Tactical Reroll
⋯

The three previews measured (T-682 rivals, T-683 typical version, T-684 Gargoyles' tricks)

From reports/tiers/2026-10-05-previews-measured.md , rendered when the site is built.

In plain English. We tried three ideas from Jordan's Tyranid notes as game-wide dials, each on top of Track A, and asked whether they make our rankings agree better with the reviewers on armies they were never tuned on. None of them is a clear win. One part of one idea, judging a unit by the average of its legal squad sizes rather than its best size alone, nudges agreement up a little (about +0.3 points held out, +0.4 on the full exam) and beats its placebo, so it passes our bar, but it is inside the noise. Everything else is flat or worse:

No ledger value, setting or ledger file changed. tools/previews.js does the sweep (node tools/previews.js, then --analyse build/previews.json --pairs); tools/exam.js --post gives the intervals.

Method. Track A (tools/ledger-benchmarks/2026-10-05-step34b-baseline.json, wCarry 0), snapshot 8eb69bc7af (inputs hash a20dbd2c3316). Each army is computed once and passed through each preview's exam hook (T-685) at strengths 0.25, 0.5, 1, re-sorted as the exam re-sorts, and scored on its 11th-edition lists (T-691): pooled over 15 armies (Track A 59.1), the Tyranids' three 11e headline lists beside (82.4). Held out by army: T-670/T-675's three seeded splits of the 15 armies into 10 (choose the strength, 0 included) and 5 (score the gain over Track A); mean ± SE over the three splits. Placebos, each through the same held-out choice, mean of seeds 1 to 3: rivals with the likeness matrix scrambled among the army's units; typical with random spans placed by width (each unit keeps its own span, best form down to its weakest own solo form, and the fractions of the span the dial moves are dealt to other units at random, so the size of the moves is kept and only which unit is "typical" is forgotten); steal and jail with each dial's qualifying worth dealt to as many random units of the army. The bar: a dial helps only if it holds out above 0 and beats its placebo.

T-683's size rule (built here). tools/typical.js typicalSizeRule: smallest (phase 1's typical squad, the default), mean (per detachment, the mean over every size the unit has a solo form at, each at its datasheet wargear) and best (the size of the unit's best solo form, at its datasheet wargear: only the wargear goes typical). Every rule at 0 is today to the bit (tests/unit/typical.test.js).

Rivals (T-682, rivalW)

rivalWPooled 15 ΔPlacebo ΔTyranids Δ (82.4)PlaceboArmies better
0.25−0.5−0.2+0.0−0.05/15
0.5−0.3 [−1.6, +1.1]−0.8−0.5−0.37/15
1−0.3−2.0−3.4−3.06/15
Held out−0.73 ± 0.73 (picks 0; 1; 0)+0.00 ± 0.00 (0; 0; 0)

The real likeness does beat its scrambled one at 0.5 and 1 (it costs less), but never beats Track A: the cut mostly shrinks weak units the reviewers already rate low. Biggest swings at 0.5: Votann +4.8, Agents +2.7 against Blood Angels −4.8, Drukhari −3.9. Not kept. Each army's top rival pairs are listed at the end for Jordan to judge.

Typical version (T-683, typicalSize × typicalDet × typicalSizeRule)

Setting (size rule, typicalSize, typicalDet)Pooled 15 ΔPlacebo ΔTyranids ΔPlaceboArmies better
det only 0.25 / 0.5 / 1−0.1 / −0.2 / −0.3+0.0 / +0.1 / −0.6−0.2 / −0.1 / −0.05/15 each
smallest 0.25 / 0.5 / 1, det 0+0.2 / −0.1 / −0.6−0.2 / −0.1 / −0.3−0.0 / +0.1 / +0.19, 8, 7/15
mean 0.25, det 0+0.4 [−0.3, +1.2]−0.3+0.2+0.08/15
mean 0.5, det 0+0.3 [−0.9, +1.7]−0.2+0.1+0.19/15
mean 1, det 0−0.8−0.6+0.5+0.27/15
best 0.25, det 0+0.4 [−0.3, +1.2]−0.2+0.1+0.18/15
best 1, det 0 / det 1−0.8 / −0.7−0.6 / −1.0+1.0 / +1.2+0.1 / −0.27/15
any size with det 0.25 to 1−0.9 to +0.2
Held out (choose the setting on 10, score on 5)RealPlacebo
all three rules, both halves+0.23 ± 0.32 (mean 0.5; best 0.25; mean 0.25, det 0 each)−0.38 ± 0.31
smallest only−1.61 ± 0.61−0.82 ± 0.23
mean only+0.26 ± 0.35 (mean 0.5; 0.25; 0.25, det 0 each)−0.02 ± 0.10
best only−0.01 ± 0.57−0.19 ± 0.11
detachment only−0.38 ± 0.38+0.02 ± 0.02

Only the size half helps, and only read as the mean of a unit's sizes (or, about as well, as its best size at datasheet wargear); every fold picks typicalDet 0. It passes the bar (held out above 0, placebo below) but by less than one standard error, and the exam's interval crosses 0 (+0.4 [−0.3, +1.2], armies resampled [−0.1, +0.9]). Gains: Blood Angels, Thousand Sons, Drukhari, Orks, Death Guard (+1.9 to +4.0 at mean 0.5); losses Adeptus Mechanicus (−5.3 at mean 0.5), Votann, GSC. On Tyranids it barely moves at 0.5 (17 of 52 units move, mean 0.5 places): Tervigon 21 → 25, Hormagaunts 18 → 21, Termagants 38 → 41; Carnifexes stay 32nd (on Track A their single and pair are worth the same per point, so Jordan's Track B case does not arise here).

Gargoyles' tricks (T-684, jailW, stealW, apart)

DialPooled 15 Δ at 0.25 / 0.5 / 1Placebo ΔTyranids ΔPlaceboArmies betterHeld outPlacebo held out
jailW+0.0 / −0.0 / −0.1−0.0 / −0.0 / +0.1+0.0 / +0.0 / +0.0−0.0 / −0.0 / −0.23, 4, 3/15−0.00 ± 0.01 (0.25 ×3)+0.02 ± 0.04
stealW−0.0 / −0.0 / −0.2+0.0 / +0.1 / +0.1+0.1 / +0.5 / +0.4+0.1 / +0.1 / −0.02, 1, 1/15+0.00 ± 0.00 (0 ×3)−0.02 ± 0.03

The steal beats its placebo on Tyranids (+0.5 against +0.1; Tyranids all five lists +0.6) and nowhere else: of the other qualifying datasheets (Eliminators, Land Speeder, Pteraxii Skystalkers, Scourges), Dark Angels gain +0.1 and Adeptus Mechanicus lose −0.6. Exam at stealW 0.5: −0.0 [−0.2, +0.1]. Jailing does nothing. Neither is a game-wide dial; the steal could return as a Tyranid rule under T-676.

For the cascade

DialKeep?StrengthsFast path
rivalWno––
typicalSize, rule meanyes, weakly (inside noise)0, 0.25, 0.5 (rule best at 0.25 as an alternative)a post step after each run's form values (corners per unit, no new compute()); needs adding to crack2core
typicalDetno (0)––
jailWno––
stealWnot game-wide (Tyranids only)0, 0.5 if T-676 takes itnot needed for the game-wide cascade

Top rival pairs at Track A (for Jordan)

Each unit's rival is the same-army unit whose pull (likeness × how much better per point, capped at twice) cuts it most; letters are every panel the army's ledger carries (10th-edition lists too). Six per army.

What the pairs show: the likeness pairs characters with characters across very different roles (Chaplain with Judiciar, Memnyr Strategist with Grimnyr, Navigator with Inquisitor), and the Cults' ledger carries Astra Militarum allies (Cadian Castellan, Krieg Command Squad) as rivals. Jordan's own Tyranid pairs (Winged Hive Tyrant and Hive Tyrant, the two Warriors, Leapers and Lictors) are not the top pulls. If Jordan sends rival pairs by hand, they can be scored against these.