Event evidence — be98c0e04c005aaa
This is a PHOENIX research candidate.
Whether or not this row reflects a real fire on the ground is determined by
the reconciler over the next 72 hours via independent comparators
(Sentinel-2 burn scar, Sentinel-1 SAR change, VVF report, ANSA news). Until
that outcome is set, treat this event as unverified. PHOENIX is not
an operational alert system; this page is for technical review and audit.
Event metadata
| Cluster centroid | 36.9729, 14.8150 |
| First timestamp | 2026-06-10T23:25:56.430707+00:00 |
| Representative source | wind_diff |
| PHOENIX-led? | yes |
| Verification tier | T0 |
| Biome class | crop |
| WUI class | N |
| WUI built % | 0.0 |
| Graded at | 2026-06-11T04:35:48.803275+00:00 |
| Grader version | v2 |
Model scores (all signals shown — none aggregated away)
| Transformer (production) | not scored |
| Ensemble raw (v3+v4+v5) | not scored |
| Ensemble calibrated | not scored |
| Universal fusion (all DGX data sources) | 0.0 |
How to read these scores.
Each model is one of three trained heads from the 2026-05 bake-off
(methodology). The raw ensemble is the simple
mean of the three sigmoid outputs. The calibrated ensemble applies an
isotonic regression fit on labelled training data
(fitted 2026-06-05T10:30:32.372485+00:00, n=800 samples). A
calibrated score of 0.80 means: of all events that historically received this
score, ~80% turned out to be confirmed fires in our training distribution.
Out-of-distribution events may calibrate worse.
N-series shadow signals
| shadow_tier (N1) | T0 |
| shadow_families | phoenix_thermal |
| shadow_refutation (N3) | — |
| sole_reporter_alert | — |
Corroborator presence (computed at grading time)
| has_vigili_fuoco (VVF) | no |
| has_news (ANSA / Italian RSS) | no |
| has_burn_scar (Sentinel-2 dNBR) | no |
| has_sar_change (Sentinel-1) | no |
| has_lst_anomaly (MOD11) | no |
External-fires rows within ±72h / ±7 km of cluster centroid
| Source | Lat, Lng | Timestamp | FRP (W) |
|---|
| No external_fires rows within ±72h / ±7 km of cluster centroid. |
If this section is empty AND t72h_outcome
above is confirmed_*, it means the reconciler found corroboration
via a path the simple spatial-temporal join above doesn't surface (typically
burn-scar dNBR or SAR change at a slightly different timestamp). Check the
t72h_outcome_evidence field below for the exact corroboration trail.
FP-catalog manual adjudication
Confirmed agricultural burn (manual review)
Catalog cell at (36.95, 14.79) - 3.37 km from event cluster centroid. Wave v1.2.1, from manual_review_agricultural_burns_2026_05_29.json.
Evidence: Rural agricultural landscape with mixed crop fields and a small settlement, showing typical Mediterranean farming patterns with no industrial heat sources visible.
Race / lead-time context
| Race-strict (capable comparator overhead) | — |
| Lead vs comparator sensed time (min) | — |
| Lead vs comparator reported time (min) | — |
| Worst capable lead (min) | — |
| Below comparator floor? | 0 |
| Comparator class | — |
| Capable comparator count | 0 |
| PHOENIX had coverage? | — |
| Lead likely geometric? | — |
| Delivery advantage only? | — |
| Race note | — |
Outcome chain (t72h / t14d / t45d reconciliation)
| t72h outcome | — |
| t72h evidence | — |
| t14d outcome | — |
| t14d evidence | — |
| t45d outcome | — |
| t45d evidence | — |
| refute_strength | — |
What "defensible" means here.
Every column shown above is queried live from
event_grades on the
ground-truth SQLite — no caching, no derived smoothing. The model scores have
their training-time methodology fully documented at
/methodology. The reconciler logic that sets the
t72h_outcome is in the public PHOENIX repo. If you find a value
on this page that doesn't match independent recomputation, please file at
[email protected] — we will publish a retraction within 48 h.