Drone Watch: strike monitor

Drone Watch

Drone and missile attacks on Ukraine and Russia. Every figure is a claim.

This day

Launched at Ukraine-

Launched at Russia-

Locations reporting strikes-

What these figures do and do not count

Launched at is how many drones and missiles that country's defence ministry says were fired at it. Locations reporting strikes is how many places said something came down on them. They measure different things and will not add up; neither is a casualty count.

A range (140-160) means sources published different numbers for the same day. Both ends are shown rather than one of them picked.

A day's figure is the whole day. Both ministries report in separate windows: one overnight, then daytime updates. A news story usually quotes the overnight window alone, so the number you saw there will often be smaller than the one here. Where a day was added up from several windows, the source line under it says how many.

Shown on the map

War totals

▼

loading…

Whole-war figures as reporting states them, not this site's own tallies added up. Each carries the date it is current to; where no source publishes a total, that is said rather than estimated.

Volume per day

launched at Ukraine ↑ launched at Russia ↓ min-max range strikes reported, no count published event, hover the mark for what happened

Objects

…

Reading

Vertical scale

Key▼

Strike aimed at Ukraine

Strike aimed at Russia / occupied Crimea

Bursts and day totals are coloured by who was hit, so the hue is the target, not the country's own colour, which is why the shading further down, where a colour does mean a country, runs the other way.

Confirmed: 2+ independent outlets

Reported: single outlet

Claimed: belligerent official only

Marks you can switch on

Airspace incident: the war's drone down in a third country

Civil airport shut by the war, closure still standing

Civil airport reopened, or shut after February 2022

Military object: one chip per category, glyph as in the menu

Shadow-fleet vessel reported struck

Object reported destroyed

Every strike, all dates: one dot per reported strike

Reference chips are compiled from public reporting, not surveyed: each popup prints how precisely it is placed. A chip is where the object is, never a claim that it was hit.

The land

National borders: Ukraine, Russia, Georgia

Contested: Russian-claimed oblasts; Abkhazia & S. Ossetia

In Ukraine the hatch is what Russia claims, not what it holds: the claim is the whole oblast and the front runs through it. In Georgia (occupied since 2008) it is what it holds. Borders as recognised.

Country supplies Ukraine

Country supplies Russia

Country supplies both sides

Military aid and military-relevant trade only: weapons, ammunition, troops, dual-use components. Unshaded is neither, not neutral in general. Shade shows direction, never scale. Hover a country for what it sends.

How these numbers are made

Nothing here is measured. Every figure on this page is something somebody published — a defence ministry, a news outlet, an analyst — read out of the text of the report and stored with a link back to it. This is what happens between that report and the mark on the map.

Where the reporting comes from

The spine is the Institute for the Study of War's daily campaign assessment. It is the one source that can be asked for a date rather than scraped out of whatever a feed still happens to be holding, which is the only reason a record years long can be walked at all.

Around it: the registered news feeds, and the Telegram channels that publish nowhere else — both defence ministries, regional administrations, OSINT projects. For attacks on Ukraine there is also the Ukrainian Air Force's own daily record of what was launched and what was shot down, taken from a machine-readable dataset rather than from prose. Where it covers a day it replaces what reading sentences produced, and the text passes are locked out of that day afterwards.

For older days, archived captures of those same feeds are replayed through the same parser, alongside the outlets that still publish a dated index of everything they filed that day.

Every article is stored with its URL and its outlet. Sources in the panel behind this one lists all of them, what each one is, and when each was last read.

How a day's launch figure is decided

Counting happens sentence by sentence, never over a whole article. An assessment names Russia on every other line regardless of who was attacked, so scoring the document as a whole put the day's figure on the wrong country.

A number is recorded only if the noun it modifies names a weapon class. "89 Shahed drones and 4 Kh-101 cruise missiles" is two facts, not a total of 93, and a number whose noun will not classify is dropped rather than assumed to be drones.
Interceptions are matched first, and their words are then closed to the launch patterns — otherwise "shot down 12 drones" also reads as twelve launched, and one number becomes two.
Multi-day and running figures — "over the past week", "since Monday", "so far this month" — contribute nothing. No source says how the week split into nights, so the figure is excluded rather than divided.
The Russian defence ministry publishes a day in consecutive windows: overnight, then daytime updates. A day is the sum of its windows, not the largest of them. The source line under a figure says how many windows it was added from.

A row is a whole calendar day. An overnight attack lands on the morning-after date, which is the convention the Air Force's own daily reports use.

Silence is not a zero. A day nobody published a figure for has no row and is drawn as a gap. A source that positively says none were launched writes a real zero, with its citation. The two are different claims and the record keeps them different.

Why a day sometimes shows a range

When sources disagree about the same day, both ends are kept and shown as 140-160 rather than one of them being picked or the two averaged. On the chart that is the min-max band. Cumulative totals accumulate each bound separately, so the band widens as the total grows: disagreement compounds over a range, it does not cancel out.

Where a day could only be estimated — its share of a period figure, once the days in that period with their own tallies are subtracted — the row is marked as an estimate and says so under the figure. A real number replaces an estimate the moment one is found, and estimates are never quoted in a strike description.

How a place gets a mark on the map

Place names are read from the same sentences and resolved against an offline gazetteer: the cities anyone would name unprompted, plus the smaller settlements the stored reporting actually names, each one resolved against the GeoNames populated-place record. A name earns a pin only when a single settlement can be meant.

A place is recorded only when it lies in the country that sentence says was attacked. That is what keeps a launch site out of the map: "drones launched from the direction of Bryansk" puts a strike verb in a sentence with a real place, and without the country test it became a target.

How sure the record is, per location:

Claimed — a belligerent's own channel, and nothing else.
Reported — one independent outlet.
Confirmed — two editorially independent outlets, same place, same date. ISW alone counts as confirmed by itself: it geolocates and cross-checks against both sides before publishing, and making it wait for a second outlet would throw away verification already done.

A mark sits on the settlement something was reported at, not on a surveyed impact point. When the parser stops standing behind a pin, that pin is retracted rather than left on the map to age.

What the language model is and is not allowed to do

A small model is optional, and it is allowed exactly one job: deciding which places in a sentence were hit, which were launch sites, and which were only mentioned. It never touches a number. The split is deliberate — a pattern cannot tell a target from a launch site, and a model asked for counts will confidently return a figure that is not in the text.

Every verdict it returns is checked before it is believed: the place must resolve in the offline gazetteer, its country must match, the evidence it quotes must be a real substring of the text it was shown, and any weapon or number must be findable in that evidence. A hallucination therefore costs a pin; it cannot invent one.

With no model reachable at all, every pass still runs on the rules alone. The difference is how much gets found, never whether a figure is trustworthy. No sentence on this page was written by a model.

How each strike's one-line description is written

By assembling fixed fragments in a fixed order — drones, missiles, intercepted, hits, fire, casualties, what was hit, what the place is, days in a row — and dropping every slot with no fact behind it. Every word is a literal in the code. Nothing is quoted out of an article and no model writes them, so the same facts always render the same line and it cannot inherit an outlet's framing.

A number is attached to a place only from a sentence about that place alone. A country-wide tally can be shown on a pin, but only labelled as such — "89 drones nationwide" never becomes 89 drones over one town.

How old days are kept level with new ones

Today is deep because today has feeds: a recent day can rest on twenty to fifty separate articles. A day from last October rests on whatever can still be asked for that date. Left alone, the record would say more about how hard each day was looked at than about what happened.

So a timer works old days one at a time, thinnest first, giving each the treatment today gets: that date's index at the outlets that publish one, that date's archived feed captures, the assessment re-read under the current rules, and the day's stored text re-read as a whole so a figure that only appears when two reports are put side by side is caught.

Re-running is safe by construction: every write is keyed on the day and the place, so a pass converges on the same rows instead of duplicating them. The record is rebuilt from stored text, never edited by hand.

What is compiled by hand rather than scraped

Five things, all of them switched off by default and all of them reference rather than event data: the military and industrial objects, the airports the war closed, the airspace incidents in third countries, the shadow-fleet vessels reported struck, and the whole-war headline totals.

They are compiled from open published sources, not surveyed. Each object states how precisely it is placed — the facility itself, or only the town it is in — and a chip on the map is where the object is, never a claim that it was hit. Where an object has strikes reported near it, the popup gives the distance and the date so the reports can be checked, and says nothing stronger.

The war totals are what published reporting says the whole-war figure is, kept beside their citations and the date they are current to. They are deliberately not this site's own rows added up, and the two are never merged.

What these figures are not

Not casualties. Nothing on this page counts people.
Not the front line. Ground control is not tracked here; the hatched areas are what is claimed, not what is held.
Not live status. The airports layer is the closure record of the war, not whether anything is flying tonight.
Not adjudicated. Both sides publish figures about their own air defence. This records who claimed what, on what date, with a link — it does not decide who is right.

Every figure is a claim, and every claim keeps its source. The current day is not shown until it has ended.