The leader audit (T-714, 5 Oct)

T-709 judged leaders by what our engine says each one adds to a squad, and those numbers looked wrong. This report finds out why. It is a report only: nothing is rebuilt and no ledger number moves. The tool is tools/leaderaudit.js and its numbers are in 2026-10-05-leader-audit.json. It reads compute()'s run at Track A (step 34b, wCarry 0, as tools/exam.js runs it) on the 16 armies' built ledgers, snapshot d8b5a1bb14 (hash c5e7d85b43f8, BSData 374f505). The hand checks rebuild each joined form inside the ledger's realm (the frozen inputs) and run the ledger's own kernel. Rules were read on Wahapedia's 11th-edition pages to understand them; nothing here is their wording.

In plain English

• The dice are right. The pricing is wrong.
  • We rebuilt all 636 leader-and-squad pairings and ran the ledger's own kernel on each, with and without the leader's effects. The effects reach the right weapons with the right conditions. 624 of the 636 rebuilt forms match the stored Kill exactly; the other 12 are mixed detachments, where we rebuilt only their main part.
  • The leader's dice lend comes out at about two-thirds of the textbook per-roll gain. For example, re-rolling every hit at 4+ should add 50% to hits; the Carnifexes' Kill rises 34–36%. Extra wounds partly land on models that are already dead in the engine's finite enemy army. This is the same for every leader, so it doesn't change who ranks first.
• What T-709 called "synergy" is mostly not the leader's buff. It was the joined form's Value minus the squad alone and the leader alone, per 100 of the leader's points. Five things fill that gap, none of them the dice the leader lends:
  1. The leader's own army-wide rules multiply the whole squad.
     • Command points (×1.88), healing (×1.805), Synapse (×2.18) and buffs (×1.12) multiply the joined form's whole Value, so the squad's worth is scaled by a rule that belongs to the leader.
     • This is 57% of all the synergy T-709 measured.
     • The Apothecary's 6,113 is 5,189 from Gene-seed Recovery's command points multiplying ten Hellblasters.
     • Ûthar's 2,730 on Hearthguard is 2,161 from Grim Efficiency (a once-a-round Stratagem discount) multiplying the Hearthguard. His Rampart Crest adds nothing there, because the Hearthguard already have the same 5+ invulnerable save and Deep Strike.
     • The Iron-master's 3,111 on Thunderkyn is 2,558 from a repair rule that only reaches vehicles and Steeljacks.
  2. Kill is read on two scales.
     • With burstKill on, a solo form's Kill is its Burst: one activation of its natural mode only.
     • A joined form keeps the engine's Kill: both phases, every turn.
     • So the squad's second phase is credited to the leader. This is 27% of the synergy measured, and 141 leaders are moved by more than 200 a hundred points.
  3. Support characters compete with Leaders.
     • In 11th edition an Apothecary, a Lieutenant, an Ancient or (in BSData 374f505) a Master of Executions is a Support character. It joins a squad beside its Leader, not instead of one.
     • Our forms hold one character each, so 37 such characters (160 pairings in 9 armies) fight Captains and Chaplains for the squad.
  4. What joining does for any leader is divided by the leader's points.
     • Joining changes OC times survival, sheltered wounds and shared targets. Those changes would be there whichever character joined.
     • Dividing them by the leader's points favours cheap characters. The Einhyr Champion lends no dice and still scores 269 on Hearthguard after fixes 1 and 2.
  5. Missing data.
     • Some grants are noted but never computed: the Judiciar's Fights First for its unit, the Kâhl's Teleport Crest (Deep Strike for Hearthkyn Warriors) and the Ancients' +1 OC.
     • 41 leaders' lends are written wholly as vouched conditions, which count as never met.
• On the reviewers' check, fixes 1 and 2 together bring the engine to chance; pricing the lend on its own beats chance. The check: does the engine's best leader hold the reviewers' best letter? (T-709 reported 30% of 129 cases. Our count differs a little; see the table.)
  • Today it does 25.8% of the time.
  • With fixes 1 and 2 it does 36.4%, against 36.6% by chance.
  • Scoring the lend at the squad's own rates gives 40.4%.
  • Scoring the joined form's Value on one scale with no leader multipliers gives 45.7%.
• The Winged Hive Tyrant is rightly not a leader. BSData 374f505 gives it no Leader ability, and neither does the 11th-edition page. Only the walking Hive Tyrant (Tyrant Guard) and the Swarmlord lead.
• How many leaders are mis-measured:
  • 168 of the 182 characters with pairings in the 16 armies are off on their best squad by more than half and by more than 200 a hundred points, against their lend priced at the squad's own rates.
  • 126 of the 182 move by more than half from fixes 1 and 2 alone.
  • In short, about nine in ten of the game's leader synergies are wrong by more than half. The worst are the cheap characters with a command-point, healing or Synapse rule.

How it was measured

For each pairing, the tool takes the joined form T-709 takes: the best g over sizes, loadouts and detachments, with "alone" read as ledger.js partnerAlone() reads it. It then reads four versions of g:

"Hand" in the tables is the textbook per-roll gain of the lend on the squad's main weapons, worked from the profiles (BSData 374f505), so it can be set beside the engine's Kill change.

Votann

At fixes 1 and 2 the Kâhl is Hearthguard's best leader (942, against the Champion's 269 and Ûthar's 155), as the reviewers have it. He is also the best on Warriors (442, against the Iron-master's 399, the Grimnyr's 292 and Ûthar's 19). The Iron-master keeps Thunderkyn, which no one else may lead.

Space Marines

Tyranids

• Hormagaunts are led by the Tyranid Prime with Lash Whip, not the Winged Prime. In BSData 374f505 the Winged Tyranid Prime leads Gargoyles and Tyranid Warriors.
• The Winged Hive Tyrant is right as it stands. It has no Leader ability in BSData 374f505 (the Library's datasheet) or on the 11th-edition page. Its own rules are a Stratagem discount and a Fight-phase psychic debuff, so T-709's "not a leader in our data" is correct.

World Eaters

The Lord on Juggernaut has no dice to lend. What makes him worth taking is speed, and the ledger counts extra movement at ×1 on lands, so he can't win a leader contest on today's data.

The causes, ranked by reach

Across the 16 armies: 636 pairings, 182 characters.

Not a cause: the dice. The lend reaches the right weapons. For example, Lethal Hits reaches every Hearthguard weapon and the Kâhl's own. The re-roll of wound 1s reaches only ranged weapons. The vouched conditions stay at zero, as the ledger rules them.

How each measure does on the reviewers' check

The cases: each squad with two or more candidate characters, in each 11th-edition list that grades at least two of them with different letters. A hit is when the character with the best value holds the best letter among those graded; a tie for the best value shares the hit. Chance is the share of the graded candidates that hold the best letter.

There are 151 cases (T-709 reported 129). The difference is how they are counted. The ordering of the measures doesn't depend on it: counting every edition's lists (160 cases), or taking the best among all candidates rather than the graded ones, moves each figure by 1 to 2 points.

Per army, at fixes 1 and 2 against chance: Adeptus Mechanicus 9 of 9 (chance 3.1), Death Guard 4 of 4 (2.3), Space Wolves 6 of 7 (2.7), Agents 8 of 15 (4.7). The armies that stay low are Astra Militarum 0 of 6, Genestealer Cults 0 of 8 and Orks 1 of 6. In all three the candidates are mostly Support characters, or leaders whose worth is a noted rule.

Proposed fixes

None is applied; they wait for Jordan. Each gives its expected effect, read off this audit's numbers.

1. A leader's army-wide multipliers on its own share only (tools/ledger.js and the page's valueForm).
   • In buildForms, tag each rule with the member it comes from: baseRules = dedupeRules(b.members.flatMap((m, k) => unitRulesOf(m.p).map((r) => ({ ...r, from: k })))). Mark a joined form's army-kind rule leaderOnly when the squad lacks it.
   • In valueForm, a joined form keeps such a kind out of m.total and adds (x − 1) × the leader alone's weighted lines. The form would carry them as leaderLines, from partnerAlone at the same detachment.
   • For leadercomp alone, the same thing in one line: score g with those kinds off on both the joined form and the leader alone. That is this audit's "mult off".
   • Expected: the Apothecary on Hellblasters 6,113 → 924, Ûthar on Hearthguard 2,730 → 569, the Iron-master on Thunderkyn 3,111 → 553, the Broodlord 2,671 → 448, the Master of Executions 4,447 → 689. In all, 225 pairings and 79 leaders move.
   • Alone it doesn't help the reviewers' check (25.8% → 23.2%), because the Kill-scale gap then decides.
2. One Kill scale for joined forms.
   • Either give joined forms a Burst (tools/burst.js on the joined unit with its lend: its natural mode, swapped by burstView like a solo form), or have leadercomp read every form with burstKill off.
   • Expected: with fix 1, the reviewers' check goes to 36.4% (chance 36.6%), and the Kâhl leads Hearthguard (942 against the Champion's 269 and Ûthar's 155).
3. Support characters as Support.
   • In matrix.js contextEnv, keep BSData's supports in a separate map from leads (BSD.canJoin already tells them apart, and bsdata.js has the role).
   • tools/ledger.js would then build squad + Support forms, and squad + its best Leader + Support forms (three members; the kernel and the allocation stages take members in order).
   • leadercomp would compare Support only with Support, on the squad's best led form.
   • Expected: the Apothecary, the Lieutenants, the Ancients and the Judiciar stop taking Captains' and Chaplains' squads; 160 pairings and 37 characters are re-cast.
4. Healing scoped to whom it heals, and counted once (data: unitfx.json, tools/soakfx-rules-space-marines.json, spawn rules).
   • A heal rule for vehicles, Steeljacks or the bearer alone carries a to (keywords) or self scope, and the ledger lists it only on forms holding such a model: Forgewrought Expertise, the Enginseer's and Iron Priest's repairs, the Meks' repairs, Unstoppable Monster.
   • Narthecium is counted in one place: inSpawn in soakfx-rules-space-marines.json, so healMultOwn leaves it out. The Dark Angels' Apothecary gets the same Spawn rule instead of a ×1.805.
   • Old One Eye's self-heal is already on Soak (healMultOwn), so it shouldn't also be a heal kind.
   • Expected: 26 leaders lose a heal multiplier on squads it never reaches.
5. Lone Operative on a joined form only when every member has it (unitRulesOf over the members).
   • Grants (Locus, Lelith Hesperax) stay.
   • Expected: 7 pairings (the Enginseer, Lucius, Snikrot) lose ×1.93 on Soak.
6. Price the lend apart from joining (tools/leadercomp.js).
   • A leader's synergy would be its lend, measured as here (the realm rebuild with and without its effects) at the squad's own rates, plus its own rules' worth from its solo form.
   • This replaces "joined minus both alone", whose joining effects favour cheap characters.
   • Expected: 40.4% on the reviewers' check (45.7% with the joined form's Value on one scale), against 36.6% by chance. This is the first measure here above chance.
7. The missing grants (data):
   • The Judiciar's Tempormortis as Fights First for the unit (kind "first").
   • Teleport Crest as Deep Strike for the unit, so the charge model moves the joined form's lands.
   • Banners and Regimental Standards as +1 OC per model in the joined form's ocPer100.
   • "The unit charged this turn" on a lend settled at the form's lands, not counted as never met: the Captain with Jump Pack, Astorath, the Lord of Contagion and Lord Marshal Dreir among them.
   • Expected: lifts leaders the reviewers value for these rules. Not measured here.
8. partnerAlone at the joined form's own loadout. "Alone" is the same Auto-equip squad the joined form carries, not the best solo loadout.
   • Expected: removes loadout swings like the Reivers'. Not counted game-wide.

Order: fixes 1, 2 and 3 carry most of the error. Fixes 4 and 5 are data and small. Fix 6 changes what T-709 measures. After the fixes T-709 should be re-tested: they move the inputs, not its dial.

Doubts

• Rules were read on today's 11th-edition pages; the ledger is pinned at BSData 374f505, before the 30 Sep update. The Master of Executions is Support in 374f505 and shows as a Leader today. The Judiciar is both in 374f505 and a Leader today.
• The lend is priced at the joined form's own rate for a point of Kill or Soak. That rate carries the squad's own multipliers and Track A's scales. A defensive lend (Feel No Pain, an invulnerable save) is cheap because Soak weighs little at Track A; the Grimnyr's +36% Soak comes to 26 a hundred points.
• The reviewers' check uses 151 cases. Many squads have only Support characters or noted-rule leaders as candidates, so a few armies (Astra Militarum, Genestealer Cults, Orks) can't be won by any dice measure.
• The hand-worked gains are per roll against one or two target profiles. The engine's lower figures are the kernel's finite field, as above; that wasn't separately proven for every leader.

Reproduce

node --max-old-space-size=2500 tools/leaderaudit.js --hand --json reports/tiers/2026-10-05-leader-audit.json # every pairing (about 15 min, one realm process per army) node --max-old-space-size=2500 tools/leaderaudit.js --army leagues-of-votann --pairs "Hearthguard|,Warriors|" --hand --weapons # one army's pairings, the lend weapon by weapon node tools/leaderaudit.js --from reports/tiers/2026-10-05-leader-audit.json # rescore the reviewers' check from the saved rows