Counting rule: a **detection** is one raw FIRMS thermal pixel. An **incident** is a deterministic spatiotemporal cluster of detections (`incident-v1`, time-ordered greedy chaining, eps 0.03 deg, temporal gap 24 h) and is **not** a confirmed fire. Each detection joins at most one incident, so the incident count is an upper bound on distinct fires and can merge two nearby fires into one cluster. Incident IDs are `INC-` and are stable across rebuilds. - Wilaya statistics are counted from the detections loaded in the browser, not from a national total, and the page says so. - Priority is `priority-v1`: score = 0.40*frp_norm + 0.25*persist_norm + 0.20*count_norm + 0.15*sat_agree, caps FRP 100 MW (log), persistence 72 h, count 20; tiers LOW<0.25<=MODERATE<0.5<=HIGH<0.75<=CRITICAL. It is deterministic, rule-based, and **never consumes model output**. Factors with no dataset (settlement proximity, critical infrastructure, terrain accessibility, forest context, visual verification) contribute zero and are listed explicitly. - Verification is a real model call: `GET /detections/{id}/ai` runs `predict_proba` on a verified artifact and returns FIRE / NON_FIRE / UNCERTAIN with a probability. When the artifact cannot be verified the API returns 503 with **no probability, no model and no verdict**, and the page then shows the unavailable state rather than substituting a value. - Satellite thermal anomalies are not confirmed fires, and an absence of detections is not evidence of no fire. - Emergency: Protection Civile **14** (or **1021**). Reference information only. This application is never a dispatch, staffing, availability or response-time service. - Language: English only. There is no locale switch. ## Surfaces - `/` — the application. Focus one wilaya with `?wilaya=`. - `README.md` — what this is and how to run it. - `provenance.md` — every source, its authority and its terms. - `index.md` — markdown twin of the page. - `data/wilayas.geojson` — display-only wilaya boundaries. ## Backend endpoints consumed | Endpoint | Use | |---|---| | `GET /system/status` | FIRMS / verifier / database / data-state chips | | `GET /detections?limit=` | map points and counts | | `GET /detections/{id}` | authoritative selected-detection record | | `GET /detections/{id}/ai` | live-v1 verdict, probability, threshold | | `GET /incidents?limit=` | incident clusters and the priority-v1 result | | `GET /incidents/{id}/report` | report-v1 document with limitations | Nothing else. There is no fixture, no generated detection and no fallback data. ## Scope, stated plainly The verifier is an **experimental structured satellite-data model**, not a validated Algerian wildfire detector. Its training labels contain no independently confirmed southern Algerian wildfire positive, and the negative class is predominantly catalogued gas flare. Test recall is 0.81 (1,977 of 10,323 fire detections missed). Southern test metrics are recorded as UNAVAILABLE for insufficient class coverage rather than as zero. Do not read this as national validation, and do not read it as separation of wildfire from industrial flare. ## Licence No site-wide data licence is declared. Application code is MIT. Data terms are stated per layer in `provenance.md`; check that file before reuse.