How we score: every factor, formula and constant

The canonical spec of Tactical Reroll's maths (T-535). The public summary is /about/'s "How the scores work"; this note is the long version, for anyone who wants to check our working. Everything here is our own words and our own numbers: no Games Workshop rules text. A change to tools/build_rankings.js, tools/targets.js or an effects file (detfx, leaderfx, unitfx, armyfx) updates this note in the same commit, and tests/unit/scoringdoc.test.js fails if a target, attacker, Utility part, trait, job or delivery constant in the code isn't named here.

Live means it's what the site publishes today. Proposed means it's under test in the Lab (T-523, Tier system v2) and not on the site.

1. One attack, and the table we play on

One attack (live). An attack hits on its BS or WS, wounds by comparing Strength with Toughness (twice the Toughness or more: 2+; more: 3+; equal: 4+; half or less: 6+; otherwise 5+), and is saved on the better of the armour save worsened by AP and the invulnerable save. Each unsaved wound does its Damage. Feel No Pain then ignores each point of damage on its roll. Damage beyond what a model has left is lost; it doesn't spill to the next model. Weapon abilities change the chances along the way: Lethal Hits, Sustained Hits, Devastating Wounds, Anti-X, Twin-linked, Torrent, Blast, Heavy, Rapid Fire, Melta, Lance, Ignores Cover, Psychic, Hazardous. Each keyword is read whole, the way the data spells it (T-582): an ability followed by keywords counts only against a target that meets them ("Lethal Hits: non-Monster/Vehicle" against neither a MONSTER nor a VEHICLE; "Sustained Hits 2: Monster/Vehicle" against either); Anti-X takes a slash list and "non-" the same way, with or without its hyphen or its "+"; Blast X and Cleave X add X attacks for every five models in the target (Cleave on the one target we always pick); Sustained Hits D3 rolls its extra hits on each critical hit, and Rapid Fire D3 or D6+3 rolls its extra attacks at half range; "Twin Linked" is Twin-linked; Conversion makes hit rolls of 4+ critical hits beyond 12" (read as beyond half range: every Conversion gun reaches 24" or more); Psychic Assassin makes 6 attacks against a PSYKER and isn't Psychic; Overcharge fails its hazard roll on 1 to 4. Not read, because they aren't dice: Harpooned (a charge bonus), Reverberating Summons (models returned to another unit), Linked Fire (measuring from another Fire Prism), Plasma Warhead (how the Deathstrike picks targets), Defensive Array and Hive Defences (Overwatch for no CP).

Two engines that agree (live). The calculator rolls the dice (app.js rollSingleSequence, thousands of runs). The rankings use exact averages (docs/ev.js), which walk the whole spread of unsaved wounds through the target's models, so "one unsaved wound against a 2-wound model" kills nothing on average. The unit tests hold the two together (tests/unit/ev.test.js, volley.test.js).

The neutral table (live), the same for every unit (tools/targets.js OPTS):

2. The archetypes

Targets (live, tools/targets.js TARGETS): stand-ins for what an army has to kill. Each is matched to a real unit in the calculator (app.js STAND_INS, kept in step by tests/unit/standins.test.js). Each has equal weight today; under Lenses (T-531) the weights would follow the meta.

Attackers (live, ATTACKERS): the profiles units get shot and hit with, per hit.

Proposed (T-529). Swap the heavy vehicle for a big target with an invulnerable save (T11 Sv3+ 4++ W18), add a tough horde (10 × T5 Sv5+ W1), and add a mortal-wounds attacker (D2, ignores saves, Feel No Pain still counts). Weights stay equal. In the test this raised agreement with list share from 0.119 to 0.147. Weighting the archetypes by the meta lowered it, to 0.088.

3. The live scores

Each unit is scored at its smallest size and its points, among every non-Legends unit (the pool).

• Offense. For each target: the expected wounds removed per 100 pts, the better of the unit's shooting and its melee. Melee uses its best weapon per model plus any extra-attacks weapons. Each target is scored with its own best legal squad (section 5) where that beats the default wargear. Each target becomes a percentile over the pool; Offense is the percentile of their mean.
• Defense. For each attacker: the hits needed to destroy the unit per 100 pts, models × hits per kill ÷ (chance to wound × chance to fail the save) ÷ pts × 100. Hits per kill walks the attacker's damage, after the unit's own Feel No Pain, through each model's wounds. Then percentiles and a mean, as for Offense.
  • Toughness (T-546). A unit whose models have different Toughness is attacked at the highest T among them (the core rules' Multiple Toughness; on a unit with a Leader attached, the highest among its bodyguard models, never the leader's). Defense scores the unit alone, so it is the highest T among BSData's kinds of model on the datasheet (bsdata.js models), the first statline when there are none: a Decimus or Aquila Kill Team T6 (its Gravis Veteran), a Traitor Enforcer T6 (its Ogryns). Black Spear Task Force's tMajority (T-545) is a detachment rule and Defense counts no detachment. The calculator and ev.js still take the first statline (open, below).
  • Heals (T-556). A unit's own heal in each of its Command phases (the coverage files' heal rules to itself with per: "cmd", with kindfx.json's numbers: Reanimation Protocols on every Necron datasheet that lists it, Old One Eye's regeneration) multiplies each attacker's hits to destroy by the heal's closed form below, healLift(models × W, T0, h, 5), the tier test's own. T0 is the turns it lives without the heal, as the tier test counts them: each attacker's hits to destroy the whole unit (def × pts ÷ 100) over half the pool's median, between 1 and 5, the mean of the five. h is healPerTurn's. A unit that dies in its first enemy turn gains nothing (k = 1), so most lone Necron characters move little. Not counted: heals that need a kill or a fight (Swallow Whole, Digestion Spine), heals from other units (a Technomancer's or a Ghost Ark's repair, a joined character's), and heals the data has no numbers for. rankings.json's def stays the bare hits; hl is the lift and hr the rules behind it; Defense's percentiles, the unit pages' Defense table and the app's Rankings use def × hl. Feel No Pain was already in (above).
• Value = (Offense + Defense) ÷ 2. Utility and Meta sit beside it, never in it.
• Utility. Four parts, each a percentile over the pool; their mean is ranked within the unit's own army (T-198).
  • oc, board control: OC × models per 100 pts.
  • act, action economy: 100 ÷ pts (cheap bodies for actions).
  • reach: Move, +2 with Fly, +6 with Deep Strike, +6 with Infiltrators, +4 with Scouts.
  • sup, support:
    • a Leader counts 1, or, when its effects are known, 0.5 + min(1.5, its lift on its best squad × 5) + 0.5 if it protects the squad, at most 2;
    • its auras (at most 2);
    • 1 each for Transport, Lone Operative, Stealth, revives and Fights First.
• Traits (read from keywords and ability names, shown as badges): action (65 pts or less), ds, inf, scouts, fly, leader, aura, transport, lone, stealth, ff, revive.
  • Rule links as a bare unit has them (T-565, docs/bsdata.js hiddenBare): a datasheet's rule counts only when the unit has it with nothing taken and nothing joined. A link hidden unless an enhancement is taken or a leader is attached (the Chameleonic enhancement's Stealth, Lord Invocatus's Deep Strike, a Technomancer's Feel No Pain) is not the unit's own; a link hidden only by the army's choices (a detachment, a game mode: the Lictor's Deep Strike outside Ambush Predators) stays. 76 unit rows lost such a rule, 21 of them a trait (Stealth, Deep Strike, Scouts, Infiltrators).
  • Sizes (T-565, _raisedByModel): a model's maximum raised once another model in the unit reaches a count (a second Nodebeast at 20 Neurogaunts) counts in the unit's size: Neurogaunts 11 to 22.
  • Wargear costs are not read. bsdata.js reads a unit's own points (its size breakpoints) and no option's cost; BSData prices about 70 options at the pinned commit (the Tyrannofex's rupture cannon +20 and acid spray +10, the Riptide's ion accelerator +25, unit upgrades). The best-loadout solver picks a paid weapon at the unit's base price (section 8).
• Meta. The share of the lists that could field the unit that take it, taken ÷ (lists + 3). The 3 lists of doubt stop one list of one from reading as 100%. Meta itself is shown against the army's most taken unit (100).
• Today's letter (T-219; the meta lift, T-226, 26 Sep):
  • The rule: tier score = the higher of Value and 40 + 100 × share; S 80+, A 65+, B 50+, C 35+, D below.
  • Why: Jordan: "if 40% of winning armies are taking it, it's probably S tier".
  • The cost: 264 units wear S (T-524), and a unit's letter mixes the maths with how often it's taken.

4. Effects: leaders, detachments, units' own abilities, army rules

Each file holds what a rule does to the dice, in our own short words and numbers, never its text. Each effect has:

• a side: att when attacking, def when attacked;
• mods, for example +1 to hit, re-rolls, Lethal Hits, Feel No Pain 5+, −1 to be hit;
• conditions the engine checks: melee or ranged, the target's keywords, the unit's keywords, a weapon's name. Conditions it can't check start on, and the player can switch them off. An effect marked off: true (a damaged bracket, a once-a-battle or rarely-met condition, T-541) starts off instead, and the player can switch it on. The build tools (rankings, Banner, the tier test) read no unitfx effect yet; a tool that does should leave off effects out.

• Heals between turns (T-540). A heal rule's numbers: n (a dice expression), per (cmd: each of its own Command phases; kill: n per enemy model its weapons weapon slay in a fight; fight: once after a fight it slew in), models (slain models come back, one point a model with 1 wound, once none is wounded), to (self; unit: the unit a character leads; near: other units, keywords who). In the calculator they're chips of their own (on unless off) and play only To the death, turn by turn: the side whose turn it is heals first, then fights; a kill heal follows the blows. The answer says how many wounds came back a fight on average; the played-out fight's round strip says who healed how many. ev.js has the closed form (below).
• Mortal wounds outside attacks (T-540). A mortal rule's numbers: n dice (pm: for each model left), each on on or more (0: no roll) dealing d mortal wounds; when it fires (move, shoot, arrive, fightEnd with its turn's shooting; charge, fight with its fight; shock, react start switched off), range, vs (the target's keywords it needs), near (each enemy unit in range), max, and v (the tier test's expected mortal wounds per use into the five targets, mortalfx.json's field; worked out from n, on and d when missing). In the calculator each is a chip and a "weapon" of its own in the volley: its dice are attacks, each mortal wound a bullet of one damage that skips the wound roll and the save (Feel No Pain and Feel No Pain against mortal wounds apply; damage reductions don't), spilling over models. The funnel's wound gate and the answer read "+N mortal". In a shoot-then-fight fight to the death, a move or shooting rule fires only in the first turn (the turns after are fights).
• Auras on other units (T-540). A coverage rule's aura object (range, to, not) says who receives it; what they get is aura.effects (kindfx.json, for an aura noted in words) or else the rule's own effects. In the calculator a joined character's auras reach the squad it stands with when the squad's keywords match (range taken as met, as a leader's bonus is), and its enemy-side auras (a rule of kind enemy named as an aura, a debuff on whoever attacks it) come with it; a rule leaderfx.json already gives the squad isn't given twice. The unit's own aura reaches itself only through aura.effects (its own effects are already its own). Each is a chip like the rest.
• The vocabulary the coverage kept asking for (T-545). Attacking: tMinus (the target's Toughness −N for the wound roll; not +N Strength, which differs at the doubling steps: S4 against T9 wounds on 6s with −1 T, on 5s with +1 S), rrOneHit / rrOneWound (one hit roll and one wound roll re-rolled a volley, shared by its weapons, fresh each turn of To the death; ev.js adds P(a miss) × a fresh die's chance, per weapon). Attacked: tPlus (+N Toughness), svUp (the Save improved by N, never past 3+: Take Cover!), tMajority (a mixed unit uses the Toughness most of its kinds of model have; BSData gives no counts, so each kind counts once, the lower on a tie). Conditions: sVsT gt / ge / lt / le (the attack's Strength against the target's Toughness, the datasheet's or edited values before other modifiers; unknown, the player's call) and rapid (Rapid Fire weapons only). kindfx.json gives noted rules effects, adds an option's extra, patches an effect or option by its label (Death Guard's contagion from +1 S to −1 T; 24 "Strength above its Toughness" conditions the player had to vouch for), and lets an army rule reach units that don't list it (Orders). Army rules (armyfx.json) the datasheet lists, or reaches, are chips in the calculator now, each starting switched off (they hang on the board), a rule's choice picked in the Army part. Not in the live rankings (they read neither kindfx.json nor army rules).
• A friendly unit nearby (T-551). The calculator's Army part has a Nearby pick: one unit of the side's army that stands near its unit (range taken as met, as for a joined character's aura). Its auras that reach the unit's keywords count (aura; kindfx.json now gives numbers to 11 Imperial Knights Bondsman duties and Ion Aegis, to an Armiger, and 7 War Dog auras of the Chaos Knights, to a War Dog), and so do the marks it lends its friends: a rule's lend (effects, the friends' keywords to), each against the marked unit only (T-548's Marked chip): Guide, Doom, Death Hex, Blight Bombardment, the Storm Speeders', Incursors', Exocrine's, Psychophage's and 15 more (23). A unit's own mark (a Norn's prey) isn't lent. Kept per army and unit (hre.det.v1). The unit pages' aura line (T-547) reads the new Knight auras too (15 pages); no score counts them.
• Tests a unit takes on itself (T-550). One shape, selfTest { on, fail, mw }: d6 fails on fail or less (Hazardous: 1 or 2, one mortal wound, three for a MONSTER or VEHICLE; ev.js hazardPer now goes through it, the same numbers), ld is 2D6 against the unit's Leadership, failing under it (Dark Pacts), double fails on a double of 2D6 (the Thousand Sons' Psychic test, each Ritual). Its cost is P(fail) × the mean of mw, each time it uses the rule (ev.js selfTestFail, selfTestCost; a line's selfTest, added to its hazard, when opts.selfTests names them). In the calculator a chip, in play while the rule's dice effects are (a Pact taken, a Ritual cast), and a line under the answer: Leadership 6+ fails 10 times in 36, D3 mortal wounds, 0.56 a volley, once for its shooting and once for its fight. Not in the live rankings (no army rule is).
• Resource pools (T-549). A rule that lets a unit replace one of its rolls with a chosen result once a phase (Acts of Faith's Miracle dice, Strands of Fate's Fate dice) is the mod swapSix, valued by one rule only: the expected gain of turning the volley's worst die of a roll into a 6, on the roll where that gains most (ev.js swapGain). For n dice of a roll each failing with chance q, a failure happens with chance 1 − (1 − q)^n, and the 6 adds one success of that roll carried through to damage: on the hit roll a critical hit (Lethal Hits: a wound at once; Sustained Hits: its extra hits), on the wound roll a critical wound (Devastating Wounds: past the save), each times the damage of an unsaved wound; the target's own swap turns its first failed save of the volley into a pass. Per volley: unitVsUnit gives the attacker's swap to the weapon and roll that gain most and the target's to the save of the weapon that costs it most; the calculator's dice put it on the volley's first failed roll of the roll sixStage picks by the same rule (the target's on its lowest save). A pooled die is rolled when it's gained, so the 6 is its best case: an upper bound, no other constant. A pool spent on buffs (Pain tokens, Yield points, the Blessings' eight dice, the Cabal's rituals) keeps the coverage data's hand pick; its rules carry pool (kindfx.json) and their chips say they need the pool. Not in the live rankings.
• Marks on an enemy unit (T-548). Condition chosen: the effect needs the target to be the unit the attacker's rules marked (Oath of Moment's target, a Farseer's guided unit, a Norn's prey, a hexed or judged unit, a quarry). In the calculator the defender's card carries a Marked chip while the attacker has such a rule; it starts off, Oath of Moment's included (T-554, Jordan 2 Oct; T-548 had started an army rule's mark on), the chip deciding; tapped, it holds for that pairing this visit. Condition codex ({ dets, not }): Oath of Moment's +1 to wound by the 2 Oct errata, in play with one of the Codex detachments (the codex's and the Faction Pack's, as Wahapedia's 11th-edition Space Marines page lists them; none picked counts as Codex) and no Blood Angels, Dark Angels, Deathwatch or Space Wolves unit in the attacker's list (or army). kindfx.json patches 42 effects to chosen (any other words of their condition stay the player's to vouch for). Where the target isn't known (the tools, the list comparator, the tier test), both stay the player's call, as before; the tier test keeps Singular Purpose in tierfx.json. Not in the live rankings (they load no kindfx.json).
• Damaged X (T-544). The datasheet's own threshold, read from BSData by bsdata.js (damaged: the Orks' core rule "Damaged" with X appended to its name, or an older datasheet's "Damaged: 1-X wounds remaining" profile whose effect is -1 to hit; brackets that only halve attacks or lower OC aren't taken; 481 datasheets). In the calculator a chip on each side whose unit has it (on; switching it off drops it); the attacker's and the defender's own attacks take -1 to hit (inside the ±1 cap; psychic attacks and "ignores hit modifiers" skip it) while the squad's last model has X wounds or fewer, so it bites in Fight back's reply and To the death's later rounds, never in a first volley. The coverage's hand-set damaged brackets (unitfx.json, off) aren't shown for a unit that has it, unless the coverage's bracket is something other than -1 to hit (Skarbrand's extra attacks): then the generic reading yields to it (detfx.js ownDamaged). ev.js damagedAt(W, X, T) = T × (W − X) ÷ W is the turn-time it's damaged from under typical fire; the tier test's kdmg (--kinds) scales each later turn's damage of a single-model unit by what -1 to hit leaves against each target. Not in the live rankings (one volley at full wounds).
• The heal's closed form (ev.js healLift, healPerTurn). A unit with W wounds in all that the enemy's fire destroys in T turns (the same damage D = W ÷ T each enemy turn) and that gets h back in each of its own turns dies in the first enemy turn k with W − kD + (k − 1)h ≤ 0, so k = (W − h) ÷ (D − h), or lasts the game when h ≥ D; its durability is ×k ÷ T, both at most 5 rounds. h is n's mean, at most one model's wounds less one in a squad that returns no models (a lone model's wounds); a kill heal is n × the models its weapons slay in a fight × the share of its turns it fights. The live Defense uses it since T-556 (Command-phase heals to itself, above); the tier test, behind --kinds, also counts kill and fight heals and heals lent to squads nearby.
• Leader lift (live, T-062). A character's effects on each squad it can lead, against MEQ, by Offense's rule. Its best squad's gain feeds Utility's support part. BSData's Supporting links aren't read yet; T-532 fixes that.
• Detachment synergy (live, T-058). Each detachment's effects on each unit of its army, against MEQ. The units it lifts most are listed on the detachment's page.

5. The best-loadout solver (live, docs/loadout.js)

• Legal squads only. Every squad of the unit's smallest size that BSData's wargear choices and caps allow, such as "1 per 5 models" or two weapons that can't go together.
• One kit at a time. A kit is what one model carries. It's scored once into each target, and a squad is the sum of its kits.
• Three answers:
  • the best squad into each target;
  • the best average share of each target's best;
  • the all-rounder, whose weakest share is strongest.
• Limits: 400 kits per kind of model and 200,000 squads per unit and size; past these, the best of those checked.
• Hazardous is priced at the squad's points per wound, one model's share rather than the whole unit's.
• Not cleanly read: a unit whose best squad can't match its own default wargear into every target has choices we didn't read right, and keeps its default.

6. Tier system v2 (proposed, not live)

Tested in the Lab: reports/tiers/, made by tools/tiertest.js, tools/banner.js, tools/archetypes.js. The letter and the job tag would be what counts. Offense and Defense stay the published raw pair under any version.

One army until the variables are found (Jordan, 2 Oct, via the supervisor): "For the purposes of ironing this out, stick to Tyranids until we can identify all the variables." The tier work now iterates on Tyranids alone, judged by the Auspex panel list, Tyranid list share and Jordan's own reading (he plays Tyranids, so he can say when a number is wrong); each missing variable is named and modelled one at a time until the Tyranid table is explained. The Space Marine and Ork panels run only as a check at the end of a round, never to steer it.

The four jobs. The names are Jordan's, our own words, not GW keywords. Each job is a percentile over the pool.

• Hammer, hitting:
  • V1: Offense.
  • V2: damage before death, damage per turn × turns alive under typical fire, at most 5. Turns alive: each attacker's hits to destroy over half the median unit's, so the median unit lives 2 turns; at least 1.
  • V2b: damage delivered before death (below).
• Anvil, holding:
  • V1: Defense's hits to destroy per 100 pts × OC.
  • V2: Defense per 100 pts × OC per 100 pts × the chance to pass Battle-shock (2D6 against Leadership) × 1.15 for units that heal or return models.
  • Known problem: Battle-shock weighs too much here (T-530), because it multiplies all of a unit's durability.
• Runner, the secondaries (V2): the mean of four percentiles: reach, units per 100 pts, a small footprint, and Leadership.
  • Footprint is models × base size: infantry 1, mounted 2, walkers 3, light vehicles 4, monsters and battle tanks 5, aircraft 6, Titanic 12; swarms count double.
• Banner, enabling others:
  • V1: today's Utility parts, weighted.
  • V2: a Leader's lift, auras (up to 2), Transport and revives, per 100 pts.
  • T-527's method: the score from section 7.

Composite and letters.

• Composite = 0.75 × the best job + 0.25 × the mean of the jobs.
• A unit's job is its best one, plus a second where another scores 80 or more. Its tag reads "S · Hammer".
• Letters go by quota on the composite's rank: S the top 7%, A the next 15%, B 28%, C 30%, D the rest.
• S and A are split across the jobs by their share of the winning lists' points: lists taking the unit × its points.

V2b's delivery model (tools/tiertest.js V2B). A five-round game on a 44" × 60" table.

• Movement: each turn the unit moves (+2" with Fly; a unit still out of range advances 3.5" more).
• Where guns stop: a gun unit stops at its longest range, or at half range when that's worth over 15% more (melta, Rapid Fire). A melee unit keeps coming.
• Damage each turn:
  • its ranged damage when in range (the half-range figure inside half range);
  • its melee damage × the chance of making the charge (2D6 against the gap left);
  • Overwatch.
  • Each turn is weighed by the chance the unit is still alive.
• Detachment credit: a character carrying the keyword a detachment's effects need shares that detachment's best lift with the army's other characters carrying it, up to 1.5 support points. It applies only where 6 or fewer characters carry it.
• Result in test: V2b didn't help (T-524 round 2). Damage per turn still dominates, and assuming every Infantry can ride lets slow elites that can't use most transports reach the fight too early.

V3: delivery modes, the bundle and the round-3 fixes (T-538, proposed; Jordan, 1 Oct (Pacific), via the design session). tools/tiertest.js V3. V2's four jobs; Hammer is damage delivered before death, as in V2b, but each unit is scored under every way it may legally reach the fight and keeps the one with the best composite. The unit page would say which: "S · Hammer, by Deep Strike" or "Best delivered by Rhino (+34% Hammer, 215 pts together)".

• Walk: on the table from turn 1. Scouts move first; Infiltrators start closer.
• Deep Strike: off the table turn 1, arrives turn 2 just over 8" away (the 11th-edition rule: more than 8" from every enemy unit). It shoots at once if its guns reach. A charge from there needs 9 on 2D6, 28%; no re-roll is assumed.
• Strategic Reserves: off turn 1, arrives turn 2 within 6" of a table edge and more than 8" from the enemy (the 11th-edition core rules, read 2 Oct). The 50% points cap on reserves is a list-level limit: noted, not enforced per unit.
• The bundle (the transport rule): the unit plus its best legal Dedicated Transport, priced at both together. Jordan: "when you're buying Berserkers you're also paying for a transport, and when you pay for the transport you also get its guns and its mobility." Its terms:
  • delivery: the pair moves at the transport's Move on turn 1, and the unit is out and on its own from turn 2;
  • protection: the unit can't be shot for V3C.SHIELD round inside, so it keeps that turn of its life;
  • the transport's guns: they add to the pair's Hammer (Firing Deck isn't counted yet: BSData gives the number only in text);
  • the transport's body: its hits to destroy and OC add to the pair's Anvil;
  • cost: everything per point of the pair.

If the bundle doesn't beat the unit alone, the transport isn't part of the unit's story. Who may ride what is read from BSData's capacity lines at run time (tools/transports.js): the capacity, the keywords a passenger needs, the models it can't carry, and the ones that take two or three spaces. Nothing is stored. Only Dedicated Transports form bundles: a Land Raider or Battlewagon is a unit in its own right, scored alone, and bundling it handed every passenger the big tank's guns. Characters aren't bundled: they ride with the squad they join. Allies are skipped for now.

• Survival: only turns the unit can be shot use up its life (V2's turns alive). A unit off the table or inside a transport keeps it. No terrain model: every shot is taken to have a line (an exposure factor of 1), to be calibrated against the panel.

The standard table (distances, from the 11th-edition Chapter Approved 2026-27 mission deck and core rules on Wahapedia, read 2 Oct; no terrain):

The round-3 fixes (each switchable in the tier test, measured on its own):

• (a) Mortal wounds are damage. Abilities that deal mortal wounds outside an attack add their expected mortal wounds a round to Hammer: reports/tiers/2026-10-02-mortals.json, 34 abilities read on 11th-edition datasheets, in our own words. Devastating Wounds was already in ev.js.
• (b) Characters that join a squad aren't Runners. A unit with Leader or Support links, and a unit with no OC, scores 0 as a Runner. Its value is in Banner, as part of the squad.
• (c) Battle-shock only bites below half strength. Anvil's Leadership factor is 1 − V3C.BELOW × (1 − the pass chance), not the pass chance.
• (d) Transport capacity in Banner. Tested and left out: it lowered every yardstick.
• (e) Indirect Fire is in range from turn 1.

Result (2 Oct, T-524 round 3). V3 (every term but d) fits list share at ρ 0.164: V2 0.127, V2b 0.087, V1 0.174. Within the army: Space Marines −0.18 (the least wrong of every system), Orks 0.18. Under V3, 579 units walk, 220 Deep Strike, 233 use Strategic Reserves and 129 ride a Dedicated Transport. reports/tiers/2026-10-02-round3.md.

The Context rule (T-539, a design rule for v2; Jordan, 1 Oct (Pacific), via the design session).

• The letter is the unit's best legal context, because that's what gets fielded:
  • its best loadout (today);
  • its best delivery (V3);
  • its army's best detachment for it: the detachment that most lifts its best job, defensive effects included;
  • under the Current meta lens and the meta's disposition mix.
• Leaders don't inflate a squad's letter. Banner already credits the character; the squad page shows "with <leader>: +x%" as information.
• Always with the asterisk: one line under the letter, "S · Hammer, in <detachment>, delivered by <transport>".
• The floor as a range: next to the letter, the letter with no detachment, alone and under the Even lens, shown as "B–S". A wide range means "needs its build"; a narrow one means "good anywhere". The tier list sorts by the ceiling, with a toggle for the floor.
• Axes are shown one at a time, never crossed:
  • a unit page has a small table of its letter per detachment of its army (precomputed) and per disposition (recomputed in the browser from stored cells);
  • each axis is computed against the others' defaults;
  • a specific crossing is the calculator's job.
• The same cells give pages per axis: "best units for <detachment>" on each detachment page, and "best units for <disposition>".
• On /about/ when v2 ships: "we score the build you'd actually field, we say what that build is, and we show how far the unit falls without it."
• Cost, measured 2 Oct: 1,761 units × their armies' detachments with effects = 24,999 cells, plus 8,805 disposition cells, plus 21,132 archetype cells at 6 + 6.
  • Today's detachment pass (one target) runs 16,329 cells in 1.4 s. Five targets would add about 11 s, taking the rankings build from 28 s to about 40 s, under a minute.
  • rankings.json would grow from 356 KB to under 400 KB gzipped.

V2c: V2b with round 3's six steps (T-524 round 3, 2 Oct, tested, not live). tools/tiertest.js --round3; the report is reports/tiers/2026-10-02-space-marines.md, "Round 3". Each step was measured on top of the last:

• (a) Mortal wounds are damage. Devastating Wounds were already in ev.js: a critical wound becomes mortal wounds with no save, and Feel No Pain still applies. New: data/mortalfx.json holds every 11th-edition ability that deals the enemy mortal wounds without a weapon. Each entry is the expected mortal wounds per use against each target, when it fires (move, shoot, charge, fight, react, arrive) and its reach. The delivery model adds them, per 100 pts, on the turns they can fire.
  • MW.NEAR = 1.5: an ability that hits every enemy unit within a few inches catches 1.5 units once it's in the fight.
  • MW.REACT = 1/3: one that answers being fought fires on a third of the turns the unit is in combat.
  • MW.RECHARGE = 0.5: one that fires on a charge fires again on half the later turns it fights, as it charges a new target about every other turn.
• (b) Characters who join a squad (Jordan, 2 Oct):
  • A Support character (one that joins only through Supporting links) can't take Runner or Anvil as its job. Its value is Banner plus its own attacks inside the squad.
  • A Leader may be scored alone, since a Captain or a Lieutenant can run missions on his own. But Runner now counts survival: Runner = (reach + turns alive + Leadership + V2C.RUNNER_HALF × (units per 100 pts + small footprint)) ÷ 4, each a percentile.
    • V2C.RUNNER_HALF = 0.5, so cheapness and footprint together can't carry Runner.
    • Turns alive is Hammer's own term, at most 5. A Lone Operative's counts V2C.LONE = 2 times, because it can only be shot from close by.
  • Lone Operatives keep every job.
  • A unit with OC 0 can't be an Anvil.
• (c) Battle-shock bites only below half strength. Anvil's Leadership term is 1 − b × (1 − the chance to pass). b = (min(5, T) − T ÷ 2) ÷ 5, never below 0, is the share of a five-turn game a unit that lives T turns spends below half strength: it loses about 1 ÷ T of itself a turn. T is turns alive, uncapped.
• (d) Banner per point, beats bodies, for every unit. Banner = a Leader's multiplier on its best squad (tools/banner.js's l) + auras + transport capacity + the detachment credit.
  • An aura (at most 2) lifts V2C.AURA_UNITS = 2 units. Each is lifted by the median Leader's lift on its squad (32%) and is worth the median unit's points (120), all against the bearer's points.
  • A transport's capacity is what riding adds to the best non-character squad of its army it can carry and that passes the bundle test (below), against the same points of that squad. At most 2.
  • Revives leave Banner. They're the unit's own durability, already 1.15 in Anvil.
• (e) Indirect Fire weapons are in range from turn 1. No change in practice: 62 of the 70 units with one already reached the 30" gap.
• (f) Delivery fixed:
  • Durability per point. An army's fire spreads over the enemy roughly by points. Turns alive therefore come from Defense's hits to destroy per 100 pts against half the median unit's (the median unit still lives 2 turns; between 1 and 5), not from the unit's whole durability.
  • Exposure. A unit uses up its life only on turns it's on the table outside a transport.
  • Who rides. Two rules, each with its own job:
    • May it ride (legality): a Dedicated Transport of its army can carry it, by capacity, the keywords it takes, what it can't carry and what takes more room (tools/transports.js reads these from BSData at run time). This is what players buy for a squad.
    • Does it ride (Jordan's bundle test): the pair (passenger riding plus transport) must beat, per point, both the passenger walking and the transport alone, on Hammer's damage delivered before death per 100 pts. Riding changes nothing else we score. The test is what stops a passenger inheriting a letter from a transport that's better on its own (Scouts in a Land Raider Crusader). The legality list alone doesn't.
  • V2C.RIDE_COST = 1 is the first cut this replaced (a transport costing no more than its passenger). It's kept only to count what it would have allowed.
• Result:
  • Space Marine ρ with the public list −0.24 → −0.08, distance 1.57 → 1.43.
  • ρ with list share 0.087 → 0.137.
  • Orks 0.06 → 0.01.
  • Mortarion C → S.
  • Most of the Space Marine gain is durability per point. The report has every step.

V4: the two strands merged, Runner's points gate, solo characters and step g (T-524 round 4, 2 Oct, tested, not live). tools/tiertest.js --round4 (inside --round3's run); the report is reports/tiers/2026-10-02-space-marines.md, "Round 4", and reports/tiers/2026-10-02-tyranids.md. Every part is a switch, measured by leaving it out.

• From V3: the arrival modes (walk, Deep Strike, Strategic Reserves, the Dedicated Transport bundle) and the constants V3C. The bundle now has to pass Jordan's beats-both-parts test: the pair's Hammer per point must beat the passenger's best way alone and the transport's best way alone. The unit keeps the mode with the best composite.
• From V2c: durability per point, exposure only while on the table, mortal wounds from data/mortalfx.json (absolute, on the turns they fire), Battle-shock damped, the Support/Leader rule (with survival in Runner), Banner per point (V2c's raw Banner), and Indirect Fire from turn 1.
• No charge after an advance (11th-edition core rules, read 2 Oct: a unit that advanced can't declare a charge). A melee unit advances only when a normal move leaves it more than V4C.CHARGE_REACH = 12" away, the farthest a charge can reach.
• Runner's points gate: a mission unit must be cheap enough to spend. Runner's raw mean × a factor of 1 up to V4C.GATE_FULL = 75 pts (tools/hill.js's Light bound), falling in a straight line to 0 at V4C.GATE_ZERO = 260 pts (twice its Middle bound, 130), applied before Runner is ranked. That's 86% at 100 pts, 70% at 130 and 32% at 200. Mortarion (375 pts) scores 0 as a Runner.
• Solo characters pay for being alone: a single-model character that joins squads (Leading or Supporting links), scored on its own without Lone Operative and not a MONSTER or VEHICLE, is the enemy's first target. Its turns alive × V4C.SOLO = 0.5, in Hammer's delivery, Runner's survival and the Battle-shock share.
• Step g, Hammer as enemy points destroyed, weighted by the meta (Jordan, 2 Oct, via the supervisor):
  • Each target's damage delivered before death counts as enemy points destroyed: wounds × the stand-in's ppw (tools/targets.js), per 100 pts of the attacker.
  • Target classes: the meta's weights are the points share of winning lists' units (lists × points, data/inclusion.json) by the class each unit resembles:
    • vehicles, monsters, walkers and aircraft are a heavy vehicle with V4C.HV_W = 14 or more wounds, and a light one below that;
    • every other unit is chaff with 1 wound, TEQ with a 2+ save or 3 or more wounds, and MEQ otherwise.
    • On 2 Oct the weights are GEQ 10%, MEQ 6%, TEQ 38%, light 22% and heavy 23%.
  • Hammer = 0.75 × the peak target + 0.25 × the meta-weighted mean. The peak counts a target in full when its class holds at least an even share (20%), and in proportion below that.
  • The tag: "S · Hammer (anti-tank)", "(anti-elite)" for MEQ or TEQ, "(anti-chaff)", from the peak's class.
• Result (2 Oct):
  • Space Marine ρ 0.04 (V2c −0.08, V3 −0.18), the first positive one, with both distances at 1.33.
  • Tyranids 0.19 (both strands 0.21).
  • Orks −0.19 (V3 0.18).
  • List share 0.104 (V2c 0.137, V3 0.164).
  • Mortarion S (V2c) → B, as a Hammer, not a Runner. The Lieutenant with Combi-weapon stays C. Eradicators D → C (anti-tank).
  • The Chaplains stay B or A: now Banner carries them, not Runner.
  • A panel list's ρ has a standard error of about 0.15 to 0.19 (30 to 45 units), so most single-panel differences are noise; list share (1,105 units, about 0.03) is the reliable yardstick, and V4 is below both strands on it.

V5: eight variables, Tyranids first (T-524 round 5, 2 Oct, tested, not live; Jordan's variables, via the supervisor). tools/tiertest.js --round5 (inside --round4's run) with --colist (tools/colist.js's counts of units listed together in winning Tyranid lists, grimstat-corpus, CC BY 4.0); the report is reports/tiers/2026-10-02-tyranids.md, "Round 5". V4 plus eight switches, each measured alone and stacked, on the Tyranid units only. Each run ranks the Tyranid units against one all-units run at its scale: the switches that rescale every unit (1, 3, 9) rescore the pool; the others are Tyranid rules ranked against V4's pool. With every switch off it reproduces V4 exactly.

1. Survival absolute (Jordan: incoming fire from a typical army turn doesn't scale with the unit's cost). Turns alive come from each attacker's hits to destroy the unit, over half the median unit's, between 1 and 5, as V2's, not from hits per 100 pts. Damage stays per point. It's used in Hammer's delivery, Runner's survival and the Battle-shock share; Anvil stays per point.
2. Arrival rules (data/tierfx.json arrival). A unit that goes back into Strategic Reserves at the end of its own Fight phase when unengaged (the Red Terror) is off the table in the enemy turn. Its fight ends unengaged when the charge failed or the target died: its melee wounds against the target's total, the five targets' mean. A unit that goes back at the end of the enemy turn (Raveners) has already been shot; it re-arrives only when that beats staying, which in our game it never does. Both reach any objective each round, so Runner's reach part is 100.
3. A mortal-wound attacker and the meta's incoming fire. A sixth attacker, mortal: V5C.MW_D = 2 mortal wounds a hit, no save. Feel No Pain counts, and Feel No Pain against mortal wounds too (tierfx.json proposed; the Norn Emissary's 4+).
   • The six are weighted by the meta's incoming fire: every weapon on the default wargear of every unit in winning lists, weighted by the lists taking it.
   • Each weapon is sorted by its profile: Damage 3.5 or more, or Strength 10 or more, is anti-tank; Damage 2.5 or more is elite; Damage 1.5 or more is plasma; AP or Strength 5 or more is special; the rest is small arms.
   • Each is weighed by its expected damage into the five stand-in targets. Devastating Wounds' critical wounds and data/mortalfx.json's mortal wounds go to mortal.
   • On 2 Oct: small arms 8%, special 20%, plasma 37%, elite 7%, anti-tank 22%, mortal 6%.
   • The weights apply to turns alive and to Anvil's Defense mean.
4. Enemy debuffs and Battle-shock pressure (tierfx.json enemy, shock), into Banner as points lent per point of the unit.
   • A −1 to hit on enemies near it touches V5C.NEAR = 50% of the attacks on the unit itself (its own Defense) and on its 2 nearest friends. It's worth 2 × 50% × (4/3 − 1) of the median unit's points; a −1 to wound below half strength adds the below-half share.
   • A forced Battle-shock test fails at the meta's rate: the Leadership of winning lists' units, 36% at no modifier.
   • Each Battle-shocked enemy unit-round is worth V5C.SHOCK = 10% of the median unit: its OC, stratagems and actions for one round of five, about half its round. That's ×V5C.SAB = 2 while the Neurolictor's aura is near (−1 to hit them, +1 to wound them).
   • An army-wide test made one harder (the Neurotyrant's Psychic Terror) counts over V5C.ENEMY_UNITS = 10 enemy units.
   • The rounds a unit spends in range come from its V4 delivery profile. A Lone Operative working from 12" or more counts its turns alive × V2C.LONE.
   • A modelled aura replaces V2c's stand-in for it (tierfx.json replaces).
5. Buff auras through the engine (tierfx.json aura).
   • The two squads most often listed with the bearer in winning Tyranid lists that it can lift at all (tools/colist.js) are run through docs/ev.js with and without the aura. A squad's lift is on its whole output: the better of shooting and melee into each target.
   • Each squad's lift × its points is lent. An aura on the unit the bearer shot (the Exocrine, the Psychophage) counts V5C.MARK = 50% of two friends' attacks.
   • Synapse (+1 Strength in melee) is credited the same way to every SYNAPSE unit, shared over the Synapse units an average winning list carries (5.7 on 2 Oct).
6. Leader pairs: each squad's best leader and its lift on the squad's damage (leaderfx.json; the Hyperadapted Raveners' from unitfx.json). Shown, never in the letter (the Context rule).
7. Detachment context (the Context rule). Each unit is scored under each of its army's detachments (detfx.json) and keeps the best: the ceiling. The floor is no detachment under the Even lens (no meta weights).
   • Dice effects whose condition the engine can't check count for a share of the game: within Synapse range V5C.SYN = 75%; below half strength V3C.BELOW; below Starting Strength the same, for units of more than one model; a Battle-shocked target 20%.
   • A pick made each battle round counts each option for a round in five.
   • Movement and OC rules come from a rule's tier key: arrival "tunnel" (reserves arrive V5C.TUNNEL = 7" away), "advCharge" (charge after an advance), "ds" (Deep Strike for who); oc, protean, upgrade, regen, surge. The Tyranid coverage card (T-541) wrote them under a rule's move key, which V5 doesn't read: renamed to tier they move 15 Tyranid letters and take V5's ρ with Auspex from 0.38 to 0.28 (tunnels and Crusher Stampede's OC most), so the constants behind them come first (reports/tiers/2026-10-02-tyranids-coverage.md). Round 6 reads them (mv, below).
   • Units' own abilities (--own-out, T-541, a report only): each Tyranid unit's dice effects from unitfx.json in every context. Effects marked off are left out, conditions are counted at the shares above (an objective 50%), and Singular Purpose, the enemy-side rules and Feel No Pain against mortals are left to tierfx.json. On 2 Oct it moved one letter (Venomthropes C → B) and left ρ at 0.38.
   • Singular Purpose (the Norns, tierfx.json chosen) runs here because the Talons of the Norn Queen's swap needs it. It's re-roll hits and wounds into the best chosen target all game, or OC 15 and Feel No Pain 5+ on an objective all game; with the swap, half each.
8. Small fixes.
   • A one-use unit (tierfx.json oneUse) or an immobile one (Move "-", not AIRCRAFT) can't be a Runner or an Anvil.
   • A CP is worth V5C.CP_PTS = 15 pts, so CP gained, saved or cost to the enemy count in Banner. A CP-cost aura replaces its stand-in.
   • A unit that spawns another instead of shooting (Biovores' Spore Mines) spawns on turns 1 to 4 and shoots on turn 5; the better of that plan and shooting counts. It adds V5C.SPAWN = 50% of the spawned unit's own Runner for each round one is on the table.
9. Banner's smooth start (found this round). Most units lend nothing, so any positive Banner raw ranked above all of them: a 0.04 jumped to the 57th percentile. Below the median positive raw, Banner's percentile now rises in a straight line from 0.

• Result (2 Oct, Tyranids):
  • ρ with Auspex's order 0.19 → 0.38; the same letter 14 → 16 of 44, within one 26 → 31; ρ with Tyranid list share 0.12 → 0.24.
  • Today's letter: 0.65, 21, 37.
  • Alone, the largest gains are the small fixes (ρ 0.26, list share 0.19) and the detachment context (ρ 0.25). Survival absolute lowers ρ (0.15) but raises the letters (16 and 29).
  • The end-of-round check: Space Marines ρ 0.04 → −0.01, Orks −0.19 → −0.12, list share 0.104 → 0.129.

Engine kinds on V5 (T-540, --kinds, tested, not live). The kinds the dice engine now plays, read from the coverage files with kindfx.json's numbers, switched in on top of V5 one at a time and cumulative, on the Tyranid units against round 5's pool; without --kinds V5 is unchanged. The report section is "Engine kinds" in reports/tiers/2026-10-02-tyranids.md.

• heal: a unit's own heal lifts its durability by ev.js's healLift on the turns alive it has without it (so turns alive and Anvil both rise), and replaces the flat ×1.15 revive for that unit; a unit with the revive trait and no heal in the data keeps ×1.15. A kill heal counts the models its weapons slay in a fight (ev.js, the five targets' mean) × its engaged rounds over its turns alive. A heal on units near it (the Tervigon's Termagants) is lent in Banner like an aura: the lift on the two squads most often listed with it that it reaches, × their points.
• aura: variable 5 reads a Tyranid bearer's buff auras from unitfx.json (range, to, not; effects from the aura object or the rule), each through the engine on the two co-listed squads it reaches; tierfx.json's aura stays for the marks (the Exocrine, the Psychophage).
• mortal: V4's delivery reads a unit's mortal wounds outside attacks from its numbered coverage rules (after a Fight phase counts as once a round in range, as mortalfx.json had the Harpy; Battle-shock and Fall Back triggers left out); mortalfx.json stays for every unit without them.

Every army scored (T-552, --every, tested, not live). Round 5 scored only the Tyranids with their own abilities, kinds and detachments; the pool's other units were bare, so every percentile and quota was set against bare units. With --every each unit of the pool is scored as a Tyranid is, four switches (ev { own, kinds, army, det }), cumulative: its own abilities (unitfx.json, a chapter's from the Space Marines' entry; effects marked off left out; the same condition shares as round 5's variable 7), the engine kinds (heal, mortal through mwkOf, aura on the squads its army's winning lists take most where no co-list counts exist, Damaged X, T-545's vocabulary), its army rules (DetFx.armyRules, a chapter's with the Space Marines' too; Synapse's pilot shape, an aura from a keyword, stays variable 5's; Oath of Moment's Codex +1 to wound only outside the excluded chapters; a rule with a pick takes its best option per target and per attacker, as a leader's options are taken; swapSix goes through ev.js's swapGain) and its army's best detachment (the Context rule: each of the army's detachments detfx.json has, the best composite against the bare pool at the same scale). No constant changes. --own-kinds adds the Tyranids' own abilities and every kind as one row of the Engine kinds table; --v5dump <file> writes every unit's V5 letter, job, jobs, percentiles and detachment. --tyr4-out now keeps hand-appended sections when it regenerates (writeKeeping). Results: reports/tiers/2026-10-02-tyranids.md, "Every army scored"; reports/tiers/2026-10-02-v5-every.json.

Round 6: Jordan's settings and the movement rules (T-557, --round6, tested, not live; Jordan, 2 Oct, via the supervisor). Four switches on top of the every-army run (T-552's end state), each measured alone and cumulative: cp0, tmax, pos, mv. Report: reports/tiers/2026-10-02-tyranids-round6.md.

• A CP is worth V6C.CP_PTS = 30 pts for now (round 6b, Jordan, 2 Oct; round 6 ran it at 0). A CP gained, saved or cost to the enemy counts 30 pts in Banner, per point of the unit as every lent worth. A CP-cost aura still gives up its V2c stand-in. T-558 is to replace this with a CP's worth taken from the stratagems each unit can use, measured with the engine. --r6yard <file> writes the yardsticks (round 5, round 6, today's live letters) as JSON. Report: reports/tiers/2026-10-02-tyranids-round6b.md.
• Tunable constants, kept as they are and to revisit if they count too much or too little (Jordan, 2 Oct):
  • Battle-shock's worth, V5C.SHOCK = 10% of the median unit for each enemy unit-round Battle-shocked.
  • The enemy debuffs' share, V5C.NEAR = 50% of the attacks on the unit and its two nearest friends.
• A mixed unit's Toughness (tmax). The tier test's own hits-to-destroy (htk5: the defensive effects' ratios, Singular Purpose's hold) uses the highest Toughness among the unit's kinds of model, as the live Defense does since T-546.
• The position track (pos): the same ways in as Hammer's delivery, followed round by round on the line from the unit's start to the enemy's (V3C.GAP):
  • walking: Infiltrators start V3C.INFIL from the enemy and Scouts move V3C.SCOUT first; each turn the unit moves its Move plus an Advance, V3C.ADV;
  • Deep Strike: turn 2, V3C.DS from the enemy;
  • Strategic Reserves: turn 2, V3C.RES from the enemy;
  • tunnels: turn 2, V5C.TUNNEL from the enemy, for either.
  • An objective X" out is held from the first round the unit stands at X or beyond. The unit keeps its best way in.
  • Anvil × (V6C.POS_FLOOR + (1 − V6C.POS_FLOOR) × the share of the game it holds the midfield objective, V6C.OBJ_MID = 15", and the one in the enemy's half, V6C.OBJ_FAR = 21"). V6C.POS_FLOOR = 0.5 is what an Anvil is worth without getting anywhere: its home objective and the nearest ground.
  • Runner's reach is that share over the two objectives and the enemy's deployment zone, V6C.OBJ_DEEP = 27", as a percentile over every unit. It replaces Utility's fixed reach (Move, +2 Fly, +6 Deep Strike, +6 Infiltrators, +4 Scouts).
  • Transports aren't on the track yet.
• The detachments' movement rules, read in full (mv). A rule's move key (the Tyranid coverage's, T-541) is read as tier always was, so its reach feeds three jobs through the closed forms above, not Runner alone:
  • Hammer's turns in reach: tunnels bring an arrival to V5C.TUNNEL; a Deep Strike grant adds that way in; a surge adds to Move; a charge after an Advance (Vanguard Onslaught) lifts V4's no-charge-after-Advance rule.
  • Anvil's position and Runner's reach: the same changes on the track.
  • The non-movement move keys are read too:
    • Crusher Stampede's +2 OC at Starting Strength, at the coverage's 0.9 share;
    • the Norns' re-pick;
    • Feed the Swarm, now through the heal closed form, EV.healLift: one regeneration a Command phase (wounds back on one model, an Infantry model returned, or up to 3 models of an Endless Multitude) on whichever of the two squads most often listed with the Harvester it helps most, × that squad's points, lent in Banner like the Tervigon's heal.
  • Only the Tyranid coverage writes move keys so far. Enhancements' movement (re-deploys, +2" Move) and Synaptic Nexus's +1 to Advance and Charge aren't written as keys yet.
• Result (2 Oct):
  • Tyranids: ρ with Auspex 0.37 → 0.34; same letter 13 → 12 of 44; within one 23 → 26; ρ with Tyranid list share 0.28 → 0.27.
  • Every unit's list share 0.155 → 0.169. Space Marines −0.06 → −0.03. Orks −0.14 → −0.16.
  • The position track alone gives 0.41 and 0.36.
  • On top of it, the movement keys hold (tunnels: ρ 0.40, list share 0.36). Crusher Stampede's OC (0.35) and Feed the Swarm (0.38) cost agreement.
  • Read in Hammer only, the same rules give 0.27.

Round 7: best-only grading (T-559, the r7 switch, --r7json <file>, tested, not live; Jordan, 2 Oct). Round 6b's run with three changes:

• Best-only grading (Jordan's rule). A unit's score is its best allowed job's percentile plus a bonus that can only add: for each other allowed job above the bar V7C.BAR = 60, V7C.BONUS = 0.25 × (percentile − 60); at most 100. It replaces the composite (0.75 × best + 0.25 × mean of four) in round 7. 92/10/10/10 scores 92; 80/80/80/80 scores 80 + 3 × 5 = 95; the Tyrannofex's 89/93/16/0 scores 93 + 7.25 (capped at 100). A weak job never lowers a unit. Allowed jobs are the ones a unit may hold (no Anvil at OC 0; Hammer and Banner only for a joining Support character; no Anvil or Runner for a one-use unit), as for the best job. Letters stay by rank quota on the score, S and A split by job demand.
• Feed the Swarm's lend capped. The healed squad keeps the regeneration's worth; the Harvester is lent at most what V2c credits an aura on V7C.REGEN_UNITS = 1 squad (the median Leader's lift on the median unit's points, 38 pts), one squad a phase. The healed squad's own Anvil doesn't count the heal yet.
• A one-use transport delivers once. deliver's drop: the cargo is off the table on turn 1 and set down on turn 2 at V3C.DS, as Deep Strike, then walks; a transport with tierfx.json oneUse (the Tyrannocyte) has its Banner capacity credit from that, not from V2c's ride on turns 1 and 2.
• Yardsticks, every round from now on: ρ with the panel's order and its standard error (Bonett and Wright's); strict on our ruler (the panel's graded units in its order, each letter a tied block, cut at our quota; a block that straddles a cut takes the letter most of it falls in; then the same letter as ours), with the raw same-letter count beside it; lenient (within one letter, raw letters); and beside the strict and lenient counts their shuffle baselines, the mean over 2,000 permutations of our letters against theirs. Ungraded units stay out.
• Result (2 Oct): Tyranids ρ with Auspex 0.39 → 0.50 (SE 0.12); strict on our ruler 19 → 20 of 44 (chance 10.3); raw 12 → 12; within one 27 → 29 (chance 20.4); Tyranid list share 0.32 → 0.39; every unit's list share 0.171 → 0.164; Space Marines −0.03 → −0.05, Orks −0.16 → −0.13. 291 of 1,161 letters move. Report: reports/tiers/2026-10-02-tyranids-round7.md.

Round 8: step grading on fixed cuts (T-560, the r8 switch over r7, --r8json <file>, --r8sliders <file>, tested, not live; Jordan, 2 Oct). Round 7's run (its Feed the Swarm cap and one-drop Tyrannocyte kept) with its grading replaced:

• Steps, no cap (Jordan's rule). Score = the best allowed job's percentile + V8C.ELITE_STEP = 5 for each other allowed job at V8C.ELITE = 90 or above, + V8C.STRONG_STEP = 2 for each at V8C.STRONG = 80 to 89, nothing below. Never capped (90/90/90/90 scores 105). Ties and the order inside a letter go to the mean of the other allowed jobs. 90/10/10/10 → 90; 90/70/70/70 → 90, ranked above it; 95/10/10 → 95 beats 85/60/60 → 85; 90/90/10/10 → 95; 90/90/90/10 → 100; Tyrannofex 93/89 → 95; Norn Assimilator 96/77 → 96; Hormagaunts 99/64 → 99.
• Letters, all armies (the default): fixed cuts on the score, S ≥ V8C.S = 95, A ≥ V8C.A = 88, B ≥ V8C.B = 80, C ≥ V8C.C = 65, D below; no quota, no job-demand split.
• Letters, within this army: the same scores ranked inside one army, lettered by position: top V8C.AS = 10% S, next V8C.AA = 15% A, next V8C.AB = 25% B, next V8C.AC = 30% C, the rest D (rounded to the nearest unit; an army of 10 or more units has at least one S).
• Yardsticks: strict on our ruler compares the within-army view with the panel re-cut at the within-army quota (a public tier list is a within-army view); the all-armies view keeps round 7's ruler; raw and lenient for both views, each with its shuffle baseline. The Lab's sliders page (/lab/reports/tiers/sliders/) re-letters the Tyranids live from the four bar and step values.
• Result (2 Oct): Tyranids ρ with Auspex 0.50 → 0.47 (SE 0.13); strict on our ruler 20 → 20 within army (chance 10.6), 16 all armies on round 7's ruler (chance 8.6); raw 12 → 19 all armies, 13 within; within one 29 → 31 all armies, 30 within; Tyranid list share 0.39 → 0.41; every unit's list share 0.164 → 0.162; Space Marines −0.05 → −0.18, Orks −0.13 → −0.12. 561 of 1,161 letters move (520 up): 246 units are S on the fixed cuts, against round 7's quota of 81. Report: reports/tiers/2026-10-02-tyranids-round8.md.

Round 9: Anvil divides by points once, a survival gate, three panels (T-562, the r9 switch over r8, --r9json <file>, --sweep, tested, not live; Jordan, 2 Oct, from the independent critique's points 4, 5 and 9). Round 8's run with Anvil changed:

• One division by points. Round 8's Anvil was Defense per 100 pts × OC per 100 pts, i.e. hits to destroy × OC ÷ points², so halving a unit's cost quadrupled it. Round 9: hits to destroy × OC ÷ points (Defense per 100 pts × the unit's OC), × the same Battle-shock, revive and position factors. Not the geometric mean of the two per-point terms: √(hits × OC) ÷ points is the square root of round 8's product and ranks every unit the same, so a percentile wouldn't move. The measure now grows with a unit's models (twice the bodies, twice the hits and OC, for twice the points), so large minimum units gain.
• The survival gate (removed in round 10). × 0.5 when the unit was removed in one typical enemy turn, × 0.8 in two, × 1 for three or more, on Hammer's own turns alive Tu. Round 10 took the gate and its constants out of the code; the r9 switch now runs round 9 without it.
• Constants named for the sweep (values unchanged): Hammer's weight on the peak target V9C.PEAK = 0.75 (step g's split); the flat revive lift V9C.REVIVE = 1.15; the closest a transport's cargo gets out to the enemy after the first turn's ride, V9C.SETDOWN = 3".
• --sweep: each hand-set constant at ×0.5 and ×1.5 (a share capped at 1; the bonus bars ±10 points; the letter cuts ±5, re-lettering only), alone, against round 9: letters moved over every unit (all-armies view) and the Tyranid ρ with Auspex. reset9 drops what was worked out with the old value between runs. Not swept: V3C.ROUNDS, V3C.GAP, V5C.MW_D and V4C.HV_W (read when the pool is built), and V2c's transport constants (worked out before round 5).
• The three-panel history: every version still runnable (V2b, V3, V4, round 5, T-552, round 6 at CP 0, 6b, 7, 8, 9), rerun in one run, against Tyranids (Auspex), Space Marines (Tista Minis) and Orks (Sprues & Brews), ρ with SE, plus every unit's list share; a chart beside the report. From round 9 every round's report leads with all three panels.
• Cut calibration (a proposal; the default stays S ≥ 95, A ≥ 88, B ≥ 80, C ≥ 65): the scores that give 8 / 15 / 27 / 30 / 20% of all units under round 9: S ≥ 98.79, A ≥ 94.49, B ≥ 86.13, C ≥ 71.15, to be set once and frozen as absolutes if Jordan takes them.
• Result (2 Oct): Tyranids ρ 0.47 → 0.49 (SE 0.13), Space Marines −0.18 → −0.10, Orks −0.12 → −0.12; every unit's list share 0.162 → 0.077; Tyranid list share 0.41 → 0.45; 13 Tyranid letters and 240 of 1,161 move. Anvil: Neurogaunts 99 → 89, Termagants 100 → 96, the Norns 96 → 100, Tyrannofex 92 → 96. The sweep's biggest movers are Runner's points gate (399 letters), the Advance (288) and the objectives' distances (286); a CP's worth moves 3. ρ between an army's S share and its top-quarter lists: −0.37 (round 8), −0.39 (round 9, current cuts), −0.24 (proposed cuts). Report: reports/tiers/2026-10-02-tyranids-round9.md.

Round 10: Anvil as OC-turns per point, a fieldable pool, the final rank on frozen cuts, the list-share headline, the explainer (T-563, the r10 switch over r8, --r10json <file>, --explain, tested, not live; Jordan, 2 Oct, from the method appraisal's experiments 1 to 4). Round 8's run (step grading) with:

• Anvil = OC × min(turns alive, V10C.TMAX = 5) ÷ points (per 100 pts) × Battle-shock × position × revive, no gate: objective-control turns bought per point. Turns alive are turns5's T (each of the six attacker classes' absolute hits to destroy against half the median unit's, held to 1 to 5, mixed by the meta's incoming fire; × 0.5 for a lone character), so survival stops counting at the game's length and a 3,500-pt Titan can't buy its way up. A transport bundle: the pair's OC and points, the squad's turns alive. The Norns' chosen-objective branch: OC 15 on the variant's own turns alive.
• The pool: fieldable units only. A unit over V10C.MAXPTS = 2,000 pts at its listed size can't be fielded in a 2,000-pt list and leaves the pool before any percentile (Legends were never in): 8 units (the six Titanicus Titans over 2,000 pts, the Phantom Titan, the Manta), 1,161 → 1,153. The medians fixed when the pool is built (half the median unit's hits to destroy, Runner's Leadership and footprint percentiles) keep the whole pool.
• The final score: the pool percentile of the composite (best allowed job + round 8's steps): the share of the fieldable pool scoring below it, 0 to 100. A monotone transform: no ρ moves (ties aside), and the best-of-four pile-up (369 units in the 90s, 50 above 100) spreads to about 115 a bin.
• Frozen cuts (calibrated 2 Oct on round 10's run, the lowest final score inside each share S 8%, A 15%, B 27%, C 30%, D 20% of 1,153 units): S ≥ V10C.S = 92.02, A ≥ V10C.A = 77.02, B ≥ V10C.B = 49.96, C ≥ V10C.C = 20.03, D below. Set once: a later round that lifts scores shows as more S. The within-army view is unchanged (position inside the army, round 8's shares).
• The list-share headline (from round 10 on): over the units some winning list could field (armies with 5 or more winning lists), the median of each army's ρ with list share (armies with 10 or more such units), the Tyranids' own ρ beside it (the overfit gauge), the AUC for units winning lists take against units they never take, the hit rate (taken among our top quarter, against the base rate) and ρ among taken units only. The pooled ρ stays as a secondary line.
• The explainer: node tools/tiertest.js --explain "<unit>" [--army <army>] [--out]: one unit's points, letters (both views), final score, four job percentiles, the raw quantities behind each (damage per 100 pts against the five targets, hits to destroy under the six attackers, OC, turns alive, position share, reach, Leadership, points lent by kind), best detachment and way in, list share and panel letter, beside its three nearest peers (same best job, nearest in points, its own army within ×2 first); then each raw quantity's biggest gap from the peers' median (a log ratio with a small floor per quantity). Derived numbers only.
• Result (2 Oct): Tyranids ρ 0.49 → 0.51 (SE 0.12), Space Marines −0.10 → −0.07, Orks −0.12 → −0.18; every unit's list share (pooled) 0.077 → 0.157; per-army median 0.113 → 0.165 (round 8 0.188); AUC 0.536 → 0.582; Tyranids' own ρ 0.45 → 0.46. Anvil's ρ with points 0.41 → 0.19 (the target was ±0.15). No unit over 1,000 pts is S. The pool cut moves every panel ρ by under 0.01. S per army against top-quarter lists −0.39 (round 9) → −0.16. Report: reports/tiers/2026-10-02-tyranids-round10.md.

The Tyranid workbench (T-564, tools/tierbench.js over --rawdump tyranids; a report, not live; Jordan, 2 Oct: no more every-army runs until the Tyranid table is within one letter of Auspex for 85% of the 44 graded units). The method appraisal's layers, built for one army.

• Layer 1, the raw table: node tools/tiertest.js --rawdump tyranids (about 35 s) runs round 10, then scores every Tyranid in every context tyrRun tries (no detachment and each detachment, the Norns' branches) at every datasheet size and with every leader BSData's canJoin allows (the pair as one unit: the leader's attacks inside, its leaderfx.json effects on both, its defensive effects on the squad, its wounds behind it, both OCs and costs), plus each leader's own attacks delivered with the pair's turns alive ("inside" rows), plus every pool unit's chosen context. It writes build/tier-raw-tyranids.json, reports/tiers/<date>-tyranids-raw.json and a flattened .csv. Columns: id, army, unit, ctx {det, branch, kit, size, leader, inside}, pts, models, allow {H, A, R, B}, per way in (ways): Hammer raw, damage per 100 pts into the five targets, inRange per round, Anvil raw, effective OC; hits to destroy per attacker class (htd), turns alive per class before the floor (x: hits to destroy ÷ half the median unit's), T, Tu, the solo factor, OC, position share (hold), reach per objective band, Leadership pass, below-half share, units per 100 pts, footprint, lends in points by kind, Runner's parts, Banner raw and its leader multiplier with the fixes' alternatives (banAlt, tools/banner.js's onSquad with lendOnly or models), list share, most-listed size (reports/tiers/2026-10-02-tyranids-sizes.json) and leader, the Auspex letter. Null: the kit (one loadout), the Battle-shock lend (inside the debuff lend), the spawn lend (Runner's).
• Layers 2 to 4, compose(raw, config): pure; rebuilds the pool's percentiles from the pool rows, scores each row by job5's rule (best allowed job + steps, the tie key), picks each Tyranid's best context, letters within the army and on the frozen cuts, and gives the yardsticks against Auspex (within one, the same letter on our ruler, raw, ρ with SE, the gaps). Config knobs: target.w, hammer.peak, fire.w, price.cp (price.shock, price.debuff inert), anvil.tmax, runner (half weight, the gate's ends, Lone Operative factor, spawn share), steps, ceiling, letters. With every fix off it reproduces round 10 exactly (all 52 Tyranids and every other unit's composite).
• The fixes (one per residual report's finding, default off): bannerLendOnly, bannerModelCount, turnsHarmonic (Anvil's turns alive and Battle-shock share as 1 ÷ Σ wᵢ ÷ turnsᵢ, no floor), reArriveAfterKill (tiertest.js goRe's alt), bestSize, bestLeader, reArriverGateExempt, auraToNeighbours, spawnShare75, offTableSurvival.
• First pass (2 Oct): baseline 31 of 44 within one (70%); all ten on 30 (68%, ρ 0.51 → 0.71); best subset 36 (82%); the 85% line is 38. Report: reports/tiers/2026-10-02-tyranids-bench-1.md.
• The bench's banOf (T-565): the Banner switches take off the stored multiplier (banLead), not today's banner.js value (banAlt.cur). --rawdump takes build/tier-raw-tyranids.lock and writes through a per-process temporary file renamed into place; --raw-out <file> writes elsewhere.

Round 11: the workbench's fixes made permanent, every army (T-565, o.r11 over round 10, --r11json <file>, tested, not live; Jordan, 2 Oct: a unit's context is its best legal size, turns alive capped at 5, a CP 30 pts). Round 10's run (V10C, the fieldable pool, frozen cuts) with each fix a switch in o.r11 (the code's R11F names them), all on, applied to every unit of the pool, not only the Tyranids:

• bannerLendOnly (lend): Banner's leader multiplier is tools/banner.js's onSquad with lendOnly (the character's own attacks left out), today's value replacing the --banner file's; the leader's own attacks get an "inside" context, its Hammer delivered on the turns alive of the pair with the squad that keeps it alive longest (at that squad's listed size). Its Anvil and Runner stay its own.
• bannerModelCount (models): banner.js's models, a multi-model character at its real model count.
• turnsHarmonic (harm, turnsH): Anvil's turns alive and its Battle-shock share are 1 ÷ Σ wᵢ ÷ turnsᵢ over the six attacker classes (no per-class floor or cap), capped at V10C.TMAX = 5 in Anvil (Jordan: the game's length). Hammer and Runner keep turns5's.
• reArriveAfterKill (kill): a fight-phase re-deployer goes back to reserves only after a kill (goRe's alt).
• reArriverGateExempt (gate): a unit that re-arrives each round skips Runner's points gate.
• lendOnBestSquads (best2): every lend (a data aura, tierfx.json's auras and marks, Synapse, a heal for others, the −1-to-hit shield, Feed the Swarm) is worth its lift on the two squads of the bearer's army it lifts most by points lent (lift × points), from every non-character unit of that army in the fieldable pool that it reaches. A defensive lift counts as the turns alive it adds inside the game, min(5, T × (1 + l)) ÷ min(5, T) − 1. Feed the Swarm is the mean of its two best squads (one regeneration a phase), round 7's cap gone. A unit whose lending is measured loses V2c's aura stand-in (and the CP block's stand-in swap); a unit with none keeps the stand-in, counted with the widened aura pattern (auraN11: "(Aura)", "(Aura, Psychic)" and the like). One rule in place of the benches' Hive Tyrant, Pyrovores and Maleceptor special cases.
• stealthSurvival (stealth): a unit with the Stealth trait takes 4/3 the attacks to destroy from each ranged attacker class (htk5's cover factor; Stealth gives the benefit of cover, T-131), unless its own abilities already give it cover. ev.js has the hook (a defender's cover), but the tier test's incoming fire is htk5's closed form, so this lives in the tier test only.
• bestSize (size): every size on the datasheet (sizeVariant: damage, durability and OC scaled per model), in every detachment; the best by the tie key.
• bestLeader (leader): each leader BSData's canJoin allows (pairVariant), at each size, in the squad's best unpaired detachment. The pair's Banner is the squad's plus the leader's lending other than its multiplier, per point of the pair. A leader takes the pair as its own context when it is a multi-model unit or (Tyranids) joins that squad in most lists that take both.
• sameLoadout (kit, a bug fix): an effect's ratio into each target is worked out on that target's best squad (rankings.json's lt/lw, the squad the damage uses) at the listed size, not on the default wargear (kitAt, through ratioOf and lendRatio).
• A context at another size or as a pair holds its Banner per point of its own cost.
• Result (2 Oct, every army): Tyranids ρ 0.52 → 0.66 (round 10 on today's data), Space Marines −0.07 → 0.14, Orks −0.19 → −0.15; per-army median 0.139 → 0.173, AUC 0.582 → 0.585, ρ among taken units 0.094 → 0.105; the Tyranid table within one 31 of 44 on the within-army ruler, 36 on Auspex's shape (the 85% line is 38, not reached); 537 all-armies letters move. Alone, sameLoadout lifts the Marines most (−0.07 → 0.10), bannerLendOnly the Tyranids (0.62); without bestSize the Tyranids would read 0.70 and 39 on the shape. Report: reports/tiers/2026-10-02-tyranids-round11.md.
• Charge distance (checked 2 Oct, T-566): in the 11th edition a charge target must be within the 2D6 roll, so a charge from a gap of d" needs a roll of d, not d − 2 (engagement range, 2", is where the move must end). pChg stands.

Round 12: compose-only experiments (T-581, tools/round12.js over tools/tierbench.js; tested, not adopted). Three new compose knobs, all at today's values by default (--check still reproduces round 10 exactly):

• cost.alpha (α, default 1): Hammer, Anvil and Banner become raw ÷ points^α. The raw holds raw ÷ points, so it is multiplied by (points ÷ 100)^(1 − α).
• cost.runner (default "both"): Runner's two points terms, the points gate (75 to 260 pts) and the cheapness part (units per 100 pts, act). "gate" keeps the gate and drops the cheapness part (Runner = (reach + survival + Leadership + half × footprint) ÷ (3 + half) × gate). "cheap" keeps the cheapness part and drops the gate.
• composite (default "best"): "allowed" takes each job's percentile among the pool units allowed that job; "log2" scores each job as log₂(raw ÷ the median positive raw of the units allowed it), the floor −8 at zero, the best allowed job with no steps.
• fixDet: the Tyranids scored in one detachment only (the option-premium what-if).
• Result (3 Oct): nothing passes the gate. With the gate dropped (the choice rule's pick), ρ(score, points) goes −0.16 → +0.10 pooled, but unseen agreement falls 0.27 → 0.16 and big models return as mission units (round 3's failure). With the cheapness part dropped it goes −0.16 → −0.10, agreement is unchanged (+0.01), and 116 letters move. α 0.75 alone is cost-neutral pooled. No knob makes the per-army lean neutral. Report: reports/tiers/2026-10-03-round12.md.

Round 13: the every-army raw table, the config file and compose over every army (T-583; tested, not adopted).

• Layer 1 for every army: node tools/tiertest.js --rawdump all records, while round 11 runs (o.r11 all on), every context it scores: the bare pool that picks each unit's context (base), every context everyRaw tries for every fieldable unit (p rows: detachments, other legal sizes, joined pairs from either side, a leader's attacks inside), every context tyrRun tries for a Tyranid (t rows: the Norns' branches too), and the joined pairs on every other plain context. Each row holds, per way in, Hammer's and Anvil's raws (and the damage per 100 pts into the five targets, v), Anvil's bare raw, turns alive, Runner's parts (reach, Leadership and footprint percentiles, units per 100 pts, the points gate, the re-arriver exemption), Banner, the jobs allowed, a spawn's parts, the engine's key, and features for the factor analysis (f: turns alive per attacker class, OC, Leadership, footprint, models, lends by kind). Full precision; build/tier-raw-all/<hash12>.json (the hash covers every input file, line endings normalised), written through a per-process temporary name; latest.txt names the newest. It also stores the engine's own keys for B0 and B1 as a check. At 374f505: 70,376 rows, 73.6 MB, about 40 s, peak memory 2.0 GB.
• tools/tierconfig.json: every knob of layers 2 to 4 with value, default, source, date, owner and layer (compose: true where compose reads it): runner.half, runner.gateFull, runner.gateZero, runner.lone, runner.spawn, cost.alpha, cost.runner, curve, bsmooth, composite, steps.*, ceiling.*, letters.cuts, letters.within, context.bestSize, context.pairs, tolerance; the layer-1 constants it bakes in (round 11's fidelity switches, the turns cap, the lends, a CP's price, the target and fire weights, Advance, the objectives' and reserves' distances, the revive, below-half, Synapse and near shares) under layer1 for provenance. Profiles: B0 (round 11 exactly) and B1 (Jordan's frozen conventions: minimum size, context.bestSize false; the turns cap and the lends on the two best squads are round 11's already).
• compose over every army (tools/tierbench.js composeAll, loadConfig(profile), loadRawAll()): the bare pool's percentiles pick each unit's context (ties: the engine's 1e-9), the chosen contexts make the pool's percentiles, a Tyranid takes its own pick under them (ties 1e-12), the final score is the pool percentile of the key, letters by the frozen cuts and the within-army shares. Knobs baked into layer 1 throw. Reproduces round 11's published JSON with 0 differences over 1,153 units and 52 Tyranids, and both stored engine runs with a key gap of 0; the battery rerun from compose is identical. About 130 ms a run.
• Result (3 Oct): frozen round 11's ρ(score, points) is −0.414 (Spearman, every fieldable unit); B1 −0.30. Round 12's experiments on B1, chosen by rules that read no panel: Runner's cheapness part dropped and α 0.6 take the lean to −0.01 pooled (16 of 29 armies outside ±0.15, from 24) and the unseen pooled ρ 0.33 → 0.36, a rise inside its interval: not adopted. Report: reports/tiers/2026-10-03-round13.md.

Army rules (T-530, proposed).

• Aura grants: a unit that grants an aura always has it, and one that receives it is in range 75% of the time.
• Effects: melee mods lift Hammer, more Battle-shock dice lift Leadership, and granting an aura counts in Banner.
• Not in letters yet: only when every army's rules are written, because the first written inflates its army.

Lenses (T-531, proposed). A lens is a weight vector over the stored targets and attackers. Letters recompute in the browser: 35 ms for 1,241 units at phone speed. The default lens is the one that best predicts what wins. The custom lens is a future supporter perk; the built-in lenses stay free.

Disposition lenses (T-537, tested, not recommended yet). tools/dispositions.js sorts each primary's VP by the job that earns it: holding an objective is Anvil; taking new ground, their half or their home is half Runner and half Anvil; an action is 70% Runner and 30% Anvil (it completes only if the unit still holds the objective); a kill is Hammer. The counts assume a five-round game: a line every round counts 5, from round 2 counts 4, rounds 2 to 4 count 3, once at the end counts 1. "Each objective" means 2 besides your home, or 3 with it; "each kill" means 1.5 a turn; "each unit that acted" means 1.5. A disposition's demand is its five primaries, weighted by how often the field plays the opposing disposition each one answers. The lens's S and A split gives each job half today's split plus half the demand, with Banner kept at today's share. The lens moves only the S and A quotas; scores don't change. The check failed (2 Oct): the lists that play each disposition don't carry what its primaries pay for (Anvil ρ −0.70 over the five), because an army can only pick the dispositions its detachments offer.

7. Banner, scored from effects (T-527, proposed)

For each character and each squad it can lead or support:

• Damage: the joined unit's wounds removed into the five targets, with the character's effects on and its own attacks included, against the squad alone.
• Durability: the hits to destroy it, with the character's defensive effects on the squad and its own wounds behind it, against the squad alone.
• Bodies: the same points spent on more models of the squad, at its next size up or its own rate per point.
• Multiplier: (damage added ÷ what bodies would add + durability added ÷ what bodies would add) ÷ 2. Above 1, the character beats bodies.
• Score: the best squad's multiplier, percentiled; the squad is recorded.
• Auras, measured (T-547): where the coverage data gives an aura its effects, tools/aurafx.js plays them on each squad of the army they reach; the 2 best squads' multipliers vs bodies count, × AURA_IN_RANGE = 0.5 for being in range.
• Proxies: any other aura is worth half the median leader's gain over a bare body, across 2 units. A transport is worth a quarter of a squad's damage.
• Not modelled: army-wide effects and CP.

7b. The attacker × defender matrix (T-568, built, not live)

tools/matrix.js; the build spec is design/matrix.md (sections 2 to 5), the plain-English account design/tier-engine.md. Nothing here reaches the site or the live rankings; T-569's field, allocation and jobs will read it.

• Profiles. An attacker profile is one weapon's dice with its name thrown away (kind, A, skill, S, AP, D, the keyword flags ev.js reads, plus Lance, Indirect Fire and Assault, and every keyword not read in full kept as x:<keyword>: since T-582 only the six that aren't dice; a condition on an ability is when:<ability>:<keywords>, Blast X blast:X, Cleave X cleave:X, dice Sustained Hits and Rapid Fire keep their dice). A defender profile is one allocation group of a datasheet: T (the unit's highest), Sv, Inv, W, Feel No Pain, its own unconditional defensive effects, Stealth, and the keywords an Anti or a condition names. Equal keys are one row or column.
• The cell, one attack into a fresh defender, by ev.js profileVs as is (section 1): u unsaved wounds, r damage with each wound capped at a model's W, e wounds removed in the long run (the renewal count: N(w) = [1 + Σ P(D = k) N(w − k)] ÷ (1 − P(D = 0)); e = u × W ÷ N(W), normal and mortal wounds apart), v = E[K] Var(X) + Var(K) E[X]² with K the unsaved wounds of one attack and X = min(D, W).
• Contexts. Precomputed layers: base, cover (BS worse by 1), half range (Melta), half range in cover, the charge (Lance +1 to wound). Every other context is a pair of modset keys, H = the first 16 hex of SHA-256 of the canonical merged effects; a unit's contexts are its own and army rules always × no detachment or each of its army's (every way to answer their choices, at most PICK_CAP = 32 a source) × alone or joined; marks off.
• Vouched conditions (round three; round two met them all game). A condition the engine can't check (when.manual) is a flag on the modset, vouch: [class, id], and a cell read for that modset mixes the branches with and without it by the share of the game it holds, independently per condition (effects with the same words are one condition): near an objective VOUCH.objective = 0.5, in Synapse range VOUCH.synapse = 0.75 (V5C.SYN), below half strength VOUCH.belowHalf = 0.35 (V3C.BELOW; "isn't below half" 0.65), below Starting Strength VOUCH.belowStart = 0.35 (never for a unit of one model), a Battle-shocked target VOUCH.shocked = 0.2, any other condition VOUCH.other = 1 (met all game, as the tier test reads it). The mixture's u, r and e are means; its v holds the between-branch part.
• Reading a volley from the cells. n models with weapon g into m models of W wounds: A = n × (mean A + Blast's one per five models) × One Shot's fifth; E = Σ A e, R = Σ A r, V = Σ A v. Wounds removed μ = E + (1 − E ÷ R) × min(R, VOLLEY.FRESH × W): the model in progress when the volley ends has had no killing blow, and holds on average half its wounds (VOLLEY.FRESH = 0.5, the renewal argument's mean, not fitted; 0.75 and 1 tried and worse on Terminators). Then the cut at the unit's m × W wounds, the volley taken as normal: μ − (μ − c) Φ(z) − σ φ(z), z = (μ − c) ÷ σ, σ = √V. The one thing a sum of cells can't see is mixed damage: one weapon's chip damage making another's big hits overkill (killDP plays the order).
• The cached build. build/matrix/<hash12>/ (git-ignored), named by a hash of the BSData commit, the engine files, the effects files, rankings.json and the scenario; the three newest kept. Incremental (step 6): a cell depends on its two keys and the cell hash alone (the engine, the effects files and the layers; not the commit, not rankings.json), so a build copies every cell an older folder with the same cell hash holds and computes only keys it lacks; the unit table likewise by row (the whole datasheet record at its size) and column (target key and points per wound). An army's catalogue edited: its new rows and columns only, and the result equals a build from scratch.
• The check (step 7, --check). CHECK.CELLS = 10,000 random real cells: r against profileVs to CHECK.TOL = 1e-9 in memory, CHECK.F32 = 1e-6 relative read back from the float32 files. Today's Offense: each pool unit into the five stand-ins with the squad the unit pages score into each (rankings.json lt), summed from the cells and read as above, against rankings.json off: Spearman ρ ≥ CHECK.RHO = 0.995 and median |difference| ≤ CHECK.MEDIAN = 3% per target, every pair beyond CHECK.OUTLIER = 15% named with its cause. On 3 Oct: ρ 0.9999, 0.9988, 0.9968, 0.9995, 0.9998; medians 0.02% to 1.57%; 2 pairs beyond 15% (the Corvus Blackstar into Terminators, mixed damage).

8. Known gaps

Fed by the calibration reports (T-533) and the public-list comparisons (T-524).

9. Changelog