The big crack's exam: the Tyranid net on the other armies (T-662)

5 Oct 2026. tools/ledgercrack2.js --net then --exam --armies … (the user's framing: if one code works for all armies, it is among the Tyranid winners). Fit on Tyranids only; every stored combination scored as it is on the other armies; the bar is pairwise ≥ 70% on an army. The exam is staged: a gate of armies at a time, only a gate's survivors going on to the next, as T-668's per-army inputs land. A report: nothing is adopted. Every dial is one value for the whole game, never per unit.

On one page: rung (c) and the ceiling scan

• Rung (c), Track A (155 game-wide knobs: the sliders, the fast dials and every datasheet feature as a slope on its family's lines; searched, 1,162,592 evaluations, about 4 minutes): the five Tyranid lists 78.9% → 92.9% ⚑, above the reviewers' own agreement; the same search on shuffled letters reaches 91.5% (gap +1.4), the fake worlds gain +1.2 to +2.2; held-out units 79.3% → 80.2%; the 14 armies 58.9% → 57.1%, Space Marines 64.2% → 61.7%. It fits whatever letters it is given: the finding is that 52 units cannot identify 155 knobs (the placebo families showed the same at twelve).
• Rung (c), Track B (152 game-wide knobs: the sliders, the fast dials and every datasheet feature as a slope on its family's lines; searched, 1,162,463 evaluations, about 4 minutes): the five Tyranid lists 77.5% → 92.8% ⚑, above the reviewers' own agreement; the same search on shuffled letters reaches 92.6% (gap +0.2), the fake worlds gain +1.0 to +3.0; held-out units 77.6% → 79.8%; the 14 armies 59.9% → 54.0%, Space Marines 61.6% → 57.2%. It fits whatever letters it is given: the finding is that 52 units cannot identify 152 knobs (the placebo families showed the same at twelve).
• The ceiling scan, G2 (all 5,121,913 stored combinations on all 15 armies, in-sample on purpose): the best mean over the 15 is 62.7% (1 of 15 armies at 70%), against Track A's 59.4% and B's 60.1%; 12,530 combinations reach 62%, 0 reach 63%. With the 5 exam armies held out: the best on the other 10 (62.1%) scores 60.2% on the 5 (Track A 57.5%, B 59.4%); the 100 best on the 10 average 60.0% on the 5.
• The ceiling scan, G3 (all 11,764,811 stored combinations on all 15 armies, in-sample on purpose): the best mean over the 15 is 61.6% (0 of 15 armies at 70%), against Track A's 59.4% and B's 60.1%; 0 combinations reach 62%, 0 reach 63%. With the 5 exam armies held out: the best on the other 10 (63.5%) scores 57.4% on the 5 (Track A 57.5%, B 59.4%); the 100 best on the 10 average 57.3% on the 5.
• Reading: even in-sample, the best combination of this knob space averages 62.7% over the 15 armies and clears 70% on few of them: the weights, switches, dials and family slopes cannot reach 70% broadly. The next gain has to come from new ingredients (what the ledger measures), not new weights.

The wide net

Every combination of each grid at or above its floor on the five Tyranid lists: the lower of (its best − 2 bootstrap SE) and 75%, whichever keeps more. Stored in build/crack2/ (git-ignored; node tools/ledgercrack2.js --no-grids --net makes it again from the checkpoint).

Gate: Space Marines and World Eaters (the whole net)

Bar 70% on every army of the gate; scored on each army's letters over all its lists (as T-668's tools/exam.js), an army's score the mean over its lists, the mean the mean over the gate's armies (each once; T-668's "pooled"), and its SE from each army's bootstrap SE (graded units resampled, 300 draws), √(Σ SE²) ÷ armies. "Pairs pooled" counts every pair of every list of every army once (big armies weigh more).

Inputs: every army carries T-668's per-army inputs (burst, preferred targets, life, measured support; Space Marines in docs/lab/ledger/space-marines.json, the 14 in build/ledger/<army>.json from tools/ledger-all.js), so every switch acts on every army (checked: preferredTargets, the attached Burst, lifeKill and supportMeasured each change Values on all 16). The family dials act through the datasheet features, which every army has; a feature no Tyranid unit varies on stays at 1× ("unset by Tyranids", the main report's section 5).

G2: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (3,378,291): 0.0721, 62.3%, 63.0%; 78%–80% (1,606,270): 0.076, 62.7%, 63.4%; 80%–82% (137,156): 0.0576, 62.6%, 63.5%; 82%–84% (196): 0, 62.8%, 63.6%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

G3: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (7,652,512): 0.0437, 56.0%, 59.0%; 78%–80% (3,153,154): 0.0041, 54.6%, 58.0%; 80%–82% (897,193): 0.0024, 54.9%, 58.1%; 82%–84% (61,818): 0.0024, 56.0%, 59.2%; 84%–86% (134): 0, 56.3%, 60.5%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

G3's outer settings among them: #0 G2 (200): supportMeasured 0, kill burst, lifeKill 0, arrivalKillW 0, secondaryVPW 0, supportMeasuredW n/a, shockPressureW 1, threatSoakW 1, scoreFloorW 0, pointsExp 0, wSoak 1×, wScore 1×, wActions 1×, wHold 1×, wSpawn 1×, wPresence 1×.

Space Marines: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

World Eaters: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Gate: all 15 other armies (the whole net)

Bar 70% on every army of the gate; scored on each army's letters over all its lists (as T-668's tools/exam.js), an army's score the mean over its lists, the mean the mean over the gate's armies (each once; T-668's "pooled"), and its SE from each army's bootstrap SE (graded units resampled, 300 draws), √(Σ SE²) ÷ armies. "Pairs pooled" counts every pair of every list of every army once (big armies weigh more).

Inputs: every army carries T-668's per-army inputs (burst, preferred targets, life, measured support; Space Marines in docs/lab/ledger/space-marines.json, the 14 in build/ledger/<army>.json from tools/ledger-all.js), so every switch acts on every army (checked: preferredTargets, the attached Burst, lifeKill and supportMeasured each change Values on all 16). The family dials act through the datasheet features, which every army has; a feature no Tyranid unit varies on stays at 1× ("unset by Tyranids", the main report's section 5).

G2: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (3,378,291): 0.8344, 43.7%, 60.6%; 78%–80% (1,606,270): 0.9607, 44.8%, 61.0%; 80%–82% (137,156): 0.933, 45.6%, 61.1%; 82%–84% (196): 0.102, 45.5%, 60.9%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

The ceiling scan (in-sample on purpose: the upper bound of this knob space across armies). The best mean over the 15 any of the 5,121,913 combinations reaches: 62.7% (Tyranids 78.2%; 1 of 15 armies at 70%): supportMeasured 0, kill prefer, lifeKill 0, arrivalKillW 3, secondaryVPW 0.5, supportMeasuredW n/a, shockPressureW 0, threatSoakW 1, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 0×, wSpawn 0×, wPresence 2×. Per army: Space Marines 63%, Adeptus Mechanicus 63%, Agents Of The Imperium 59%, Astra Militarum 79%, Blood Angels 65%, Dark Angels 58%, Death Guard 57%, Drukhari 57%, Emperors Children 65%, Genestealer Cults 66%, Leagues Of Votann 64%, Orks 67%, Space Wolves 61%, Thousand Sons 50%, World Eaters 69%. Combinations at or above 60 / 61 / 62 / 63 / 64 / 65%: 1,292,729 / 237,923 / 12,530 / 0 / 0 / 0. The mean over the 15, combinations per point from 50% to 70%: 0 / 0 / 0 / 0 / 0 / 2,076 / 61,403 / 419,961 / 1,422,893 / 1,922,851 / 1,054,806 / 225,393 / 12,530 / 0 / 0 / 0 / 0 / 0 / 0 / 0.

With the 5 exam armies held out (Astra Militarum, Death Guard, Genestealer Cults, Leagues Of Votann, Space Wolves; fixed by seed before any run): picked on the other 10 (62.1% there, against Track A's 60.4% and B's 60.4%), the pick scores 60.2% on the 5 (Track A 57.5%, B 59.4%); the 100 best on the 10 average 62.1% there and 60.0% on the 5. The best any combination reaches on the 5 alone: 66.1%. The pick: supportMeasured 0, kill burst, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW n/a, shockPressureW 0, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 0×, wScore 1×, wActions 0×, wHold 0×, wSpawn 0×, wPresence 1×.

G3: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (7,652,512): 0.7092, 41.6%, 57.5%; 78%–80% (3,153,154): 0.696, 42.4%, 57.3%; 80%–82% (897,193): 0.7188, 42.8%, 57.4%; 82%–84% (61,818): 0.8172, 43.0%, 58.0%; 84%–86% (134): 1.4403, 43.6%, 58.7%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

G3's outer settings among them: #49 G2 (10): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 2×, wHold 2×, wSpawn 1×, wPresence 1×; #42 G2 (9): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 1×, wPresence 2×; #41 G2 (8): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 0×, wPresence 2×; #44 G2 (7): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 1×, wPresence 2×; #43 G2 (6): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 0×, wPresence 2×; #47 G2 (5): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 1×, wPresence 1×; #54 G2 (5): supportMeasured 1, kill burst, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 2, threatSoakW 1, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 2×, wPresence 2×; #2 G2 (4): supportMeasured 1, kill burst, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 1, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 2×, wPresence 2×.

The ceiling scan (in-sample on purpose: the upper bound of this knob space across armies). The best mean over the 15 any of the 11,764,811 combinations reaches: 61.6% (Tyranids 75.5%; 0 of 15 armies at 70%): G2 #6631679 (supportMeasured 1, kill burst, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 2, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 0×, wPresence 2×); families move 0.25, defence 0, tough -0.25, objective 0, ranged 0, melee 0, weaponkw 0.25, leader 0.25, synapse 0.25, size 0, body 0.25, loadout 0. Per army: Space Marines 61%, Adeptus Mechanicus 64%, Agents Of The Imperium 62%, Astra Militarum 66%, Blood Angels 65%, Dark Angels 58%, Death Guard 59%, Drukhari 68%, Emperors Children 64%, Genestealer Cults 48%, Leagues Of Votann 69%, Orks 65%, Space Wolves 53%, Thousand Sons 57%, World Eaters 66%. Combinations at or above 60 / 61 / 62 / 63 / 64 / 65%: 74,292 / 2,713 / 0 / 0 / 0 / 0. The mean over the 15, combinations per point from 50% to 70%: 236 / 29,963 / 248,027 / 757,140 / 1,504,918 / 2,180,290 / 2,539,732 / 2,353,411 / 1,493,691 / 583,111 / 71,579 / 2,713 / 0 / 0 / 0 / 0 / 0 / 0 / 0 / 0.

With the 5 exam armies held out (Astra Militarum, Death Guard, Genestealer Cults, Leagues Of Votann, Space Wolves; fixed by seed before any run): picked on the other 10 (63.5% there, against Track A's 60.4% and B's 60.4%), the pick scores 57.4% on the 5 (Track A 57.5%, B 59.4%); the 100 best on the 10 average 63.2% there and 57.3% on the 5. The best any combination reaches on the 5 alone: 63.7%. The pick: G2 #6631655 (supportMeasured 1, kill burst, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 2, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 1×, wPresence 2×); families move 0.25, defence 0.25, tough -0.25, objective 0, ranged 0, melee 0, weaponkw 0.25, leader -0.25, synapse 0.25, size 0, body 0.25, loadout 0.

Space Marines: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Genestealer Cults: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Leagues Of Votann: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

World Eaters: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Gate: all 15 other armies (what passed Space Marines and World Eaters)

Bar 70% on every army of the gate; scored on each army's letters over all its lists (as T-668's tools/exam.js), an army's score the mean over its lists, the mean the mean over the gate's armies (each once; T-668's "pooled"), and its SE from each army's bootstrap SE (graded units resampled, 300 draws), √(Σ SE²) ÷ armies. "Pairs pooled" counts every pair of every list of every army once (big armies weigh more).

Inputs: every army carries T-668's per-army inputs (burst, preferred targets, life, measured support; Space Marines in docs/lab/ledger/space-marines.json, the 14 in build/ledger/<army>.json from tools/ledger-all.js), so every switch acts on every army (checked: preferredTargets, the attached Burst, lifeKill and supportMeasured each change Values on all 16). The family dials act through the datasheet features, which every army has; a feature no Tyranid unit varies on stays at 1× ("unset by Tyranids", the main report's section 5).

G2: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (555): 2.8721, 40.4%, 61.5%; 78%–80% (439): 2.7745, 40.5%, 61.2%; 80%–82% (6): 2, 42.3%, 62.2%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

G3: no survivors: the best available

Does a better Tyranid fit go with the other armies? Mean armies cleared, lowest-army score and pooled pairwise by Tyranid band: 75%–78% (1,000): 1.147, 42.6%, 61.3%.

The best 15 (by armies cleared, then the mean over the gate's armies; ± the bootstrap SE for the first ten):

The code (the best 200 (none survive), 200 combinations): each dial's levels among them.

G3's outer settings among them: #0 G2 (155): supportMeasured 0, kill burst, lifeKill 0, arrivalKillW 0, secondaryVPW 0, supportMeasuredW n/a, shockPressureW 1, threatSoakW 1, scoreFloorW 0, pointsExp 0, wSoak 1×, wScore 1×, wActions 1×, wHold 1×, wSpawn 1×, wPresence 1×; #43 G2 (11): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 0×, wPresence 2×; #44 G2 (9): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 1×, wPresence 2×; #41 G2 (7): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 0×, wPresence 2×; #42 G2 (6): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 1×, wPresence 2×; #47 G2 (4): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 0×, wHold 2×, wSpawn 1×, wPresence 1×; #48 G2 (3): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 1×, wHold 2×, wSpawn 1×, wPresence 1×; #49 G2 (3): supportMeasured 1, kill attach4, lifeKill 0, arrivalKillW 1.5, secondaryVPW 0, supportMeasuredW 1, shockPressureW 4, threatSoakW 0, scoreFloorW 0, pointsExp 0, wSoak 2×, wScore 2×, wActions 2×, wHold 2×, wSpawn 1×, wPresence 1×.

Space Marines: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Genestealer Cults: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Leagues Of Votann: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

World Eaters: which reviewer's style the best match

Per list: the pairwise on that reviewer's rows for the Tyranids (Auspex's six bands where they are finer; the mean of lists is the objective), on letters for the other armies (the exam's yardstick). Consensus: each unit's mean letter over the lists that grade it, and the pairs that mean separates. Last columns: how well that reviewer agrees with the consensus of the others (their style's distance from the room), beside how well this setting agrees with that reviewer (letters). A setting whose agreement with a reviewer runs above that reviewer's own agreement with the others is matching that reviewer's style, not the room's.

Re-run a gate: node tools/ledgercrack2.js --no-grids --exam --armies space-marines,world-eaters --skip-search --write; the next gate on its survivors: --exam --armies <the next armies> --on space-marines+world-eaters. Numbers: reports/tiers/2026-10-05-crack-exam.json.