The cascade re-run: everything that held, game-wide and army by army

5 Oct. tools/cascade2.js. Results are in build/cascade2/ (git-ignored). Snapshot 8dcd24032d, inputs hash dcd477d8a78c. The base is Track A (tools/ledger-benchmarks/2026-10-05-step34b-baseline.json). This is a report: nothing in the ledger moves and nothing is adopted.

In plain English

We tried every combination of the dials that have held so far. Ten ledger sliders and three post dials apply to every army: 16,384 game-wide settings. On top of that, each army picked its own values for its own dials (screen slots, modeScarcity, exclusiveCarry, objPresence, bonusShare; for the Tyranids also charHome and charShockAmp). An army's own values counted only where they beat what the same freedom gains on scrambled letters. The armies were then taken one at a time, Tyranids first and Dark Angels last. A setting stayed in only if it scored at least as well as Track A does today on that army (and at least 60%).

• The Tyranids keep about two thirds of the settings: 10,752 of 16,384 clear Track A's 89.5% on their consensus pairs, best 93.5%. No scrambled setting gets through (best 51.6 / 61.9 / 54.9%). No Tyranid-only value (charHome, charShockAmp, screens, transports, objectives, bonus share) beats its scrambled control: the Tyranids keep the shared values.
• Twelve armies pass, then there is a wall. The 10,752 settings pass World Eaters, Votann, Space Marines and Drukhari untouched. Orks, Blood Angels, Emperor's Children, Astra Militarum, Genestealer Cults, Space Wolves and Thousand Sons then cut them to 213. The wall is Agents of the Imperium: none of the 213 reaches 60% there (best 58.3%; Track A 56.1%). Adeptus Mechanicus (best 68.3% against 71.4%), Death Guard (58.8% against 60) and Dark Angels (59.3% against 60) would stop them too. Death Guard reaches 60% under no setting at all (best 59.7%).
• Without single-list armies, the four multi-list armies (Tyranids, World Eaters, Space Marines, Genestealer Cults) leave 5,806 survivors.
• Several stages are ones noise also passes. At Genestealer Cults, Space Wolves and Thousand Sons, a scrambled seed clears the bar as easily as the real letters: all 1,018 settings entering Thousand Sons also clear it on seed 12. Those are the stages that decide most of what the 213 share. Space Marines, Drukhari, Votann and Blood Angels pass on one seed too, but they cut little.
• The clean chain uses only the seven armies whose scrambled letters clear their bar under no setting at all: the Tyranids, World Eaters, Orks, Emperor's Children, Astra Militarum, Adeptus Mechanicus and Dark Angels. On it, 240 settings survive six armies, and the wall is Dark Angels (best 58.1% against 60).
• What survivors share, then, depends on which armies decide. Only Epic Heroes credited (charEpic 0.5: 61% of the 213, 75% of the clean 240), with lands and β left at today's 1.5 and 6, holds in both. The rest flips:

| Dial | Main chain's 213 | Clean chain's 240 | |---|---|---| | r.cp | 2.88 (69%) | 1.88 (100%) | | smallPiece | 0.2 (86%) | 0 (100%) | | attachChars | even | 0 (100%) | | sureFire | 0.5 without reserves (81%) | 0.25 (50%), never 0.5 without reserves | | typicalSize | 0.25 (100%) | 0.25 (43%) |

Adeptus Mechanicus decides most of the clean chain. The multi-list chain sits between: r.cp 2.88 68%, r.first 1.8 74%, pushed there by Genestealer Cults.

• Command points: the solution cascade's call for them (r.cp 2.88) comes only from the stages noise also passes. The clean stages that move it lean to today's 1.88: after Astra Militarum, 62% of survivors have it, and after Adeptus Mechanicus, 100%.
• Armies' own values that beat their scrambled control:
  • World Eaters: screens on, exclusiveCarry 0.5 (+7.3 points against +3.2 scrambled).
  • Drukhari: exclusiveCarry 1, bonusShare 1 (+16.4 against +4.8).
  • Votann: exclusiveCarry 1 or 0.5, bonusShare about half the time (+6.7 against +5.1, a narrow margin).
  • Space Marines: screens on, with transports, objectives and bonus share (+1.3 against +0.6).
  • Emperor's Children: exclusiveCarry 1 (+4.2 against +1.0).
  • Space Wolves: screens on, modeScarcity 0.2 (+5.0 against +2.4).
  • Thousand Sons: exclusiveCarry 1, bonusShare 1 (+7.1 against +2.1).

Every other army keeps the shared values, the Tyranids included.

• The top five survivors on the exam: the 15-army exam (pooled 11e lists) rises from 59.3 to 66.8 to 67.3 (14 of 15 armies better; Adeptus Mechanicus falls 71.4 → 65.2). The Tyranids' three 11e lists rise from 81.6 to 81.8 to 82.1. With every army at the shared values, #1's game-wide values alone give 63.8 (+4.5) and Tyranids 82.0. The clean chain's #1 is nearly Track A: charEpic 0.5, aircraftBoard 1, sureFire 0.25, leaderCompPen 1, typicalSize 0.25, and r.synapse 1.5. It gives 65.8 (+6.5) pooled and Tyranids 84.0 (+2.4), and Adeptus Mechanicus improves too (72.0). All of this is in sample: the settings and each army's values were picked on the same letters they are scored on.
• Each army on its own, past the wall, does better than today in every army. The headline per army is the last table:
  • Biggest gains: Votann +28.6, Thousand Sons +25.6, Drukhari +23.2, World Eaters +16.9.
  • Smallest: Adeptus Mechanicus +1.9, Tyranids +3.9.
  • Where one-list armies jump, their scrambled letters often jump nearly as far: Space Wolves 65.0 against 67.9 on one seed; Genestealer Cults 67.9 against 75.9; Thousand Sons 63.8 against 68.6.

How it was run

• Grid: each army gets 1,024 game-wide slider combinations times its own sliders (4, or 16 for the Tyranids): 77,824 computes in all. Each is a compute() of the ledger page (tools/exam.js runOf(), wCarry 0) with ten post calls, one per non-zero post value.
• Post steps: each reads compute()'s run, and their changes to a unit's Value add up. That is the same as chaining them to first order, and it lets one compute serve all 288 post combinations.
• Run time: about 2 h 10 min with 8 workers. Space Marines took 30 min and Dark Angels 50 min; the objective-presence probe doubles the compute in armies that split.
• Score check: Track A's score matches tools/leadercomp.js's consensus score (multi-list) or the exam's own (one list) on all 16 armies, and the pooled exam reproduces 59.32 and the Tyranids 81.61.

The score. Armies with two or more 11e lists are scored on their consensus pairs: pairs graded by at least two lists, where every list grading both puts them the same way round. Those armies are the Tyranids, World Eaters, Space Marines and Genestealer Cults. A one-list army is scored on its list's own pairs. Ties count wrong.

The control. Three scrambled letter sets (seeds 11, 12, 13) keep the consensus structure. Each army's 11e-graded units are permuted jointly, so each takes another unit's whole row of letters, and the pair graph stays the same. Each stage also reports its own control: of the real settings entering the stage, how many would clear that army's bar on its scrambled letters. This matters because the scrambled chains themselves all die at the Tyranids.

An army's own values (the gate).

• For each army, compare its best score with its own values against its best game-wide-only score.
• Do the same on each scrambled seed.
• The army's own values are allowed when the real gain beats the scrambled seeds' mean gain. Each scrambled seed is gated against the other two.
• When allowed, each game-wide setting takes the army's best own values. Among ties, the one with fewest dials off the shared value wins.
• Two other readings are shown beside: the gate taken per game-wide setting, and the gain shrunk by the scrambled gain in place of a gate. Gating per setting keeps the wall at Agents. Shrinking the gain brings it forward to Thousand Sons. With game-wide values alone it comes at Space Wolves.

Stages and bars.

• The order is the Tyranids, World Eaters, Votann, then the rest, easiest first, with Dark Angels last. "Easiest" means the largest share of the 16,384 settings that clear the army's bar on its own.
• Every bar is Track A's score on this snapshot, at least 60%.

Stage table

All armies. The table counts survivors out of the 16,384 game-wide settings. The column "On its scrambled letters" counts, out of the real settings entering a stage, how many would clear that army's bar on each scrambled seed.

Rows 14 to 16 score the 213 that reached the wall against those armies, each on its own.

Multi-list armies only (single-list armies are weaker evidence):

Here is how the 5,806 multi-list survivors fare on the single-list armies, each on its own:

The clean chain. It runs only the armies whose scrambled letters clear their bar under no game-wide setting (with their own values gated as usual), so no stage is one noise can pass:

Here is how the clean 240 fare on the other armies, each on its own:

Which game-wide sliders move which armies (from the stored scores; a slider that moves nothing in an army can't be chosen by it):

How an army's own values are counted (survivors after each stage; with every reading, the scrambled chains are 0 after the Tyranids):

With game-wide values alone, Drukhari already cuts the 9,674 to 712. Drukhari need their own transports (exclusiveCarry 1) to sit near their reviewer.

What survivors share

Each value's share among the survivors. Every value is equally common in the grid: 50% each, or 25% for sureFire.

On their own, the Tyranids mostly don't care. About two thirds of every value passes; they mark down smallPiece 0.2, attachChars 1 and lands 1.25.

The main chain and the clean chain agree on three things: Epic Heroes credited (charEpic 0.5), lands 1.5 and β 6, the last two today's values.

They disagree on r.cp, smallPiece, attachChars, sureFire and typicalSize.

• In the main chain, the push comes from Genestealer Cults (r.first, r.cp), Space Wolves (sureFire) and Thousand Sons (typicalSize, lands, smallPiece). Those are the stages whose scrambled letters clear the bar as easily as real ones.
• In the clean chain, it comes from Adeptus Mechanicus: today's r.cp 1.88, smallPiece 0, attachChars 0 and β 6, every one of the 240.
• r.first 1.8 leans the same way in both (82%, 58%), but it moves only eight armies.

Each army's own values among the survivors at its stage:

The top five surviving settings, and the exam

The ranking is by the mean margin over the bars of the 12 armies passed. All five share charEpic 0.5, smallPiece 0.2, r.cp 2.88, lands 1.5, sureFire 0.5 without reserves, leaderCompPen 0 and typicalSize 0.25.

The multi-list chain's top five share charEpic 0.5, smallPiece 0.2, attachChars 1, aircraftBoard 1, r.cp 2.88, r.first 1.8, lands 1.25, β 6, sureFire 0 and leaderCompPen 1, and differ in armyEnabler, r.synapse and typicalSize.

The exam. Each army's score is the mean pairwise over its 11e lists; the pooled score is over the 15 armies. Every army takes its own values where its gate allows them.

Per army, Track A → #1 → #1 with game-wide values only → the clean chain's #1:

Everything here is in sample: the cascade chose the settings on these letters. The exam's own bootstrap and held-out folds were not run on them, and the gain is not a held-out estimate.

Beyond the wall: every army on its own

Each army's best game-wide setting, with its own values where its gate allows them, against Track A. The table also shows the best the same search finds on that army's scrambled letters, and how many of the 16,384 settings clear its bar on its own (real, then the three scrambled seeds).

Scores more than five points above every scrambled seed: the Tyranids, World Eaters, Votann, Orks, Astra Militarum, Emperor's Children, Drukhari and Blood Angels.

Scores below a scrambled seed: Genestealer Cults, Space Wolves, Thousand Sons and Death Guard.

Scores within about two points of a scrambled seed: Agents, Adeptus Mechanicus, Space Marines and Dark Angels (Space Marines and Dark Angels on one seed).

Each army's best game-wide setting differs. Where settings tie, the first value of each dial in the grid wins: r.first 1.8, r.cp 1.88, r.synapse 1.5, lands 1.25, β 6 and every 0. So a dial that doesn't move an army shows its first value there. r.synapse, for one, moves only the Tyranids and Genestealer Cults (checked on the stored scores). The settings below list only the dials away from today's values, leaving out r.synapse 1.5 in the other armies:

What this says, and what it doesn't

• The one game-wide value that holds wherever we look: Epic Heroes credited, charEpic 0.5. It is the majority in the main, multi-list and clean chains, and the best setting of 11 of the 16 armies takes it (two can't move it). Lands 1.5 and β 6, today's values, also hold.
• Command points are not settled. r.cp 2.88 (the solution cascade's call) wins only where Genestealer Cults, Space Wolves and Thousand Sons decide, stages that scrambled letters also pass. Adeptus Mechanicus, a clean stage, takes today's 1.88 in every survivor. The same split holds for smallPiece, attachChars, sureFire and typicalSize: none is a shared value yet.
• Transport delivery holds army by army. exclusiveCarry (what only a transport can carry) beats its scrambled control in Drukhari, Votann, Emperor's Children, Thousand Sons and World Eaters (0.5 there), and with screens in Space Marines. The army is the right level for it, not a game-wide value.
• Screens are a World Eaters, Space Marines and Space Wolves setting. They are never on for the Tyranids.
• Not supported anywhere:
  • objective presence (extras + enabler) as an army's own value (at most a minority choice beside transports);
  • charHome and charShockAmp for the Tyranids;
  • any Tyranid-only value.
• The wall is narrow.
  • Main chain: Agents (58.3), Death Guard (58.8) and Dark Angels (59.3) each miss 60% by under two points among the 213. Adeptus Mechanicus is the outlier: every one of the 213 loses about six points there (71.4 → 65.2 at #1).
  • Clean chain: Adeptus Mechanicus keeps 240 settings close to Track A, and they miss Dark Angels (58.1), Agents (59.2) and Death Guard (59.0).
  • Death Guard reaches 60 under no setting at all.
• The exam gains are mostly the armies' own values and are in sample. #1 adds +7.9 pooled, +4.5 of it from game-wide values. The clean chain's #1 adds +6.5 with the Tyranids +2.4: it is Track A plus charEpic 0.5, aircraftBoard 1, sureFire 0.25, leaderCompPen 1 and typicalSize 0.25, with each army's own transports and screens.
• Not tested here:
  • held-out folds (the exam's gains above are in sample);
  • chaining the post steps (their changes are added);
  • the measured Synapse and command-point replacements (supportMeasured, synapseMarginal);
  • Hidden (T-699) and Soak effects (T-696).