Ukraine — the iteration loop that set the tempo, and the capital that set the ceiling

Filed as interpretation in canon/proposals.md, and interpretation holds. The id sits at line 844, inside the interpretation block that runs from 808 to 988, and unlike most of this canon's recent arrivals the section is right about what the thing is. Almost everything below is fact — dated procurement, dated reports, dated launch counts. The interpretation is the sentence laid over those facts, and the proposal states it cleanly: capability accrues to whoever can iterate fastest, not to whoever spent the most. That sentence is not in the record. It is a reading of the record, it is contested, and this project is going to want to say something very like it about AI — which is exactly the reason ransomware-and-the-rail gives for the kind existing at all.

Why not moment. There is a date, and it is the strongest date in modern history: 24 February 2022. But the house test for moment, settled by eliza-1966 and applied since by siri-2011, darpa-grand-challenge-2005 and asimo-2000, is a date on which something visibly happened in public, and the thing this entry is about did not happen on a date. It happened over four and a half years, and the argument is a comparison across that interval — the same two armies, the same terrain, iteration cycles measured in weeks against procurement cycles measured in years. A moment entry would have to pick 24 February 2022 and would then be an entry about an invasion, which is not what the citation occasion asks for.

Why not idea. On the test samuel-checkers-1959 set and bitter-lesson-2019 applied, an idea is something you can implement, because there is a loop or a formalism in it. There is a loop here — build, field, watch it fail, revise, refield — but it is not a construct, it is a description of how any competent organisation under existential pressure behaves, and it was neither invented nor formalised in Ukraine. What is specific to Ukraine is that the loop ran, in public, against a documented adversary, for long enough to be graded.

Why not prediction. Several graded predictions cluster around this subject and section 3 applies the full prediction discipline to them — Zaluzhnyi's five conditions of November 2023 in particular, which came with a stated deadline in all but name. But the entry's subject is not a claim about the future; it is a claim about a mechanism, and the forecasts are graded inside it rather than being promoted to the file's kind.

The id names a year and the content is four and a half. asimo-2000 had this exact problem and resolved it the same way: the header keeps 2022 because the id does and because the date is real, and the reader is told up front that the finding is the interval, not the year. The observation window here runs 24 February 2022 to the date of writing, 20 August 2026.

The proposal line, corrected

canon/proposals.md reads, in part: "the defence that was not supposed to hold, and why it did in the terms this project can actually use. Not a miracle: an exchange ratio and an iteration loop. Cheap software-defined systems built and revised in weeks, destroying equipment that costs orders of magnitude more and takes years to replace, with the countermeasure and the counter-countermeasure trading places every few months. Cite when: capability is assumed to accrue to whoever spent the most."

The citation occasion is right and section 2 is built on it. Four of the surrounding claims need work, and the proposal itself demanded two of the corrections in advance — state the external support, and verify every cost and count against primaries — which is the most useful instruction anyone has written into that file.

"Cheap software-defined systems" — half right, and the wrong half is the interesting one. The systems are cheap. Most of them are not software-defined in any sense that would satisfy a software engineer, and the single most consequential adaptation of the whole war was a piece of glass. Russian and then Ukrainian FPV drones defeated electronic warfare not by getting smarter but by trailing a spool of fibre-optic cable behind them — an unjammable wire, cost a few tens of dollars, no radio link to attack, no autonomy required. RUSI recorded the trade-offs in February 2025: degraded flight performance, a range of about 10 km, and the risk of entanglement with obstacles. By September 2025 Russia was reported to be producing more than 50,000 fibre-optic FPVs a month. The lesson a technology tracker should take from this is not comfortable: when jamming broke the radio link, the winning answer was a cable, not a neural network. Autonomy was the answer that got written about.

"Destroying equipment that costs orders of magnitude more" — true, and the arithmetic everyone quotes is wrong by about two orders of magnitude. Section 1 does the sum with RUSI's failure rates in it.

"The countermeasure and the counter-countermeasure trading places every few months" — right, and it is not a Ukrainian property. Russia's institutional answer, the Rubicon Center for Advanced Unmanned Technologies, was established by order of Defence Minister Andrey Belousov on 2 August 2024 and grew from about 1,450 personnel at the end of March 2025 to about 5,000 by spring 2026 against an authorised strength of 9,000. It iterates the same way and for the same reason: Rob Lee and Dmytro Putiata report that its detachments "receive equipment directly from manufacturers and provide direct feedback, which leads to a more rapid drone innovation cycle." An entry that presents the iteration loop as a Ukrainian trait rather than a property of the situation would be teaching the wrong lesson.

"The defence that was not supposed to hold" — the most load-bearing phrase in the proposal and the one that most needs the external-support paragraph. The Kiel Institute's Ukraine Support Tracker records that in 2022–2024 the United States alone allocated an average of EUR 17.3 billion a year in military aid and EUR 13.3 billion a year in financial and humanitarian aid, and that Europe allocated more again. Zaluzhnyi himself, writing in November 2023 as Commander-in-Chief, states that "the build-up of EW capabilities was carried out at the expense of international military material and logistical assistance" — the specific capability most often credited to Ukrainian ingenuity, attributed by Ukraine's own top soldier to somebody else's money. An account of this war built only out of workshops and iteration is a myth. The proposal knew that, said so, and this file honours the instruction.

1. What it is

1.1 The claim, in its own words

The argument has no manifesto, but it has an authoritative statement. On 1 November 2023, Valerii Zaluzhnyi published Modern Positional Warfare and How to Win in It. It is not a triumphalist document. It says the war "is gradually moving to a positional form", that prolongation favours Russia, and that Russia has "almost three times more mobilization human resources". Its prescription for logistics reads: "development and capacity building of the Ukrainian defence industry; creation of an asymmetric arsenal of weapons and equipment in Ukraine; creation, production and deployment of new weapons."

That is the doctrine. Note what it is not. It is not "cheap drones will win." It is a list of five conditions Ukraine had to meet to escape a positional war — air superiority, breaching mine barriers in depth, counter-battery effectiveness, reserves, and electronic warfare — of which the asymmetric arsenal is a means to several. Section 3 grades all five.

Note also the starting position the same essay records, because it is the reason asymmetry was a necessity and not a preference. Ukraine entered the war with 120 tactical aircraft, of which only 40 were technically suitable for use, and 33 anti-aircraft missile battalions, of which only 18 had fully serviceable equipment. About 65% of the types of jamming station in Ukrainian units were Soviet-produced, with only 25 new ones. The country did not choose to fight this way. It had nothing else.

1.2 The mechanism: how a change actually reaches the front

The interesting part of the iteration claim is not that Ukrainians had ideas. It is that a workshop's revision could reach a soldier without passing through a procurement system. Four mechanisms did that work, and they are dateable.

• Army of Drones, run from 2022 out of Fedorov's Ministry of Digital Transformation, which crowdfunded and centralised drone purchase outside the normal defence-procurement channel.
• Brave1, launched April 2023, a state cluster that grants, tests, certifies and matches manufacturers to units. Fedorov's own figure for what it grew: from about 10 domestic drone manufacturers in 2022 to more than 500 by 2025.
• Brave1 Market, launched April 2025, a catalogue where units order directly from manufacturers. By August 2026 the Ministry of Defence reported more than 800 certified systems and more than 400 combat units using it.
• e-points, from 2025: units are credited for verified destruction — 12 points for an enemy infantryman, 20 for a damaged tank, 40 for a destroyed one, up to 50 for a multiple-launch rocket system — and spend the credit in Brave1 Market. By August 2026 the Ministry of Defence reported more than 500,000 drones ordered through the e-points channel.

The fourth is the one worth a canon entry on its own. It is a price signal inside an army: a unit that finds a system works orders more of it, a manufacturer that ships something useless stops receiving orders, and the loop closes in weeks rather than at the next budget. It is also a textbook goodharts-law-1975 machine, and the canon should say so before someone else does — the measure is a verified strike video, and a measure that buys you equipment is a measure under pressure. That is a criticism of the mechanism, not a refutation of it, and I have found no evidence either way on whether the distortion is material.

1.3 The counter-mechanism

Russia's answer was the opposite in structure and worked well enough to be frightening. Rubicon centralises: research, procurement, training and deployment in one directorate, seven departments, six geographically distributed combat detachments by April 2025 scaling to seventeen detachments plus two battalions and six companies by 2026. Its speciality is killing the drone crews rather than the drones — tracing control emissions back to the operator — and interdicting logistics far behind the line. It was decisive in the recapture of the Kursk salient around Sudzha and supported the Pokrovsk and Myrnohrad operations in late 2025.

The same week in June 2026 produced two respectable and opposite readings of it. Lee and Putiata (FPRI, 12 June 2026) judge that rapid expansion has "dilut[ed] the elite quality of its recruitment and stress[ed] its frontline support pipelines", and that centralisation produced organisational friction rather than decisive superiority. Leonid Sokolov (Jamestown, 11 June 2026) notes that Ukraine struck Rubicon command posts in February, March and April 2026 — the hub model's specific vulnerability — and quotes Russian commentary that new drone brigades became "dumping grounds for low-quality personnel". Both also record that Rubicon narrowed a gap Ukraine had held. The honest summary is that centralisation was slower to start, scaled harder once started, and is more fragile to decapitation, and that four and a half years have not produced a clean verdict between the two models. Anyone citing this entry to prove that distributed beats centralised is citing more than it contains.

1.4 The costs and counts, verified

The proposal warned that this is the most numerically inflated subject in the canon in both directions. It is. What follows is what I could stand behind, with the provenance attached.

Unit cost of a Ukrainian FPV drone: about $580, derived rather than quoted. Ukraine's 2025 plan was to buy 4.5 million FPV drones for more than UAH 110 billion (about $2.6 billion), of which UAH 102 billion through the Defence Procurement Agency. That is roughly UAH 24,400 per airframe. The widely repeated "$400 drone" is in the right neighbourhood and is a floor, not an average: fibre-optic and heavy-payload types cost multiples of a basic quadcopter. For all UAV types together the average is about three times higher — the Ministry of Defence stated on 29 October 2024 that in the first ten months of 2024 it had contracted 1.6 million UAVs for over UAH 114 billion, about UAH 71,000 each, a figure that includes deep-strike aircraft.

Unit cost of a Shahed-136/Geran-2: unresolved, and the range is the finding. CSIS's February 2025 study lays out the published estimates — $20,000–30,000 (an Israeli missile expert, January 2023), about $80,000 (a British analyst who inspected a captured airframe in October 2022), $50,000 (Forbes Ukraine's working figure) — and then does the discipline this canon likes: it adopts $35,000, "the midpoint between the lowest cost estimate and the more often cited $50,000", and says so. Nobody outside Alabuga knows.

Unit cost of a Russian tank: not established, and I decline to pick one. The open-source range for a T-90M is $3–4.5 million; a Russian military journalist reported 350 million rubles in April 2026. None of these is a procurement document. Every "$500 drone kills a $4.5m tank" headline rests on a number nobody has produced a receipt for.

Interception, where the numbers are genuinely good. The Institute for Science and International Security published a monthly series on 7 July 2026, compiled by Igor Anokhin from Ukrainian Air Force reporting. It is the best quantitative record of the iteration race in existence, because it holds both sides' output against each other month by month:

• August 2025 — 4,128 Shahed-type launched, 680 hits, 27.9% of strike variants getting through.
• November 2025 — 5,447 launched, 826 hits, 24.6%.
• February 2026 — 5,059 launched, 548 hits, 17.0%.
• May 2026 — 8,161 launched, the monthly record at 263 a day, 543 hits, 10.5%.
• June 2026 — 5,749 launched, 464 hits, 12.6%.

Russia doubled its monthly launch volume across ten months and halved the proportion that got through. The absolute number of hits fell from 680 to 543 while launches rose from 4,128 to 8,161. This is what an iteration loop looks like when it is winning, and it is the single most defensible piece of evidence in this entry.

What did it. Wild Hornets' Sting interceptor — a roughly $2,000–2,500 quadcopter — accounted for more than 3,000 of the Shahed-type drones downed in May 2026 alone, out of roughly 7,500 intercepts that month, with the manufacturer claiming 80–90% hit rates and production over 10,000 units a month from March 2026. Note that two independent tallies for May 2026 nearly agree (8,161 launches and 7,618 downed per the Air Force series; 8,150 and 7,476 per the reporting around the Sting figure), which is worth more than either number alone.

Territory, for scale, because the exchange ratio did not stop the advance. DeepState recorded Russia taking 4,336 km² in 2025 (ISW: 4,831 km²), under 1% of Ukraine, with no advance deeper than about 40 km in any operational direction. As of 1 January 2026, Russia held 116,165 km², about 19.25% of Ukraine.

Human cost, because a file about exchange ratios that omits it is obscene. OHCHR verified, as of 14 July 2026, at least 16,431 civilians killed — including 803 children — and 48,613 injured since 24 February 2022, and recorded that in the first half of 2026 alone 1,396 were killed and 7,978 injured, with casualties from long-range weapons up 60% year on year and June 2026 the worst month for short-range-drone civilian casualties since the invasion began. Military casualties on both sides are contested and I have not verified any figure for them. The exchange ratio this entry is about is a ratio between machines. It is not a measure of how the war went for anyone in it.

2. Why a reading would cite it

LENSES.md names warfare explicitly under AI as accelerant, and attaches to that lens the hardest test on the board: it counts "only if the source shows AI was actually used in the work — a named model or method in a methods section, a technical account, or documented forensics — and says what became possible that wasn't." This entry exists to be the standing calibration for that test, in the domain where the test is hardest to apply and most often skipped.

The concrete case has three parts.

First, and most often: as the counter-example to a spend-implies-capability claim. When a lab, a nation or an investor argues that the capability lead belongs to whoever committed the most capital, this entry is the four-year test of that proposition with real stakes. Its verdict is not the one the proposal expected, and is better: iteration set the tempo, capital set the ceiling, and mass set the floor. Ukraine's iteration loop reliably beat Russia's on cycle time and reliably lost on volume, and neither fact decided the war. The reading that a lab's twelve-month release cadence beats a rival's twenty-four-month one survives contact with this evidence. The reading that cadence alone substitutes for capital does not: the Kiel Institute records that Ukraine's defence-industrial capacity grew by a factor of 35 between 2022 and 2025 to a projected USD 35 billion, while Ukraine's own 2025 budget could fund UAH 739 billion (about EUR 15.7 billion) of it — enough to use roughly half the capacity it had built. An iteration loop with no money runs at half throttle. That is the sentence this entry contributes to the Concentration lens, and it is a correction to the open-weights-versus-closed-APIs analogy rather than a confirmation of it.

Second: as the discipline for warfare findings under AI as accelerant. A reading that encounters "AI drones" in a headline should reach for this entry before depositing anything, because the specific failure mode is documented here. The deployed AI in this war is last-mile terminal guidance — a machine-vision lock that holds a target through the final few hundred metres when the radio link is jammed — and it is real, it is cheap, and it is not what "AI drone" is taken to mean. Kateryna Bondar's assessment for CSIS on 6 March 2025 is the sentence to quote: "Autonomy… is not yet present on the battlefield in the war in Ukraine." Her caution in October 2025 is the second: "much of the AI talk in Ukrainian drones is driven by marketing" — and, more sharply, that the systems cannot distinguish "a Russian and Ukrainian soldier, or even a soldier and a civilian." Ukraine's Commander-in-Chief Oleksandr Syrskyi has said in public that AI "can make mistakes" and should support rather than replace human decisions.

Third: as the case that a cost asymmetry has a direction, and you must say which. Section 4 does this at length. It is the single most abused feature of the subject.

The entry also forbids things, and that is half its value. It may not be used to argue that AI is winning or losing this war; the evidence does not support a finding in either direction at that altitude. It may not be used to deposit anything into the ledger — it is canon, it is not evidence, and nothing in it moves the needle. And it may not be used to argue that Ukraine did this alone, which is the reading its most sympathetic tellings invite and which Ukraine's own Commander-in-Chief contradicts in writing.

One note on SCALE.md, offered as fact and not as a revision. The owner's 2025 landmark is this subject at one remove, and its wording — "terminal-guidance autonomy on strike drones becomes ordinary… and AI-cued interceptor drones become the counter to mass strikes" — is more precise than almost anything published about it, because "terminal-guidance autonomy" is exactly the capability that exists and "AI drone" is not. The research behind this file supports the interceptor half strongly (section 1.4) and supports the first half with one qualification worth the owner's attention: the dominant answer to jamming was fibre-optic cable, not autonomy. Landmarks are the owner's to set and this file sets none.

3. What it got right, and what it got wrong

interpretation does not require this section. This subject does, because it is made of dated claims and because grading them is the only thing that distinguishes this entry from advocacy.

3.1 The exchange-ratio claim

Claim: a sub-$1,000 drone destroys equipment worth millions, at ratios in the thousands to one. Made: continuously from mid-2023. Grade: direction right, magnitude wrong by roughly two orders of magnitude, and the naive version is unusable.

RUSI's February 2025 fieldwork supplies the correction: "Between 60 and 80% of Ukrainian FPVs fail to reach their target, depending on the part of the front and the skill of the operators. Of those that do strike their targets, a majority fail to destroy the target system when striking armoured vehicles." Take the midpoint failure rate and the most generous possible reading of "a majority" — 30% arrive, half of those kill — and you need at least seven airframes per armoured kill, about $4,000 in drones, before counting the crew, the training, the reconnaissance drone that cued the strike, or the electronic warfare that made the sortie possible. Against a $3–4.5m tank that is still roughly 800:1 to 1,100:1 — an extraordinary asymmetry, and about one hundredth of the 9,000:1 the headlines imply.

There is a second correction, and it is bigger. Drones mostly do not kill tanks. Ukraine's Unmanned Systems Forces published their own tally on 6 January 2026 for a stated seven-month period: 532 tanks against ~2,500 guns and howitzers, ~7,700 vehicles and more than 5,500 motorcycles — armour is about 3% of the claimed haul. The tank is the icon of the story because a tank is what a drone video looks best hitting.

RUSI adds the sentence that most complicates the whole argument: "One of the main reasons for such a high proportion of kills being caused by FPVs is the relative lack of artillery in Ukrainian units." A brigade defending 18 km of front had four working howitzers. A fires officer described spending hours immobilising a single tank with FPVs, against an engagement where five BONUS shells knocked out three tanks in two minutes. The famous 60–70% share of damaged and destroyed Russian systems attributed to tactical UAVs is partly a measure of what Ukraine did not have. A NATO army reading this as a recommendation to substitute drones for guns would be inverting the finding.

3.2 The iteration-tempo claim

Claim: the adaptation cycle runs in weeks, and whoever iterates faster gains. Grade: right, well-evidenced, and not a national characteristic.

Justin Bronk and Jack Watling put the maintenance cadence at updates to software, sensors and radios every six to twelve weeks in April 2024. BCG's Nikolaus Lang told Euronews on 16 April 2026 that "drones evolve technologically every three to six months" and that a system is obsolete within twelve months of contract signature — against European procurement cycles he puts at six years and US prototyping at about eighteen months. The interception series in section 1.4 is the cycle's output, measured.

What the claim gets wrong is ownership. Russia iterated the fibre-optic answer first, scaled it harder, and built Rubicon to institutionalise the loop. In February 2025 RUSI recorded that "Ukrainian forces do not yet widely employ wire-guided FPVs" — Ukraine was the follower on the single most important adaptation of the middle war.

3.3 Zaluzhnyi's five conditions, graded

Claim: stated 1 November 2023; five conditions for escaping positional warfare. Due: the essay set no date, but it was written as an urgent prescription for 2024. Graded here at 20 August 2026, thirty-three months on.

• Air superiority — not achieved. Russian glide-bomb production rose from 40,000 UMPK kits in 2024 to an anticipated 70,000 in 2025, launched from 30–90 km behind the line, outside Ukrainian SAM envelopes. Miss.
• Breaching mine barriers in depth — not achieved. No Ukrainian operational breakthrough followed. Miss.
• Counter-battery — parity at best, on the essay's own framing, and by February 2025 RUSI found Ukrainian artillery "significantly underperforming, even when it is available." Miss.
• Reserves — not achieved. Force generation remained, in RUSI's February 2025 words, the "foremost problem". Miss.
• Electronic warfare — parity, achieved and then lost and then partly regained. Zaluzhnyi claimed parity in November 2023. Jamming then drove both sides onto cable. Partial.

Five conditions, four misses and a partial — and the front did not collapse. That is the most interesting grade in this file, and the reading it licenses is narrow but real: the thing that held the line was not on Zaluzhnyi's list. It was cheap attritable mass plus the loop that kept revising it, which his essay treats as one instrument among many. A forecast can be wrong about every named condition and right about the outcome, and this is what that looks like. It is also a warning about the base rate: the most informed possible forecaster, writing with full access, named the wrong five things.

3.4 The AI claim

Claim: in its strong form, that AI-enabled autonomous weapons are fighting this war. Grade: wrong as usually stated, right in a narrow and growing form, and the gap between the two is the finding.

What is real and deployed: last-mile terminal guidance, machine-vision lock-on in the final approach, on modules costing about $70 apiece from The Fourth Law by October 2025, at a scale Ukrainian reporting describes as tens of thousands of drones a month and last-mile targeting as standard practice across the FPV fleet. The claimed effect is large — Bondar's CSIS report cites 10–20% success manually against 70–80% with autonomous navigation, and "one or two drones per target rather than eight or nine"; one brigade using Fourth Law modules reported going from 20% to 80%. Also real: Delta, Ukraine's command-and-control layer, ingesting "tens of terabytes daily" of video, photo, acoustic and text and presenting one picture on a commander's tablet.

What is not real: target selection. Bondar, March 2025: autonomy "is not yet present on the battlefield." Yaroslav Azhnyuk, whose company builds the modules, says full autonomy in one-off demonstrations would already be feasible and mass deployment is "nowhere in sight." A machine that holds a lock a human placed is not a machine that chose.

And the counterfactual matters. The problem autonomy was reached for — the jammed radio link — was solved decisively by an unspooling glass fibre. If a reading needs one sentence from this entry for the accelerant lens, it is that one: in the most instrumented, most desperate, most iteration-friendly technical environment on Earth, the winning answer to a hard problem was not the AI answer, and the AI answer arrived second, smaller, and useful.

3.5 The verification gap

Claim: none, exactly — but the subject's numbers are systematically inflated and the pattern is dateable. Operation Spiderweb, 1 June 2025: 117 drones launched from inside Russia; Ukraine's General Staff claimed 41 aircraft; satellite imagery confirmed the destruction of about 12–13; NATO's assessment was around 40 damaged with 10–13 destroyed. Ukraine's Unmanned Systems Forces claim $20 billion of Russian equipment across 168,000 targets in a stated seven-month period; Oryx's visually confirmed cumulative record over four years stands at roughly 4,390 Russian tanks, 6,429 IFVs and over 23,000 items of equipment and is explicitly a floor. These are not contradictions and they are not reconcilable either. A reading that cites a belligerent's total without naming it as a belligerent's total has made the error this entry exists to prevent.

4. Commonly misused as

limit requires this section; interpretation does not. This subject needs it more than most entries in the canon, because it is the most quoted and least read technical story of the decade.

"A $500 drone kills a $3 million tank." What is true: the cost ratio between one airframe and one armoured vehicle is enormous. What it is made to say: that armour is obsolete and defence budgets should be reallocated. What it omits: 60–80% of FPVs never arrive; most that arrive do not kill armour; armour is about 3% of what drones actually destroy; the ratio excludes crews, training, cueing reconnaissance and electronic protection; and Ukraine's drone share of kills is partly a symptom of an artillery shortage RUSI describes in detail.

"AI is fighting the war in Ukraine." What is true: machine-vision terminal guidance is deployed at scale, cheaply, with large measured effects on hit rate, and Delta fuses sensor data at national scale. What it is made to say: that autonomous weapons are selecting and engaging targets. What the record says: Bondar, "autonomy… is not yet present on the battlefield"; Azhnyuk, mass deployment of full autonomy "nowhere in sight"; Syrskyi, AI "can make mistakes" and belongs in a supporting role. This is the misuse most relevant to this project and the one a reading is most likely to commit.

"Cheap mass always beats expensive systems." The same war contains the opposite exchange, running against the side with the cheap systems, and the comparison is the most useful thing in this entry. Against a Shahed-type drone at CSIS's conservative $35,000:

• a Sting interceptor at ~$2,000–2,500, at the manufacturer's claimed 80–90% hit rate, costs roughly $2,500–3,000 per kill — about 12:1 in the defender's favour;
• a Patriot PAC-3 at over $3 million costs about 86:1 against the defender for the same target.

Same weapon, same night, same sky, two ratios that differ by a factor of a thousand and point to opposite conclusions. A cost asymmetry is not a property of a technology; it is a property of a pairing, and anyone quoting one without naming the pairing is quoting nothing. This is precisely the finding iron-dome-2011 records for a different war fifteen years earlier, which is why it is this entry's ancestor.

"Decentralised beats centralised." Rubicon is the counter-case and it is strong. Two respectable analyses published a day apart in June 2026 read it in opposite directions. The gap narrowed. There is no verdict.

"Ukraine did it alone." Kiel's figures, and Zaluzhnyi's own attribution of the EW build-up to international assistance, refute this. By late 2025, 22% of procurement-based military aid was being spent inside Ukraine's own defence industry — up from 4% a year earlier — which is a real and important shift, and it is a shift in whose factory, not in whose money.

"The needle should move on this." It should not, on this file's account. This is canon. Nothing here is evidence, nothing here is deposited, and nothing here touches the needle — rule 7, and the plainest instruction in jobs/canon.md.

descends_from, and the strength of each link

Neither link is a citation. Both are identifications I made, and the strength of each is stated rather than implied.

• iron-dome-2011 is the strong one and is nearly a direct ancestor. That entry's whole subject is the arithmetic of interception — a machine deciding which cheap incoming rockets are worth an expensive interceptor, and the standing controversy about whether the exchange ratio is sustainable. Section 1.4 and the Sting/Patriot comparison in section 4 are the same problem with fifteen more years of hardware and a different sign. If a reading needs one companion to this file, that is it.
• alexnet-2012 is weaker, and is a lineage claim about a component. The $70 last-mile module is a small convolutional tracker running on an embedded chip; it descends from the vision stack AlexNet started, and I state that as an identification rather than as anything documented. No manufacturer cited in this file names a model architecture in public.

Two candidates I considered and declined. bitter-lesson-2019 is the tempting one — "cheap general mass beats exquisite handcrafted systems" has the shape of Sutton's argument — but the bitter lesson is specifically about computation and general methods defeating human-encoded knowledge, and mass-produced quadcopters are not that. It is a resonance and I have not made it a descent. goodharts-law-1975 is the honest ancestor of the criticism of e-points in section 1.2, not of the subject, and belongs in prose.

5. Sources

Read directly, primary or near-primary.

Kiel Institute for the World Economy, Kiel Policy Brief No. 203, "Europe Steps Up: Ukraine Support After Four Years of War", Christoph Trebesch and Taro Nishikawa, February 2026 — read as PDF in full for sections 1–4. Source of: the US 2022–24 averages (EUR 17.3bn military, EUR 13.3bn financial/humanitarian per year); the 2025 US collapse to a single EUR 0.4bn package; European military aid +67% and financial/humanitarian +59% in real terms against the 2022–24 average; NATO PURL launched July 2025 with 24 donors and over EUR 3.7bn by December 2025, supplying ~75% of Patriot missiles and ~90% of other air-defence missiles delivered; stockpile-based aid falling 60%→46%→31%→7% (H1 2025) and recovering to 35% in H2 2025; the ×35 growth in Ukrainian defence-industrial capacity to a projected USD 35bn, against UAH 739bn of Ukrainian budget capable of funding about half of it; the Danish Model (18 Bohdana howitzers, ~EUR 50m, July 2024; Denmark ~EUR 830m to December 2025) and the Dutch Model (~EUR 790m in 2025, primarily drones); and procurement inside Ukraine rising 4%→12%→22%.

RUSI, "Tactical Developments During the Third Year of the Russo–Ukrainian War", Jack Watling and Nick Reynolds, published February 2025, fieldwork November 2024 and January 2025 — read as PDF, pages 1–20. Source of: the 60–80% FPV failure rate and the "majority fail to destroy" finding; the 60–70% share of damaged and destroyed Russian systems; the artillery-shortage explanation and the 18 km / four howitzers brigade; the BONUS anecdote; ~50% of brigade casualties from FPVs, artillery and glide bombs; fibre-optic trade-offs and the ~10 km range; "Ukrainian forces do not yet widely employ wire-guided FPVs"; UMPK production 40,000 (2024) → 70,000 anticipated (2025) and the 30–90 km standoff; the 1,200 km line of contact and ~800,000 Defence Forces with under 25% combat power; and the Kh-101 visual terrain-tracking upgrade. One defect recorded rather than resolved: the published text reads "Each day in August 2025, Ukrainian air defenders reported…" in a paper published February 2025 from fieldwork ending January 2025. The year cannot be right; I could not establish which year is meant and have not used the figure.

Valerii Zaluzhnyi, "Modern Positional Warfare and How to Win in It", 1 November 2023 — read as PDF via the National Security Archive's copy of the Economist website document. Source of: the five conditions; "creation of an asymmetric arsenal"; the 120 aircraft / 40 suitable and 33 battalions / 18 serviceable figures; 65% Soviet-produced jamming station types with only 25 new ones; "the build-up of EW capabilities was carried out at the expense of international military material and logistical assistance"; "we have practically achieved a parity in the EW tasks performance"; the Excalibur degradation; 15–20 km mine belts; and Russia's ~3× mobilisation resource.

CSIS, "Calculating the Cost-Effectiveness of Russia's Drone Strikes", Neil Hollenbeck, Muhammed Hamza Altaf, Faith Avila, Javier Ramirez, Anurag Sharma and Benjamin Jensen, 19 February 2025 — the Shahed cost estimates and the adopted $35,000; ~$350,000 per target struck; PAC-3 over $3m; NASAMS AIM-9X just over $1m.

CSIS, "Ukraine's Future Vision and Current Capabilities for Waging AI-Enabled Autonomous Warfare", Kateryna Bondar, 6 March 2025 — "Autonomy… is not yet present on the battlefield in the war in Ukraine"; the 10–20% versus 70–80% success figures; one or two drones per target rather than eight or nine; Delta's tens of terabytes daily; Zvook acoustic stations at ~$500.

Kyiv Independent, "AI drones in Ukraine — this is where we're at", Kollen Post, 25 October 2025 — the Fourth Law $70 modules; Azhnyuk's "nowhere in sight"; Bondar on marketing and on the inability to distinguish a soldier from a civilian; Syrskyi's "can make mistakes"; the 20%→80% brigade claim.

Institute for Science and International Security, "Monthly Analysis of Russian Shahed-136 Deployment Against Ukraine", Igor Anokhin, 7 July 2026, compiled from Ukrainian Air Force reporting — the full August 2025 – June 2026 series in section 1.4, including the May 2026 record of 8,161 launches at a 10.5% strike-variant hit rate, and Russia's 2026 plan of ~60,000 long-range plus ~50,000 decoy UAVs. The series combines kinetic kills and EW neutralisations under one heading, which the report states.

Ukraine's Ministry of Defence, 29 October 2024 — 1.8 million UAVs contracted for 2024–2025 at nearly UAH 147bn; 1.6 million for over UAH 114bn in the first ten months of 2024, with 1.28 million delivered.

OHCHR / UN Human Rights Monitoring Mission in Ukraine, release of 14 July 2026 — 16,431 civilians killed and 48,613 injured since 24 February 2022; 1,396 and 7,978 in the first half of 2026; long-range-weapon casualties up 60% year on year; June 2026 the worst month for short-range-drone civilian casualties since the invasion.

Foreign Policy Research Institute, "Inside Rubicon: The Structure of Russia's Elite Drone Center", Rob Lee and Dmytro Putiata, 12 June 2026 — Rubicon's founding order of 2 August 2024, personnel growth, structure, Kursk and Pokrovsk roles, the manufacturer-feedback loop, and the dilution warning.

Jamestown Foundation, Eurasia Daily Monitor, "Rubicon Reveals Limits of Russia's Drone Centralization", Leonid Sokolov, 11 June 2026 — the February/March/April 2026 strikes on Rubicon command posts, the Molniya cost reduction, and the "dumping grounds" quotation.

Ukrainska Pravda, 6 January 2026 — the Unmanned Systems Forces' own tally: 168,000+ targets, US$20bn, 532 tanks, ~2,500 guns, ~7,700 vehicles, 5,500+ motorcycles, 50,000+ personnel, 832,000 combat missions. Recorded as a belligerent's claim. The stated seven-month period does not map onto any USF founding date I could verify (decree 6 February 2024, decree 28 June 2024, law 3 September 2024), and I have not resolved what it is counted from.

Ukrainska Pravda, 11 June 2026 — Sting interceptors downing more than 3,000 Shahed and Gerbera drones in May 2026, against 8,150 recorded launches and 7,476 intercepts.

Euronews, Evi Kiorri, 16 April 2026 — BCG's Nikolaus Lang on the three-to-six-month evolution cycle, twelve-month obsolescence, the six-year European procurement cycle, and the $70bn/$7bn defence-VC gap.

At one remove, via search summaries and secondary reporting — check these before a reading leans on them. The 2025 FPV plan (4.5 million for over UAH 110bn / ~$2.6bn, UAH 102bn via the Defence Procurement Agency) and the H1 2026 figure of UAH 333.6bn (~$7.6bn) with ~96% to Ukrainian producers; Fedorov's 10→500 manufacturers; the Brave1 (April 2023) and Brave1 Market (April 2025) dates and the August 2026 scale figures (800+ certified systems, 400+ units, 500,000+ drones via e-points); the e-points tariff of 12/20/40/50; Russian fibre-optic FPV production exceeding 50,000/month by September 2025 and the 15–20% (July 2025) share estimate; Zelensky's escalating capacity statements (4 million/year in October 2024 and February 2025, 8 million in June 2025, "producing 10 million a year" in July 2026) — treat these as capacity claims by an interested party, not as output; Syrskyi's 404 Shahed-type per day (January 2026); the Spiderweb figures (117 drones, 41 claimed, 12–13 satellite-confirmed, NATO's 10–13 destroyed of ~40 damaged); Oryx's 4,390 tanks / 6,429 IFVs / 23,000+ items; DeepState's 4,336 km² and ISW's 4,831 km² for 2025 and the 116,165 km² / 19.25% total at 1 January 2026; Bronk and Watling's six-to-twelve-week update cadence (April 2024); the RUSI figure of ~10,000 Ukrainian drones lost per month, which comes from a report of 19 May 2023 that I did not read — it reaches this file only through contemporaneous reporting of it; and the T-90M cost range.

Attempted and failed. Foreign Policy's "Is Ukraine's Decentralized Drone Innovation a Blip or a Revolution?" (10 July 2026), the IISS Military Balance assessment of Spiderweb losses, David Kirichenko's Forbes piece of 9 August 2026 on autonomous strike, TIME's account of the points system and the defence-blog cost breakdown all returned HTTP 403 or 503 and are not used. The RUSI Meatgrinder PDF exceeded the fetch size limit. The consequence worth naming: this file's account of the state of play in the second half of 2026 is thinner than its account of 2022–2025, and a reading citing it for anything after June 2026 should check first.

Not established, and stated as such above. The unit cost of a Shahed-136 to Russia; the unit cost of a Russian tank; military casualties on either side; the proportion of the Ukrainian FPV fleet actually carrying a terminal-guidance module; whether the e-points incentive materially distorts targeting; and whether centralised or distributed drone development is the better model, on which two serious analyses published a day apart disagree.

Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.