The question. Phase 1 built a dial for each kind of rule a character carries (14 dials, plus cpOnce, all at 0; characters). Do any of them help on armies they weren't chosen on? The test runs on the new snapshot (inputs hash a20dbd2c3316) and the 11th-edition pool (T-691: 15 armies on their 11e lists, the Tyranids on their three 11e transcripts beside). The base is Track A (step 34b, wCarry 0).
What we found
- No character kind carries the cross-army miss. Only two dials passed the rule "held out above 0 and above its placebo": charEpic (+0.29 ± 0.54) and charAura (+0.09 ± 0.12). Both are well inside noise. The other twelve and cpOnce held out at 0 or below. The placebos (the same credit given to as many random characters) did as well or better for most.
- The phase 1 favourite failed. charPsychic held out −0.23 ± 0.35 against a placebo of +0.02. In sample it was best at 0.1 (+0.11), and the Tyranids fell 0.24.
- The leader kinds hurt. charLead (−0.36), charEnh (−0.29, 10 armies down, none up) and charHome (−0.46) all lost, and cpOnce lost at every strength (−0.57 at 1). Crediting leaders, by buff, slot or home, points the wrong way again, as T-670 found.
- The two that held, together: charEpic 0.5 with charAura 0.1, chosen held out, gives +0.29 ± 0.55 (9 armies up, 3 down). The exam on the setting chosen on all 15 gives +0.5 [−0.7, +1.7] (9 of 15 better). The Tyranids are −0.1 on the three 11e lists and −0.1 on held-out Tyranid units.
- With the cross-army trio (aircraftBoard 1, smallPiece 0.4, attachChars), the trio alone gives +2.8 [+0.4, +5.4] (9 of 15). The trio with charEpic 0.5 gives +3.2 [+0.7, +6.0] (10 of 15). Held out, the trio with characters is +2.99 ± 1.48 over Track A, and the characters add +0.20 ± 0.20 over the trio (9 armies up, 3 down). The aura dial drops out once the trio is on. The Tyranids pay −2.8 on their three lists either way (held-out units −3.3), and almost all of that is the trio's.
- Where the characters move: Blood Angels +4.8, Agents +3.1, Thousand Sons +2.6, Death Guard +1.3, Astra Militarum +1.1. Space Wolves −5.7 [−19.8, +4.2]; Emperor's Children −1.0.
| Dial (strength chosen on all 15) | Held out Δ ± SE (armies up / down) | Placebo held out | In sample (pick) | Tyranids, 3 11e lists (at the pick or 0.3) | Held-out Tyranid units (vs lists) | Verdict |
|---|---|---|---|---|---|---|
| charLead (0) | −0.36 ± 0.26 (2 / 7) | −0.29 | 0 | −0.29 at 0.3 | −0.6 ± 0.1 | no |
| charMove (0) | −0.06 ± 0.04 (1 / 5) | −0.14 | 0 | 0.00 | 0.0 | no |
| charArrive (0) | −0.34 ± 0.21 (1 / 6) | −0.05 | 0 | 0.00 | 0.0 | no |
| charOC (0.3) | −0.02 ± 0.05 (2 / 4) | +0.01 | +0.10 | 0.00 | 0.0 | no |
| charShock (0) | −0.13 ± 0.11 (2 / 4) | −0.00 | 0 | +0.01 at 0.3 | +0.1 ± 0.1 | no |
| charShockAmp (0) | 0.00 (never picked) | +0.01 | 0 | +0.11 at 0.3 | +0.2 ± 0.1 | no (Tyranids only) |
| charCP (0) | −0.22 ± 0.18 (1 / 5) | −0.13 | 0 | 0.00 | 0.0 | no |
| charPsychic (0.1) | −0.23 ± 0.35 (4 / 6) | +0.02 | +0.11 | −0.24 | −0.4 ± 0.1 | no |
| charAura (0.5) | +0.09 ± 0.12 (5 / 2) | +0.02 | +0.09 | −0.29 | −0.3 ± 0.1 | holds, noise-sized |
| charDebuff (0) | −0.16 ± 0.11 (1 / 4) | −0.18 | 0 | 0.00 | 0.0 | no |
| charEpic (0.5) | +0.29 ± 0.54 (8 / 3) | −0.12 | +0.48 | +0.05 | 0.0 ± 0.1 | holds, noise-sized |
| charEnh (0) | −0.29 ± 0.10 (0 / 10) | −0.15 | 0 | −0.76 at 0.3 | −0.9 ± 0.2 | no |
| charLone (0.1) | −0.16 ± 0.21 (6 / 3) | −0.13 | +0.05 | 0.00 | +0.1 ± 0.1 | no |
| charHome (0) | −0.46 ± 0.30 (4 / 7) | +0.10 | 0 | +0.31 at 0.3 | +0.4 ± 0.1 | no (Tyranids only) |
| cpOnce (0; grid 0, 0.5, 1) | 0.00 (never picked) | 0.00 | −0.57 at 1 | +0.15 at 1 | +0.1 ± 0.2 | no |
Read. What reviewers value in characters is not the kinds of rule our files list. Psykers, leaders and Lone Operatives were the cross-army misses (T-669), but a flat credit by kind doesn't fix them: any credit lifts the wrong characters as often as the right ones, and random characters given the same credit do about as well. The Epic Hero credit is the only one with a consistent lean (8 of 11 armies that moved). It works only at the grid's edge (0.5; 0.3 gives +0.04), so it looks like a step, not a slope. Two dials help the held-out Tyranid units, charHome (+0.4) and charShockAmp (+0.2, the Neurotyrant). Both are Tyranid-specific, so they belong to T-676's army layer, not to a game-wide setting. Nothing is adopted. Phase 3 (measuring what a kind does on the table, such as a squad's Deep Strike as its arrival Kill or +OC in Score) isn't worth it on this evidence, except perhaps for Epic Heroes.
How it's done: node tools/charmeasure.js (about 10 minutes, one thread). It covers each dial alone on its grid, held out three ways (3 seeds × 3 folds of 5: chosen on 10, scored on 5), with ± the SE over the 15 armies; the placebo, same procedure, mean of 3 draws; then the dials that held, together (coordinate ascent), on Track A and over the trio. It writes the settings to build/charmeasure/, and node tools/exam.js on them gives the bootstrap intervals (2,000 draws). Held-out Tyranid units: node tools/diagnose.js --on <dial>=<w> --base tools/ledger-benchmarks/2026-10-05-step34b-baseline.json (10 repeats, model (a), against the lists).