In plain English. Two things. First, a correction: the burst stage could double a unit's Kill because a detachment nudged its melee slightly above its shooting. Chief Librarian Mephiston was worth four times as much in Rage-Cursed Onslaught as anywhere else; now the gap is the size of the rule (about 14%). The headline numbers barely move. Second, a report: we varied the three least certain pieces of Kill one at a time and watched each unit's rank in its army. The Tyranids, which the ledger was tuned on, are solid: 13% of units move much. Across the 16 armies, 41% of units move enough to call "uncertain". The biggest single cause is the melee landing scale: how often a charge is taken to land. Every unit is labelled steady or uncertain in 2026-10-05-kill-robustness.json, ready for the public pages (T-697).
Part 1: the burst spike (a correction; step 45)
Measured on the ledger snapshot c5e7d85b43f8 (T-695's re-freeze). A first run on a20dbd2c3316 gave the same numbers everywhere except Death Guard, the only army that re-freeze moved.
Why it happened. tools/burst.js picked each form's "natural mode" as the phase (shooting or melee) with the larger mean Kill, and then read that phase's Burst. Mephiston's two phases have almost the same mean Kill (ranged 67.0, melee 64.0), but his melee Burst (finishing a target in one activation) is about twice his ranged one. Rage-Cursed Onslaught's melee re-roll pushed melee to 71.6, so the mode flipped and his Burst jumped from 26.5 to 51.1. The same edge case caught 24 unit-size-loadout groups across the armies.
The fix. The mode is now the phase whose Burst counts for more as the ledger page counts it (naturalMode, meleeWeight). Melee Burst is weighted by valueForm's landing share for the form, at the benchmark's lands scale, never past 1. It is also multiplied by the form's rules on melee Kill, such as Fights First's ×2.64. The comparison is against the ranged Burst, so the value read is the larger of the two and moves smoothly with the dice. --prefer and --life take the mode from the burst file, so the three stages agree; the army-rule variants (tools/armyburst.js) pick theirs the same way. Burst, prefer, life and burstfx were rebuilt and merged for all 16 armies; no frozen input moved.
The first build missed the rule multipliers. It weighed melee by landing share alone. The robustness run then caught Gaunt's Ghosts jumping from 51st to 6th on one class-mix draw. They have Fights First, so the page multiplies their melee Kill by 2.64 and a small mode flip moved them a long way. Rules on melee Kill went into the weight, and every army was rebuilt again. The numbers below are from that second build.
The switch cases, before and after (effective Burst = melee × landing share at lands 1.5; Track A Value of the form):
| Unit (army) | Form | Before | After |
|---|---|---|---|
| Chief Librarian Mephiston (Blood Angels) | Rage-Cursed Onslaught vs Liberator Assault Group | melee 43.2 vs ranged 26.5; Value 2,026 vs 503 (4.0×) | melee 43.2 vs melee 37.7; Value 2,026 vs 1,777 (1.14×); best now Librarius Conclave (Divination) 2,510, still 2nd |
| Tormentors (Emperor's Children) | Coterie of the Conceited vs Carnival of Excess | melee 48.8 vs ranged 27.1; 1,203 vs 847 | melee 48.8 vs melee 38.3; 1,203 vs 936; 5th, unchanged |
| Gunwagon (Orks) | War Horde vs Blitz Brigade | melee 49.5 vs ranged 23.5; 820 vs 421 | melee 49.5 vs melee 41.7; 834 vs 714; 20th, unchanged |
| Killa Kans (Orks) | Dread Mob vs Blitz Brigade | ranged 61.0 vs melee 37.3; 950 vs 611 | ranged 61.0 vs ranged 47.4; 967 vs 704; 14th, unchanged |
| Big Mek (Orks) | War Horde form | melee 30.4 vs ranged 35.1 | ranged 35.1 (no flip); best form 691 → 705, 29th → 31st |
| Death Company Intercessors, Marines with Bolt Rifles (Blood Angels), 5 | Ironstorm Spearhead | ranged 30.1 / 28.2 vs melee 31.3 / 29.3 | melee everywhere; best forms unchanged (807, 1,057) |
| Kasrkin 10, Mandrakes 5 and 10, Reavers 6 | the one detachment that flipped them to ranged | ranged below their melee | melee everywhere; best forms unchanged |
| Tyrannocyte (Tyranids) | Swarming Instincts | ranged 15.1 vs melee 16.9 | melee 26.2; best 224 → 228, 48th |
| Ripper Swarms, ranged Warriors (Tyranids); Cadian, Death Korps (Astra Militarum); Trukk (Orks) | the minority form | unchanged: the minority phase really is the larger Burst |
In all, 747 of 14,498 solo forms change mode (5%): Space Marines 186, Dark Angels 241, Orks 95, Astra Militarum 50, Thousand Sons 44, Blood Angels 39, Tyranids 10, Death Guard 4, Leagues of Votann and Space Wolves 0.
Headlines, Track A (step 34b), before → after:
| Before | After | |
|---|---|---|
| Tyranids, 3 11th-edition lists | 82.4 | 82.4 |
| Tyranids, all 5 lists | 79.1 | 79.1 |
| Exam, pooled 15 armies | 59.2 | 59.4 |
| Exam, 14 (without Votann) | 59.1 | 59.3 |
By army: Orks +1.8, Astra Militarum +1.3, Death Guard +0.4 (54.6 → 55.0), Space Marines −0.2, Adeptus Mechanicus −0.3, Dark Angels −0.3, World Eaters −0.3. The rest are unchanged, Blood Angels included (61.3).
Still there: a detachment that buffs only one phase can still make that phase the larger Burst, and the jump is then real. The Emperor's Children Sorcerer is 402 in Coterie of the Conceited against 251 elsewhere: Coterie lifts his melee Burst past his shooting.
Part 2: robustness (a report; no value change)
Run on Track A after the fix, with node tools/killrobust.js --measure --army <a> then --report. Each piece was varied on its own and the others held at Track A:
- lands, the melee landing scale: 1.0, 1.25, 1.75, 2.0 (Track A 1.5).
- β, target choice: 3, 12, inf (Track A 6). With burstKill on, Kill is the same Burst at every β, so β moves only role scarcity's factor.
- mix, the enemy field's class mix: six seeded draws. Each class's share is multiplied by a factor between 0.75 and 1.25, then renormalised to 2,000 points. Burst is recomputed from each form's measured chance to finish each of the typical army's 611 targets, with the mode and the army's scale chosen again. Role scarcity is recomputed with scarcity.js's own formula on the new shares.
At today's mix the measured chances reproduce the burst files within 0.01. Astra Militarum has two near-ties that read the other way (Ministorum Priest, 6.44 against 6.44).
Steady means the unit's letter never moves more than one band, and its rank range is under 15% of the army's ranked units. Anything else is uncertain. The driver is the piece whose variations alone give the widest rank range.
| Army | units | uncertain | share | driver: lands / β / mix | lands + mix only |
|---|---|---|---|---|---|
| Tyranids | 52 | 7 | 13% | 6 / 0 / 1 | 13% |
| Space Marines | 84 | 44 | 52% | 30 / 12 / 2 | 42% |
| Adeptus Mechanicus | 34 | 13 | 38% | 6 / 7 / 0 | 32% |
| Agents of the Imperium | 29 | 12 | 41% | 5 / 7 / 0 | 17% |
| Astra Militarum | 72 | 21 | 29% | 6 / 11 / 4 | 11% |
| Blood Angels | 20 | 10 | 50% | 5 / 5 / 0 | 40% |
| Dark Angels | 93 | 54 | 58% | 45 / 6 / 3 | 57% |
| Death Guard | 36 | 22 | 61% | 16 / 6 / 0 | 50% |
| Drukhari | 37 | 16 | 43% | 4 / 12 / 0 | 22% |
| Emperor's Children | 23 | 14 | 61% | 8 / 5 / 1 | 61% |
| Genestealer Cults | 90 | 13 | 14% | 5 / 8 / 0 | 8% |
| Leagues of Votann | 22 | 9 | 41% | 6 / 2 / 1 | 32% |
| Orks | 55 | 26 | 47% | 18 / 8 / 0 | 36% |
| Space Wolves | 21 | 10 | 48% | 5 / 4 / 1 | 29% |
| Thousand Sons | 35 | 17 | 49% | 9 / 6 / 2 | 40% |
| World Eaters | 30 | 14 | 47% | 2 / 6 / 6 | 37% |
| All | 733 | 302 | 41% | 32% |
Each piece on its own: lands makes 23% of units uncertain, β 15%, mix 5%. By Track A letter, uncertain are S 16 of 81, A 69 of 145, B 161 of 220, C 55 of 184 and D 1 of 103. The middle of the table swings most, and the bottom almost never.
The 10 most uncertain units (by rank range as a share of the army):
| Unit | Army | Track A | Rank range | Letter changes | Largest move | Driver |
|---|---|---|---|---|---|---|
| Death Company Dreadnought | Blood Angels | S, 1st of 20 | 1–16 | 15% | 3 bands | β (16th at β inf) |
| Forgefiend | World Eaters | B, 14th of 30 | 3–21 | 46% | 2 | β (3–21; mix 11–19) |
| Typhus | Death Guard | A, 10th of 36 | 6–26 | 69% | 2 | β (10–26; lands 6–17) |
| Captain in Terminator Armour | Space Marines | A, 26th of 84 | 8–50 | 54% | 1 | lands (8–50) |
| Baal Predator | Blood Angels | C, 13th of 20 | 6–16 | 38% | 2 | lands (6–15) |
| Azrael | Dark Angels | B, 36th of 93 | 16–62 | 31% | 1 | lands (16–62) |
| Thulia Ghuld | Adeptus Mechanicus | A, 7th of 34 | 6–22 | 31% | 2 | β (7–22) |
| Bladeguard Veteran Squad | Dark Angels | B, 38th of 93 | 17–60 | 15% | 1 | lands (17–60) |
| Einhyr Champion | Leagues of Votann | B, 8th of 22 | 6–16 | 31% | 1 | lands (6–16) |
| Judiciar | Space Marines | B, 43rd of 84 | 19–57 | 38% | 1 | β (19–53; lands 24–57) |
Reading it.
- lands decides melee characters. Lands swings melee characters and melee squads most: Captains, Lazarus, Azrael, Bladeguard, Einhyr Champion; Typhus too (lands 6–17), though β now moves him further. Their Kill is mostly melee, and the scale moves them against shooters. This is the most important uncertain piece. The charge model's share and the 1.5 on top of it are the least measured numbers in Kill.
- β is partly an artefact. With burstKill on, β only moves role scarcity, through
meta.medianFactor[β]. That median collapses in two armies: Blood Angels 0.197 at β 6, 0.086 at 12 and 0 at inf; Agents of the Imperium 0 at inf. At 0 the page treats scarcity as off for the whole army, and near 0 it inflates every factor. The Death Company Dreadnought's fall to 16th at β inf is that switch, not a real change in what it kills. Without β, 32% of units are uncertain rather than 41%. - The class mix matters little. A ±25% field moves 5% of units on its own. World Eaters (20%) and Blood Angels (15%) are the most sensitive.
Files
- tools/burst.js:
naturalMode,meleeWeightandkillMode;--preferand--liferead the burst file's mode; andburstOfForm(..., { ps: true })returns each target's chance. - tools/killrobust.js:
--measurewrites build/ledger/robust-<army>.json (git-ignored), and--reportwrites the JSON beside this file. - reports/tiers/2026-10-05-kill-robustness.json: per army, the uncertain share and the drivers. Per unit: status, Track A letter and rank, min and max rank, letter-change share, largest band move, driver, and each piece's rank range.