The solution cascade: every setting that fits the Tyranids, carried army by army

5 Oct 2026. Jordan: "tweak the ever living God out of all the settings… every set of tweaked values above 80%… goes to another army after that." The cascade started from the Tyranids (Jordan's later call), with every toggle of the ledger free. This is discovery only: which variables the reviewers appear to value, and where they agree. Nothing is adopted; no ledger value, setting, page or letter moves.

• Snapshot: c5e7d85b43f8.
• Tool: tools/solcascade.js.
• Data: reports/tiers/2026-10-05-solution-cascade.json (every solution's knobs and stage scores).

In plain English

• Hundreds of settings fit the Tyranid reviewers better than today, and almost none carry to other armies.
  • With all 134 toggles and 155 datasheet-feature knobs free, 426 distinct settings beat Track A's Tyranid consensus score by 2 points or more (bar 92.7%; the best reaches 100%).
  • Scrambled letters clear that bar too: 527 settings, from two of the three scrambles. This many knobs can memorise the 52 Tyranid units whatever the letters say.
• World Eaters is where real and scrambled part. 10 of the 426 real settings also reach 80% on the World Eaters' consensus pairs. None of the 527 scrambled ones do (their best is 78.1%). This is the one stage where the real letters carry something the scrambled ones don't.
• The next army is a wall. None of the 10 reaches today's score on Leagues of Votann (best 60.0 against a bar of 61.0). In the chain of multi-list armies, none reaches it on Space Marines either (best 54.6 against 62.2). Either way the cascade stops at stage 3.
• What the 10 share (they come from 4 lineages, 7 of them from one):
  • command points weighted up: r.cp at 2.88, the top of its grid, and the strongest single knob for passing the World Eaters;
  • support counted as a credit rather than measured;
  • scarce answers on, Kill over a unit's life off;
  • cheap scoring screens on (screenScore 0.1);
  • no movement or enhancement credit for characters;
  • in most of the 10: leaders and Epic Heroes worth more, Hold and Presence up, the raw Kill and Soak weights down.

In Jordan's terms: reviewers of Tyranids and World Eaters appear to value command-point and support rules, and character worth, more than our dice do, and objective play more relative to raw damage. The walls show the opposite for Votann and Space Marines: their reviewers rate big vehicles (Hekaton, Land Raiders, Repulsors) high and cheap characters low.

• Beyond the wall, each army scored on its own: real settings beat scrambled ones on 8 of 14 armies by median score and lose on 6 (Votann, Astra Militarum and Space Marines among them). No setting of either kind reaches today's score on Space Marines; on Dark Angels, one scrambled setting does. Nothing in the datasheet-and-toggle space is a consensus across armies.
• The prediction ("few or no real solutions survive Tyranids; scrambled ones die at the same rate or faster") was half wrong. 426 real solutions survived the Tyranids, and scrambled ones survived them nearly as well. Scrambled ones did die faster at the World Eaters (0 against 10), as predicted.

Method

Knobs: 134 toggles on coarse grids (376 levels) plus 155 feature knobs. The full catalogue is in the JSON.

• Switches: every ledger.js switch except the two profile ones (profileLines is on throughout and carries the features), and the page switches solo, raw, scarcity and soakfx.
• Weights and rules: the seven line weights at ×0.25 to ×4; every rule kind (r.*) at 0.5, 1, today and today + 1.
• Other dials: β, lands, reach, threat, fireB, tunnel, anywhere, every strength dial, κ, wCarry and the eleven transport dials, pointsExp, smallPiece, aircraftBoard, loneOpSoak, leaderCredit and armyPick.
• Army-specific: modeScarcity and the nine screen dials.
• Characters and army rules: the 14 character dials and cpOnce; the five army-rule dials.
• Previews: rivalW, typicalSize under its three size rules, typicalDet, jailW and stealW.
• gradeBy: perPoint, half or total.
• The 155 feature knobs are slopes per SD on the parts of Value each acts on, plus flat points. They are the features that vary among the Tyranids or the World Eaters, standardised within every army as tools/wecarry.js does.

Scores are exact: the page's compute() with the knobs on the data, then the preview steps, then gradeBy. A fast path (the forms' parts of Value read off one compute) is used only inside the feature fits; its median gap to the exact score is 1.0 point.

Pairs: armies with two or more 11e lists are scored on consensus pairs, the pairs that every list grading both units orders the same way. One-list armies are scored on their list.

The pipeline is identical on the real letters and on three scrambles. Each scramble moves each unit's row of letters whole, so the lists' agreement and the consensus-pair count are kept.

1. 2,400 space-filling Tyranid settings (balanced columns, features at 0; 42 minutes).
2. Boosted trees (depth 3) on those settings, for importance.
3. The best 100 settings per letter set get their feature knobs fitted.
4. An evolutionary walk of 400 steps: 1 to 4 toggles mutated, none frozen, features re-fitted.
5. Distinct solutions: different toggles, or features at least 0.01 apart.
6. Each later stage re-chooses the army-specific dials: as carried, screens off, T-677's screens, modeScarcity 0 or 0.2.

Not reached: thousands of solutions. A compute at a random setting takes 2 to 15 seconds on the Tyranids (the mission and VP-ledger reference passes). Four hours bought 426 real solutions in 64 seed lineages.

The stage table

How the search got there:

Beyond the wall (each army on its own, a fixed panel)

Read: each of the 10 World Eaters survivors passes between 0 and 5 of these 14 armies. Against the 80 scrambled solutions, the 40 other real ones pass more often on Orks, Blood Angels, Drukhari, Genestealer Cults and Death Guard, and less often on Votann, Astra Militarum, Space Wolves and Thousand Sons. The scrambled solutions come from only 8 lineages, so their column is lumpy.

What the survivors share

The lineage caveat: the 10 World Eaters survivors come from 4 seed lineages, 7 of them from one, so much of what they share is inheritance, not evidence. Two random real solutions differ on a median 82 of 134 toggles; two survivors differ on 73.

Shared by all ten, and acting on the World Eaters (see the probe below):

• r.cp at 2.88 (32% of all 426 solutions, 100% of survivors; today 1.88). It is also the strongest knob in the World Eaters stage's trees: permutation importance 13.4, then a Vehicle slope at 7.0 and r.heal at 6.0 (R² 0.67).
• supportCredit on and supportMeasured off.
• scarceAnswers on, lifeKill off.
• screenScore at 0.1.
• charMove at 0, charEnh at 0.

Shared by all ten, but inert on the World Eaters (so set by the Tyranid stage or the lineage): r.synapse 2.18 and aircraftBoard 1. joinedShapley 1 acts only with joinedMarginal on, and that is off in 8 of 10.

Shared by most (7 to 9 of 10):

• Characters: charLead 1.5, charEpic 1.5, leaderCredit 0, charCP 0.
• Support and rules: κ 2, r.heal at today's 1.805, r.mortal 0.5, armyOrders 0.
• Weights: wHold and wPresence ×4, wKill and wSoak ×0.25.
• Kill model: β 12, lands 1.5.
• Grading and size: perPoint grading, smallPiece 0.2, pointsExp 0.25, typicalSize 0.25.
• Other switches: vpLedger on; rivalW, jailW and stealW all 0.

Feature knobs: the survivors pull down units with many datasheet profiles, transports, many wargear choices and shock rules. None of the 155 sits in a tight band away from 0 across all ten.

In words: reviewers of the Tyranids and World Eaters appear to value command points and support rules about 50% more than today's multiplier does (2.88 against 1.88). They also appear to value character and Epic Hero worth more, and Hold and Presence more relative to Kill and Soak. With 4 lineages and a wall at the next army, read this as a direction, not a size.

Importance (Tyranid sample, 2,400 settings, features at 0)

Permutation importance comes from boosted trees, scored on a held-out fifth (R²: real 0.45; scrambled 0.62, 0.52, 0.46).

• On the real letters, the knobs that matter are support, command points and characters.
• On scrambled letters, they are how a unit is graded by its points, and the transport dials.
• What the two share is gradeBy and the screen slots.

Strongest pairs (split gain along a path, real): screenSlots × gradeBy; supportMeasured × r.cp; r.cp × gradeBy; supportMeasured × r.synapse; r.synapse × r.cp; r.first × gradeBy; r.lone × gradeBy; r.first × r.lone; screenBlock × gradeBy; screenArea × gradeBy; screenArea × screenSlots; r.cp × screenSlots.

Level means are the Tyranid consensus score at each level, averaged over the rest of the design (* marks today's value).

Toggles we didn't keep that come back here:

• supportMeasured off beats on (Track B uses it; Track A doesn't).
• Higher charLead and charEpic help (both were measured at 0 before).
• screenSlots 1 helps (T-677's setting, an army setting).
• Higher aircraftBoard helps.

The World Eaters stage, real (426 entering, 10 surviving):

• Top knobs: r.cp 13.4; slope on Vehicle 7.0; r.heal 6.0; slope on minimum range 4.6; slopes on Smoke, Frame and Twin-linked 2.7 to 2.8.
• Strongest pairs: r.cp × Vehicle; minimum range × Frame; r.heal × r.cp.

The World Eaters stage, scrambled (527 entering, none surviving): r.movement 22.3 and the slope on carried ranged weapons 11.5 lead; different knobs altogether.

Inert toggles per army (the probe)

Each toggle was switched through its levels from three starting points:

• Track A;
• every switch on (the mission pair aside);
• Track A with what the toggle acts through switched on.

Inert means no unit's Value moved. Inert toggles stayed free everywhere; none was fixed.

Inert on Tyranids (21 of 134): killMargin, killClassWeight, r.arrival, r.spawn, killClassWeightW, carryFrag, carryFragRef, carryPer, carryDedOnly, carryCap, carryOnce, carryArrive, carryNeed, carrySquads, armyPick, charOC, armyMarks, armyRounds, armyStates, armyEnabler, armyOrders. Compute median 226 ms.

Inert on World Eaters (41 of 134): charsJoined, synapseMarginal, killMargin, condShares, killClassWeight, dialBiovoresOnly, shockPressure, cpRefile, soakfx, wSpawn, tunnel, anywhere, r.arrival, r.overwatch, r.first, r.lone, r.synapse, r.spawn, shockPressureW, carryFrag, carryFragRef, carryPer, carryDedOnly, carryCap, carryOnce, carryArrive, carryNeed, carrySquads, aircraftBoard, loneOpSoak, armyPick, charArrive, charOC, charShockAmp, charPsychic, charDebuff, cpOnce, armyMarks, armyStates, armyOrders, stealW. Compute median 68 ms.

The walls: units to explain

Leagues of Votann (auspex; 18 graded units; 15 tournament lists for the share)

Ranked lower by the survivors than by the reviewers (the survivors under-rate):

Ranked higher by the survivors than by the reviewers (the survivors over-rate):

Space Marines (auspexFull, tobias; 84 graded units; 33 tournament lists for the share)

Ranked lower by the survivors than by the reviewers (the survivors under-rate):

Ranked higher by the survivors than by the reviewers (the survivors over-rate):

How to read the walls:

• Ranks are among the army's graded units (1 is the best).
• "Reviewers" is the consensus rank, from the mean tier over the army's 11e lists.
• List share is the share of the army's tournament lists that take the unit (grimstat-corpus, CC BY 4.0; discovery only, never in a letter).
• Lines are Track A's split of the unit's Value.
• "All ten agree": every survivor ranks the unit on the same side of the reviewers, so it is more likely an army quirk than noise.

The pattern on both walls: the survivors put cheap characters near the top and the big vehicles near the bottom, and the reviewers do the reverse.

Files and run

• tools/solcascade.js (new). Run node --max-old-space-size=3000 tools/solcascade.js <phase> --workers 8; the phases are probe, sample, surrogate, search, cascade, beyond, stages, walls and report. Work files go to build/solcascade/ (git-ignored; a stopped run resumes from them).
• reports/tiers/2026-10-05-solution-cascade.json: the catalogue, bars, stage tables, importances, beyond-the-wall scores, walls, and every distinct solution (toggle levels, feature knobs, score at each stage).
• reports/tiers/2026-10-05-solution-cascade-walls.md: the walls tables on their own.

Time with 8 workers: about four hours, including building the tool.