Every dial screened, then one sampled search

5 Oct, Pacific (T-728). tools/everydial.js; results in build/everydial/ (git-ignored). Snapshot 8dcd24032d, inputs hash dcd477d8a78c (verified). Base: Track A (tools/ledger-benchmarks/2026-10-05-step34b-baseline.json), wCarry 0. Discovery on the frozen snapshot: nothing in the ledger moves and nothing is adopted.

In plain English

Jordan's rule (5 Oct): every live dial is in the search, and off is just one of its levels. So every dial in the T-726 inventory went in, as 95 factors. Stage 1a moves each one alone, away from Track A, on all 16 armies. It is scored as cascade2 scores, against three scrambled letter sets. Stage 1b was to move all 95 at once; by Jordan's call it was stopped and ten hand-built candidates were scored in its place.

Stage 1a, the quick screen (done, 4.9 minutes):

• 36 of the 95 factors pass. "Pass" means some level lifts the seven fair-test armies above 0 and above the scrambled seeds' gain. Those seven are cascade2's clean chain: Tyranids, World Eaters, Orks, Emperor's Children, Astra Militarum, Adeptus Mechanicus and Dark Angels. This rule is weak: about half of pure noise would get through it. Most passes are a tenth or two of a point.
• The top 15 by fair-test gain over scrambled, among passing levels. Each shows the real gain, then the scrambled gain:
  1. preferredTargets, cap 16: +0.81 against −1.96 scrambled.
  2. attachChars, every leader attached: +1.11 against −1.64.
  3. sureFire 0.5: +1.12 against −1.42.
  4. lifeKill, replacing a quarter of Burst: +0.29 against −1.57.
  5. smallPiece 0.2: +0.36 against −1.05.
  6. β all (perfect target choice): +0.36 against −0.74.
  7. lands 1 (today 1.5): +0.31 against −0.62.
  8. exclusiveCarry 1: +0.12 against −0.75. It is army-specific, and passes in World Eaters, Votann, Astra Militarum, Drukhari and Emperor's Children.
  9. typicalSize 0.25 at the best size: +0.30 against −0.40.
  10. arrivalKill 1.5: +0.15 against −0.53.
  11. landsNoClamp: +0.22 against −0.33.
  12. charEpic 1.5: +0.31 against −0.21.
  13. r.synapse 1.5: +0.32 against −0.15.
  14. synapseMarginal: +0.05 against −0.34.
  15. supportCredit κ 0.5: +0.23 against −0.16.

The other nine armies mostly agree on the big ones: preferredTargets +2.9, attachChars +3.0, sureFire +2.3, lifeKill +2.6, arrivalKill +2.6 and exclusiveCarry +2.1.

• What hurts the fair-test armies at every level:
  • charEnh and charLead at 1.5: −5.3 and −5.2.
  • killByJob: −4.8.
  • screenBlock and screenArea at 1.5: −4.2 and −4.0.
  • gradeBy total: −4.2. The other nine gain +2.6 from it, as T-700 found.
  • scarcity off: −3.7.
  • modeScarcity 0.5: −3.7.
  • charHome 1.5: −3.4.
  • scarceAnswers: −3.4.
  • joinedMarginal: −3.3.
  • bonusShare 1: −2.9. It is flagged: it reads detachment and aura-bearer list shares.
  • leadersLeading: −1.9.
  • r.debuff 1.5: −1.8.
  • r.mortal 1.5: −1.7.
  • supportMeasured 3: −1.4.

The leader credits (charLead, charEnh, leadersLeading, joinedMarginal) cost the fair-test armies while their scrambled letters gain. So these are real costs, as T-727 found on Dark Angels.

• The walls, Dark Angels (Track A 53.7) and Agents (56.1). No single level takes Dark Angels to 60. Its best are:
  • sureFire 0.5 without reserves: +4.3, to 58.1;
  • arrivalKill 1.5: +4.1;
  • r.buff 2.12: +3.4;
  • lands 1 and screenBlock: +3.2 each;
  • charPsychic 1.5 and leaderComp's penalty by rank: +2.9 each.

Agents pass 60 under six single levels:

• r.heal 1: 63.7, +7.6 against +0.7 scrambled;
• supportMeasured 3: 63.2, +7.2 against +2.8;
• charPsychic 1.5: 62.8, +6.7 against +0.1;
• killByJob: 61.0;
• lifeKill without survival: 60.5;
• charArrive 1.5: 60.1.

charPsychic 1.5 is the one level that lifts both walls: Dark Angels +2.9, Agents +6.7. It costs the fair-test pool at 0.5 (−0.4), though.

• One caution on "real − scrambled". The factors that reorder the most pull every scrambled score back toward 50%. Several fair-test armies' scrambled Track A sits above 50 (Adeptus Mechanicus 67 on seed 11, Emperor's Children 56 to 66). So big movers show negative scrambled gains, which flatters their "real − scrambled". The pass rule also asks for a real gain above 0, and that guards against it. Stage 1b's least squares, with standard errors, is the firmer test.
• Track A reproduces cascade2's score on all 16 armies to the pair.

• Under a stricter bar, no single 1a level survives on the fair-test seven. The bar is the 95th percentile of all 217 levels' scrambled gains (three seeds each): +1.16 there and +1.83 on the other nine. sureFire 0.5 (+1.12) and attachChars rate 1 (+1.11) fall just short. On the other nine, ten factors clear it:
  • preferredTargets (four levels);
  • attachChars;
  • sureFire 0.5;
  • lifeKill (five levels);
  • arrivalKill 1.5;
  • smallPiece 0.4;
  • exclusiveCarry 1;
  • carry (w1, dedicated, cargo shield);
  • r.detachment 2.07;
  • gradeBy half and total.

Stage 2, the 2^7 factorial on C1 (done, 1.7 minutes; section below): nothing beats C1 + arrivalKill 1.5 cleanly. Adding charPsychic 1.5 and attachChars gives the best all-16 mean (69.5 against 69.2), but costs the fair-test seven. Adding sureFire 0.5 to that misses only two bars (Adeptus Mechanicus, Emperor's Children) and is the one setting with Dark Angels (60.1) and Agents (60.5) both past 60. In sample.

In place of 1b: ten hand-built candidates (Jordan's call; 0.6 minutes). Stage 1b, the sampled search, was stopped (see "Stage 1b" below). In its place we scored ten settings built on cascade2's headline (its clean chain's #1, each army at its own values) plus 1a's leaders, exactly as 1a.

• No candidate clears every army's bar (Track A's score, at least 60). Dark Angels and Agents stay below 60 in all ten.
• The headline, C1, reproduces cascade2 on every army. It scores 68.3 on all 16 (Track A 61.9), 74.5 on the fair-test seven and 63.5 on the other nine, with scrambled letters at 48.5.
• C5, C1 + arrivalKill 1.5, is the best candidate on all 16 and on the fair-test seven:
  • 69.2 on all 16, 74.8 on the fair-test seven and 64.9 on the other nine, where C8 does best (66.0). Scrambled letters fall to 48.2.
  • Dark Angels reach 59.7, the closest any setting here gets to 60.
  • It misses five bars: Agents 56.5, Astra Militarum 67.1 (bar 67.6), Genestealer Cults 58.9, Thousand Sons 56.3 and Dark Angels 59.7.
• The other three 1a leaders don't add on top of C1:
  • preferredTargets C16 (C2) takes Adeptus Mechanicus from 72.0 to 59.9, the Tyranids to 89.2 and Dark Angels to 53.8.
  • attachChars rate 1 (C3) takes Adeptus Mechanicus to 65.5 and Dark Angels to 54.0. Agents gain (58.7).
  • sureFire 0.5 in place of 0.25 (C4) lifts Dark Angels to 57.5 and Thousand Sons to 58.9, and drops Agents to 52.5.
• All four together (C6) are worse than C5 alone: 68.3 on all 16. Adeptus Mechanicus falls to 58.7 and the Tyranids to 88.7. Gains in Genestealer Cults, Space Wolves and Thousand Sons (all past 60) don't make up for it.
• On top of C6:
  • lifeKill replace w0.25 (C7) costs 1.4.
  • charPsychic 1.5 (C8) is the best of the three on all 16 (68.7) and the other nine (66.0). It brings Agents to 59.2, Dark Angels to 58.2 and Thousand Sons to 67.0, but costs Emperor's Children (64.4).
  • The Agents movers, r.heal 1 and supportMeasured 3 (C9), bring Agents to 59.6 but sink Dark Angels (52.6) and Blood Angels (52.4).
• What lifts the walls here:
  • Dark Angels: arrivalKill 1.5 (59.7 in C5), charPsychic (58.2), sureFire 0.5 (57.5).
  • Agents: the Agents movers (59.6), charPsychic 1.5 (59.2), attachChars (58.7).
  • No setting lifts both past 60.
• Every candidate is in sample. The leaders were picked on these letters in 1a. Scrambled letters sit at 47 to 49 throughout, so the 16 to 21 points over scrambled are real signal. But a candidate's gain over C1 is the size of 1a's noise.

How it was run

• The factors. All 95 come from tools/dial-inventory.json. node tools/everydial.js --check fails if a dial is in neither a factor nor the exclusions.
  • 85 are game-wide. 10 are army-specific, as in cascade2:
    • five screen strengths and the screen structure;
    • modeScarcity;
    • exclusiveCarry;
    • objPresence;
    • bonusShare.
  • Dials behind a switch are levels of the switch's factor: off, then the switch on at each strength.
  • Exclusive families are one factor each:
    • the VP lines: secondaryVP, missionVP, missionBudget and vpLedger, nine levels;
    • lifeKill's modes, eight levels;
    • carry: wCarry, carryDedicated and carryCargo's eleven parts, twelve levels;
    • leaderComp's halves and options, eight levels.
  • pointsExp is smallPiece's −0.2 level.
  • Rule multipliers go from 1 (off) through Track A's value to a higher one.
  • Line weights take 0, Track A's value and twice it.
  • charHome 0.2 and charShockAmp 0.5 were cascade2's Tyranid-only values. Here they are game-wide levels, like every other character dial.
  • 312 levels in all.
• Excluded, each with its reason:
  • wKill: held fixed, because only the weights' ratios matter.
  • prem: an exact duplicate of wKill.
  • pointsExp: folded into smallPiece.
  • typicalW: a shorthand.
  • r.arrival, killMargin, armyPick, charsJoined and profileDials: inert or superseded.
  • the four cuts: they never move a Value.
  • popshare: tournament popularity stays out of letters.
  • combo, profileLines and the feature family: not live on this snapshot. combo is display only. The other two need the data tools/ledgerprofile.js prepares.
  • raw (lines as measured before T-608's one scale) stays in, flagged as a units switch, since the rule keeps every live dial.
• A setting:
  1. Each factor's level is laid over Track A.
  2. The setting goes through tools/exam.js runOf().
  3. Each post module it uses runs once on that run (tools/dialinventory.js postRun, given a bsdata option for this).
  4. Their changes to a unit's Value are added, as cascade2 adds them.
  5. gradeBy, a grading choice rather than a post step, multiplies the sum last.
• The score is cascade2's own: cascade2.js lettersOf, pairsOf and countRight; consensus pairs for the four multi-list armies, the list's pairs for the rest; scrambled seeds 11, 12 and 13.
• Stage 1a: Track A, then each factor alone at each of its other levels: 217 settings × 16 armies. On Track A's own switches that means off and a higher strength.
• Timing: 4.9 minutes on 8 workers. Dark Angels took 1.7 minutes and Space Marines 1.1. The rest took 0.1 to 0.3 minutes each.
• Commands:
  • node tools/everydial.js --rows 1a
  • node --max-old-space-size=3000 tools/everydial.js --compute 1a --heap 2500
  • node tools/everydial.js --analyse 1a, then --md 1a
• Candidates (stage "cand"):
  • node --max-old-space-size=6000 tools/everydial.js --rows cand
  • --compute cand, then --analyse cand and --md cand
  • tools/everydial.js CANDS lists them.
  • C1 is cascade2's clean #1 with each army's own values, recomputed by cascade2.js analyse (all). Its score matches cascade2's stored one on all 16 armies.

Stage 1a tables

Every factor, its best level by fair-test gain over scrambled (points of pairwise accuracy; real gain, then the scrambled seeds' mean, pooled over the 7 fair-test armies, the other 9, all 16):

The stricter bar: the 95th percentile of all 217 levels' scrambled gains (three seeds each), +1.16 on the fair-test seven and +1.83 on the other nine. Above it on the fair-test seven: none. On the other nine: carry (w1, dedicated w1, cargo shield); r.detachment (2.07); arrivalKill (1.5); preferredTargets (4/4, C16, N1, N8); attachChars (on, rate 1); lifeKill (auto, auto s0, replace w0.5, replace w0.25, replace w0.75); smallPiece (0.4); sureFire (0.5, 0.5nr); exclusiveCarry (1); gradeBy (half, total).

Every level, per army (real gain over Track A, points; scrambled mean in brackets where the level moves the army):

⚑ bonusShare: reads detachment and aura-bearer list shares.

The walls (each factor's best level for Dark Angels and for Agents, its gain, scrambled in brackets, and the score; Track A DA 53.7, Agt 56.1):

Run time (rows a minute is wall time on the workers): Tyr 0.2 min, WE 0.1 min, LoV 0.1 min, AdM 0.2 min, Agt 0.1 min, AM 0.3 min, BA 0.2 min, DG 0.1 min, Dru 0.1 min, EC 0.1 min, GSC 0.2 min, Ork 0.3 min, SW 0.2 min, TS 0.1 min, SM 1.1 min, DA 1.7 min.

Candidates, in place of 1b, by Jordan's call

C1 is cascade2's headline (its clean chain's #1: Track A + charEpic 0.5, aircraftBoard 1, r.synapse 1.5, sureFire 0.25, leaderComp's penalty 1, typicalSize 0.25, each army at its own values); each later candidate adds to it as named. Scored exactly as 1a (consensus pairs or the one list's pairs; scrambled seeds 11, 12, 13). Pairwise accuracy, %; scr is the scrambled mean.

Per army (real; scrambled mean in brackets; the bar is Track A's score, at least 60; the fair-test seven first):

Italic: below the army's bar.

Stage 1b: designed, stopped by Jordan's call

• The random block. 4,096 balanced, shuffled settings plus the 453 cascade2 anchors were drawn (build/everydial/1b.rows.json). Random settings cost about 7 times T-726's estimate:
  • they miss every ledger cache;
  • the VP-lines factor nests extra computes in 8 of its 9 levels and is on in 89% of rows.

Tyranid cold computes, sampled rows:

| VP lines | Compute | |---|---| | off | 1.0 s | | secondary | 2.7 to 3.0 s | | missionVP | 4.8 to 9.7 s | | vpLedger | 12.9 to 15.4 s |

Over the 16 armies, about 19.5 s of wall time a row on 8 workers, so 4,096 rows would take about 22 h. A first block of 1,024 rows was started, then stopped (Jordan: far too many). About 450 Tyranid rows and 49 rows on four armies are on disk, unanalysed.

• The two-level screen. In its place: a Plackett–Burman design of 96 runs over the 95 factors (a Hadamard matrix of order 96, H12 ⊗ H8), folded over to 192, with each factor at off against its best 1a level. Its rows were drawn (build/everydial/pb.rows.json; --rows pb, --analyse pb, --md pb are in the tool). Jordan stopped it before any run was scored, in favour of the candidates above.
• Analysis tools built but not used:
  • least squares on the sampled rows, with the stricter 95th-percentile bar and each factor's effect with the VP lines off and on (--analyse 1b);
  • the Plackett–Burman effects, with per-factor scrambled nulls (--analyse pb).

Both were checked on fake records with a planted effect.

Stage 2: the full 2^7 factorial on C1 (Jordan's go)

The 128 settings combine seven factors on C1 (cascade2's headline, each army at its own values), each off or on: arrivalKill 1.5, sureFire 0.5 (C1's 0.25 when off), charPsychic 1.5, attachChars rate 1, preferredTargets C16, r.heal 1 and supportMeasured 3. Each was scored as the candidates were, on all 16 armies. It took 1.7 minutes.

• Nothing beats C5 (C1 + arrivalKill 1.5) cleanly. "Cleanly" means higher on all 16, on the fair-test seven and on the other nine, with no more bars missed.
• The best on all 16 is C1 + arrivalKill 1.5 + charPsychic 1.5 + attachChars rate 1: 69.5 against C5's 69.2.
  • It costs the fair-test seven: 73.5 against 74.8.
  • It gains on the other nine: 66.3 against 64.9.
  • It misses three bars: Adeptus Mechanicus 63.7, Emperor's Children 66.0 and Dark Angels 59.9.
• The best for bars cleared is that setting with sureFire 0.5 as well.
  • It misses only two bars, Adeptus Mechanicus 64.3 (bar 71.4) and Emperor's Children 63.9 (bar 69.6).
  • It is the only setting here with both walls past 60: Dark Angels 60.1, Agents 60.5.
  • It scores 69.1 on all 16, 72.4 on the fair-test seven and 66.5 on the other nine.
  • That is a trade: the walls rise, and two fair-test armies fall.
• Main effects (on − off, all 16 armies): each is under half a point, except supportMeasured 3 at −1.24.
  • Beyond all three of its scrambled effects:
    • arrivalKill 1.5 on the other nine (+0.84) and on Dark Angels (+2.7);
    • attachChars on the other nine (+0.23);
    • three that hurt: sureFire 0.5 on Agents (−3.1), and attachChars (−1.6) and preferredTargets (−3.2) on Dark Angels.
  • charPsychic (+5.3 on Agents, +1.0 on the other nine, −1.1 on the fair-test seven) and supportMeasured (+4.6 on Agents) move a lot, but their scrambled effects move as much. They are not separable from noise here.
• Interactions are small.
  • The largest is preferredTargets × supportMeasured (+0.34 on all 16, beyond its scrambled spread).
  • preferredTargets × attachChars is −0.21: with preferredTargets on, attachChars's rate does nothing. So rows #8 and #9 are the same setting, as T-661 noted.
  • arrivalKill × preferredTargets is −0.10 on all 16 and −0.21 on the fair-test seven.
• The scrambled letters' best of the 128 is 50.6 on all 16. The real 69.5 is far above it. But the differences among the top settings (a few tenths) are the size of 1a's noise, and all of this is in sample.

The top 10 by the all-16 mean (pairwise accuracy, %; bars: Track A's score, at least 60), with Track A, C1 and C5 beside:

Fewest bars missed: 2. The best such settings:

Main effects (on − off over the 128 runs, points; the three scrambled sets' effects in brackets; bold when the real effect is larger in size than all three):

Two-way interactions (the mean where the two agree, both off or both on, minus the mean where they differ, on the same scale as the main effects; sorted by size on all 16; the top 10):

Nothing beats C5 cleanly (higher on all 16, the fair-test seven and the other nine, with no more bars missed).

Best of the 128 on scrambled letters (all 16, the seeds' mean of each seed's best): 50.6. C1 matches cascade2 on every army.

Run time: Tyr 8 s, WE 5 s, LoV 2 s, AdM 2 s, Agt 1 s, AM 4 s, BA 2 s, DG 1 s, Dru 3 s, EC 1 s, GSC 3 s, Ork 4 s, SW 4 s, TS 2 s, SM 29 s, DA 33 s.

Stage 2 as first proposed

The candidates say the next search is small. A full factorial over seven factors, 2^7 = 128 settings:

• C1's base, with each factor off or on:
  • arrivalKill 1.5;
  • sureFire 0.25 or 0.5;
  • charPsychic 1.5;
  • attachChars rate 1;
  • preferredTargets C16;
  • r.heal 1;
  • supportMeasured 3.
• Each army takes its own values as in cascade2.

Near C1 a setting costs about 4 s of wall time over the 16 armies (the ten candidates took 0.6 minutes), so the 128 settings take about 10 minutes. Then a 3-level full factorial on the factors that hold (3^5 = 243 settings, about 15 minutes; 3^7 = 2,187, about 2.2 h). Both are in sample, so held-out folds (the exam's 10-choose / 5-score) should judge whatever they pick.