The leader audit's eight fixes, applied (T-715, 5 Oct)

Jordan approved all eight fixes from the leader audit (T-714) and the re-freeze the data fixes need. This report says what each fix changed, then re-measures: Track A, the reviewers' leader check, T-709's leader competition and every leader toggle. Numbers: 2026-10-05-leader-fixes.json. New snapshot: code 8dcd24032d with index.json and the panels kept as at eb3590a, inputs hash dcd477d8a78c (was c5e7d85b43f8; the T-695 lock is kept as tools/ledger-inputs.lock-d8b5a1bb14.json). BSData 374f505. Rules were read on Wahapedia's 11th-edition pages to understand them; nothing here is their wording.

In plain English

• Leaders are now measured on what they bring, not on the squad's whole worth.
  • A leader's own command points, healing, Synapse and auras now multiply only the leader's own share. Before, they multiplied the whole squad: T-709's measure gave the Apothecary 6,113 a hundred points on Hellblasters. On the same measure it is now 886, Ûthar on Hearthguard 2,730 → −57, the Iron-master on Thunderkyn 3,111 → 409.
  • Joined units now get a Burst, so a leader is no longer credited with the squad's second phase.
  • Support characters (the Apothecary, the Lieutenants, the Ancients) now join beside a Leader instead of competing with Captains and Chaplains for the squad.
• Track A barely moves, as expected: the ranking reads each unit at its best solo form. Pooled exam 59.37 → 59.32, consensus pairs 60.08 → 60.02, the Tyranid headline 82.45 → 81.61. The moves come from three data fixes:
  • Old One Eye, the Red Terror and the Rhino lose their healing multiplier (−45%). Their self-heal is already counted on Soak, so it was counted twice. Old One Eye falls 13th → 24th and the Red Terror 16th → 30th; both are graded B or better by the reviewers, so the Tyranid headline drops 0.8.
  • Units with a banner or standard gain +1 OC a model: the Astra Militarum command squads, Company Heroes, Grey Knights Terminators, Sanguinary Guard, the Kabalite icons.
  • The Dark Angels' Apothecary is now measured by the Spawn line, as the Space Marines' is.
• On the reviewers' check, the dice a leader lends now beat chance.
  • The check: does the engine's best leader for a squad hold the reviewers' best letter?
  • Before: 25.8% of 151 cases, against 36.6% by chance.
  • Now, pricing only the dice a character lends and comparing Leaders with Leaders and Supports with Supports: 50.9% of 172 cases, against 38.7% by chance.
  • Adding the grants (Fights First, +OC) to the price drops it to 35.8%. The Judiciar's Fights First for a whole squad, at Track A's fitted ×2.64, swamps every other Support. So the competition prices the dice alone.
• Leader competition (T-709) now partly holds out. The penalty half (a leader that is best on no squad loses its nearest miss) gains +0.58 ± 0.11 held out, against a placebo of +0.00. Before the fixes it lost −0.36. The whole idea (credit and penalty together) is +0.03 against a placebo of +0.05, and credit alone −0.10: neither holds.
• No leader toggle holds yet. leadersLeading, joinedMarginal, joinedShapley, leaderCredit, charLead and charHome all pick 0 or lose held out. The fixes shrink their losses: leaderCredit at full strength −3.76 → −1.88 pooled, and joinedShapley now gains +0.28 pooled, though not held out.

What each fix changed

The re-freeze: only docs/data/unitfx.json moved among the 27 inputs. detfx.json, detachments.json and rankings.json are the T-687 builds, and build_rankings.js reads none of the new fields. Every ledger was rebuilt: the Tyranid and Space Marine chains, then node tools/ledger-all.js. The burst tail ran with armyburst, and the shared stamped files (carry, secvp, missionvp, rulevals, armyrules, field classes) were restamped; only their stamps moved. The dials fingerprint was re-recorded, and 3 of its 17 hashes moved: vpLedger, "supportMeasured off" and "attachChars and preferredTargets off". The rest set the healing kind to 1 or read solo forms only.

Track A, before and after

The exam on 11th-edition lists, at Track A (step 34b), wCarry 0:

The joined-form fixes (1, 2, 3, 5, 6, 8) move no ranked number: Track A ranks each unit at its best solo form, and its line scales, threat median and reference forms read solo forms only.

The reviewers' leader check

For each squad with two or more candidate characters, in each 11th-edition list that grades at least two of them differently: does the character the measure puts first hold the best letter? Ties share the hit. Chance is the share of graded candidates holding the best letter. tools/leaderaudit.js --check <price>; before, on the copy of the ledgers built before the fixes.

• By role is the comparison T-715 makes: Supports now compete with Supports on the squad's best led form, so there are more cases (172).
• Where the dice measure still misses:
  • Captains lend no dice. Rites of Battle is a Stratagem discount, so the Chaplain is the best Leader on most Marine squads, while the reviewers grade the Captain A.
  • Genestealer Cults' and Agents' candidates are mostly Supports whose worth is a rule we only note.
  • Of the Hive Tyrant, the Swarmlord and the Neurotyrant on Tyrant Guard, only the Neurotyrant lends dice.
• The grants hurt the check:
  • The Judiciar's Fights First, at Track A's fitted r.first ×2.64 on a whole squad's melee, makes it the best Support on every squad it can join (1,904 to 5,796 a hundred points). The reviewers grade it B or D.
  • The Ancient's +1 OC makes it the best Support on Hellblasters and Tactical Squads, where the reviewers prefer the Lieutenant.

T-709 re-tested: leaders compete for the best squad

tools/leadercomp.js --measure (strengths 0, 0.25, 0.5, 1; held out by army, 10 armies to choose and 5 to score, 3 seeds; placebo seeds 11 to 13), Track A at wCarry 0:

On the dice price, the whole idea at 0.5 adds +0.33 to the pooled exam (placebo −0.26), +0.42 on consensus pairs and +0.21 on the Tyranids. The penalty at 1 adds +0.42 pooled (placebo −0.07). The lend price's penalty is larger (+1.44), but its spread is seven times the dice's and one fold gains only +0.54. It rests on the Judiciar and the Ancients, the very cases the check says it gets wrong.

The leader toggles, alone, on the fixed ledger

tools/leadertoggles.js (new): each toggle over Track A, held out by army (10 and 5, 3 seeds), against a placebo that deals each unit's change to a random unit of its army (3 draws).

None holds. Each picks 0 in most folds or loses held out.

• The fixes help the toggles that read joined forms. leadersLeading −2.22 → −1.72, joinedMarginal −2.35 → −1.39, leaderCredit −3.76 → −1.88, and joinedShapley now gains +0.28 pooled (8 armies better, 6 worse), though not held out.
• charLead and charHome don't read joined forms and are unchanged, as expected.
• Blood Angels gain 16 to 22 points under every toggle that reads joined forms, on 19 graded units. Agents of the Imperium lose 8 to 14.

The wall units

Track A rank and Value (unchanged unless named above), and each one's competition on the dice it lends (g, a hundred of its points; leadercomp --army <slug>). "Before" is T-709's g.

Decided alone

• Fix 2's way: joined forms get a Burst, rather than reading comparisons with burstKill off (above).
• Fix 3's best Leader: the best Leader for each three-member form is chosen by the engine report's slider-free measure, at each squad size. Roles come from the links: a Supporting link wins over a Leading link for the same squad. In BSData 374f505 that makes the World Eaters' Master of Executions a Leader, not a Support as the audit had it.
• Fix 4's scopes: a to heal stays on its bearer's solo form as army-wide worth and never reaches a squad's share. A self heal takes no multiplier where the Soak effects count it (Tyranids, Space Marines). In the other armies it keeps its multiplier, because nothing else counts it there.
• Fix 6's default price is the dice alone. The card says "dice gain", and the check agrees. The grants are a separate price (lend).
• Fix 7's grants are leading only where the rule says "while it leads" (Wahapedia's 11th-edition wording, read to understand). The Judiciar's solo form doesn't fight first, so its Track A number doesn't move.
• "Charged this turn" is settled only on lends (leaderfx), as the audit asks. Units' own and detachment effects with that condition stay vouched.
• The leader check's case count by role is 172. The 151 by-squad cases are kept beside it to compare with the audit.
• ledger-all's time limit is raised to 45 minutes.

Doubts

• The Judiciar's Fights First is priced at r.first ×2.64, fitted on units that fight first themselves. On a whole attached squad that is very large, and it is why lend fails the check. The calculator still doesn't read the grant.
• Track A's Tyranid headline falls 0.8, from Old One Eye and the Red Terror losing a doubled heal. The reviewers grade both well. Their worth now rests on Soak, which weighs little at Track A.
• The fingerprint was re-recorded on this PC. As before, CI may disagree on vpLedger.
• The Teleport Crest moves nothing. The charge model sees deploying on the line as equal to a turn-two Deep Strike for Hearthkyn Warriors. Its value to Jordan (a delivery that can't be screened) is T-712's delivery certainty, not lands.
• The to repairs keep their army-wide multiplier on the bearer's own forms, as every other army-wide kind does. Whether a repair is worth ×1.8 of a character's own lines is the multiplier's question, not this card's.
• Tests:
  • ledger: 110 passed, 0 failed.
  • exam 3, scoringdoc 1, leaders 12, unitfx 1, vocab 4, soakfx 4, spawn 5 and charge 6 passed.
  • matrix: 2 failures (profile counts, Offense parity), both there before.

Reproduce

node tools/ledger-inputs.js --verify # hash dcd477d8a78c node tools/exam.js tools/ledger-benchmarks/2026-10-05-step34b-baseline.json --md node tools/leaderaudit.js --check dice # also lend, lendown, joined; --dir <old ledgers> node --max-old-space-size=5000 tools/leadercomp.js --measure # --price lend|lendown|joined node --max-old-space-size=5000 tools/leadertoggles.js