The pooled readable model, version 2: design/jobs.md's variables (T-704)

5 Oct 2026. Discovery, not adoption: no ledger value, setting, letter or page moves. Version 1 (reports/tiers/2026-10-05-readable-model.md) is unchanged and reproducible (tools/readable.js --v1). The units, pairs, model, shrinkage, bootstrap, controls and the frozen snapshot (d8b5a1bb14, inputs hash c5e7d85b43f8) are version 1's. Only the variables are new: design/jobs.md's priorities 1 to 25 (T-707), five interactions and seven saturation ranks. Track A is step 34b (wCarry 0).

Tournament list share is the second target, for discovery only. It comes from grimstat-corpus (https://github.com/N041M/grimstat-corpus, CC BY 4.0): 487 11th-edition lists from 30 events, 27 June to 26 September, published by their players and tournament organisers on MiniHeadQuarters. We read counts only. List share is never in a letter.

In plain English

What was fitted.

• The same 595 graded units (733 ranked) and 9,354 consensus pairs as version 1. Single-list armies count at half weight.
• 42 columns:
  • jobs.md's 25 jobs (two of them, the army-rule engine and the transport, in two parts), plus three parts of their recipes: points, Fly and "can be Hidden";
  • five interactions players reason about;
  • seven saturation ranks (is it the army's best at a job that a list fills once or twice?);
  • a placebo.
• Each column is measured against the unit's own army (one army-SD), as in version 1.
• All 25 jobs were built. The five marked "partly" in jobs.md got these small proxies:
  • the army-rule engine: armyrules.json's enabler flag, plus the friends' Kill each unit is measured to add (Synapse coverage and support effects);
  • the action-doer: an action's price (the field's primary action lines × points a VP) less the ranged Kill it gives up;
  • VP denial: the kill-card VP a unit hands over when it dies, per point;
  • mitigation: heal and leave-and-return (Tyranids and Space Marines only), Feel No Pain and heal or spawn rules;
  • screening: units per point that still cost about one enemy unit's firepower to clear.
• Left out as too much for this pass: OC removed per activation, turn-one objective reach, the transport's protection, the screen's field-arrival share and a unit's chance of dying in a game. Each needs a new field from burst.js or deployments.json.

Does version 2 learn more than version 1? No. It reshuffles where the signal is found, but no more of it transfers.

• Leaving each army out and predicting its order from the other 15 (army mean, 16 armies):

| | Army mean, 16 | 15 (no Tyranids) | Every pair pooled | |---|---:|---:|---:| | Track A | 61.9 | 60.1 | 62.5 | | Version 1 | 60.1 | 59.5 | 58.5 | | Version 2 | 59.8 | 58.9 | 59.7 | | Version 2, main effects only | 59.7 | 58.6 | 60.2 | | Burst alone | 58.2 | 57.4 | 59.6 | | Version 2 + Epic Hero and Psyker (version 1's flags) | 61.9 | 61.1 | 60.6 | | Version 2 + Track A's Value | 60.4 | 59.4 | 60.3 |

• Scrambled letters through the identical pipeline give 50.2 (20 seeds, range 45.7 to 55.0; none reaches 59.8). The placebo comes out at +0.01 [−0.01, +0.04], with no army's deviation clear of 0. So the signal is real and it transfers, but it is the same ~60% ceiling as version 1 and T-635's 341 variables.
• The interactions and the saturation ranks add nothing held out: the main effects alone score 59.7, the same as the full model.
• The two flags jobs.md left out matter most. Adding version 1's Epic Hero and Psyker back lifts version 2 by 2.1 points, level with Track A (61.9). Without the Tyranids it beats Track A (61.1 against 60.1). Thousand Sons alone go from 55.0 to 67.6. Jobs.md has no "named character the army is built around" variable; this says it should.
• Beside Track A's Value, version 2 does worse than version 1 did (60.4 against 62.1). With 42 columns the chosen shrinkage is three times heavier (λβ 300 against 100), and Value's own weight is shrunk with the rest.
• By army, version 2 beats version 1 in 8 of 16 armies and Track A in 7:
  • better: Leagues of Votann +11.4, World Eaters +6.4, Space Marines +5.4, Orks +5.0, Emperor's Children +4.7, Astra Militarum +3.9, Tyranids +3.6;
  • worse: Blood Angels −25.0 (124 pairs, one list: 30.6%, worse than chance), Thousand Sons −10.0 (Psyker carried them), Adeptus Mechanicus −6.8, Death Guard −3.6.
• List share is where version 2 adds. Held out it predicts players' picks at 61.2%, against version 1's 58.6 and Track A's 59.8. Its board variables (OC, holding, command points, squads a leader lifts) are what players take.

What reviewers appear to value. Each figure is the weight per army-SD (β) with its 95% bootstrap interval and the share of 200 bootstraps above 0. Burst shares its credit with the four target-class Bursts, which correlate with it at 0.7 to 0.9, so read the damage block together.

• Clearly valued (interval above 0):
  • Damage into the hard targets: anti-tank +0.05 [+0.03, +0.06], 16 of 16 armies the same sign; anti-elite +0.04 [+0.02, +0.05], 16 of 16;
  • horde clearing +0.03 and Burst +0.02;
  • character hunting +0.01 [+0.00, +0.03];
  • reserves and Deep Strike +0.04 [+0.01, +0.06];
  • the army-rule enabler +0.03 [+0.02, +0.05];
  • can be Hidden (all Infantry, Beasts or Swarm) +0.03 [+0.01, +0.05];
  • command points +0.03 [+0.00, +0.05];
  • being the army's best horde-clearer, beyond the amount (saturation rank) +0.03 [+0.01, +0.04]; the best anti-tank unit +0.02 (97% above 0).
• Clearly marked down:
  • Fly −0.04 [−0.06, −0.01] (13 of 15 armies), as in version 1;
  • the cheapest bodies and the cheapest action-doers of each army, beyond what their cheapness and OC say: saturation ranks −0.03 [−0.04, −0.01] (15 of 16 armies) and −0.03 [−0.04, −0.01] (16 of 16). Reviewers don't reward being the army's cheapest unit. The opposite of jobs.md's "cheap activations" hypothesis, at least for letters.
• Leaning, 85 to 97% of draws above 0:
  • OC kept steady +0.02;
  • holding power +0.02;
  • hard to target +0.02;
  • forward deployment +0.02;
  • the engine measured +0.02;
  • durability × OC +0.02;
  • transport × rider +0.01;
  • points +0.01;
  • list dependence +0.01.
• No sign either way:
  • durability per point on its own (14% above 0, leaning down);
  • melee delivery;
  • threat range;
  • transport capacity;
  • slot competition among siblings (13%, leaning down as expected, but weak);
  • Battle-shock pressure;
  • the action-doer;
  • VP denial;
  • mitigation;
  • screening;
  • Burst × Soak;
  • delivery × melee damage (6%, leaning down).
• "Unlocks a squad": +0.01 [−0.01, +0.03] (84%), and the squad's side +0.00. The package synergy, measured as jobs.md asked, carries little for reviewers. T-705's stated tag (+1.0 letter) is on 21 units, so the reason is real for those units, but our joined-form Value doesn't pick out which leaders reviewers mean.

Where reviewers and players differ (the same model on list share; read as leans):

• Players take what leaders lift: the squad's side of the leader package is +0.10 [+0.06, +0.14] for list share, 0.00 for reviewers, and the leader's side +0.04 against +0.01. Lists are built around joined pairs more than letters show.
• Players weigh board control more:
  • OC kept steady +0.07 against +0.02;
  • holding power +0.06 against +0.02;
  • command points +0.09 against +0.03;
  • durability × OC +0.05 against +0.02;
  • hard to target +0.06 against +0.02.
• Players take units that need one detachment more: list dependence +0.06 against +0.01. A list built for a detachment takes its specialists.
• Both value the damage block, reserves, Hidden and enablers about equally. Both mark down Fly and the cheapest bodies' and action-doers' rank.
• Reviewers lean on points (+0.01) and players don't (0.00), as in version 1.

Which armies differ (the clearest deviations; the placebo's clears 0 in none of the 16):

• Space Marines: durability × OC, anti-elite, horde clearing and screening count more.
• Tyranids: Fly is marked down harder; forward deployment, anti-tank and command points count more.
• World Eaters: anti-elite and Hidden count more, and Fly is marked down harder.
• Astra Militarum: anti-tank and durability count more; action-doers and cheap bodies count less.
• Genestealer Cults: OC counts more; melee delivery and the cheapest bodies count less.
• Dark Angels: Fly is not marked down; durability is marked down.
• Orks: enablers (the bosses) count about four times the shared weight.
• Death Guard: Deep Strike counts about four times the shared weight.

Read with care.

• Heavy shrinkage (λβ 300) and strongly correlated columns share credit:
  • OC kept steady with holding power (0.99 within armies);
  • cheap activations with screening and points (0.85, −0.89);
  • the Bursts with each other;
  • transport with transport × rider. Read a group's signs together, not one column alone.
• Saturation ranks are reciprocal ranks (1, ½, ⅓…): they reward the army's top unit at a job. A negative weight on the cheapest-body rank says the army's single cheapest unit grades below its measured worth.
• The interactions are products of the two parts' within-army z-scores, so a unit low on both scores positive.
• 12 of 16 armies have one 11th-edition list. Blood Angels' collapse rests on 124 pairs from one reviewer.
• Recipes chosen alone are listed in the handoff (reports/handoff.md). They include the 30 points a CP, the kill cards' VP and the 1.2 for a heal or spawn rule.
• What it suggests (Jordan's call; nothing adopted):
  • jobs.md's variables explain list share better than version 1's, and reviewers no better;
  • the one thing they miss is the named centrepiece (Epic Hero, Psyker);
  • a version 3 might keep version 2's main effects, add a "centrepiece" variable and drop the interactions and most saturation ranks, which add nothing held out.

How it's done.

• node --max-old-space-size=3000 tools/readable.js --boot 200 --scrambles 20 --json reports/tiers/2026-10-05-readable-model-v2.json --md <file> runs in about 12.5 minutes on one thread.
• The sections below are the tool's --md output; the section above is written by hand.

The coefficients: reviewers and list share side by side

Pairwise logistic, shared across the armies, each variable standardised within its army (one army-SD). β with its 95% bootstrap interval (200 draws of units within armies) and the share of draws above 0. Burst, version 2's nearest to Kill, shares its credit with the four target-class Bursts (correlation 0.7 to 0.9 within armies), so its own β is small and ratios to it are unstable; read the β themselves (the JSON keeps the ratios). Reviewers: λβ 300, λδ 30. List share (grimstat-corpus, CC BY 4.0; discovery only): λβ 300, λδ 300. "Armies same sign": armies where β + the army's deviation has the shared sign, of those where the variable varies. A flag's median within-army SD is beside it (having the flag ≈ β ÷ SD).

Where armies differ (reviewers): per army, the size of its deviations (root mean square of δ over the variables) and its largest deviations whose bootstrap interval leaves 0 (at most four, |δ| ≥ 0.03). Calibration: the placebo's deviation clears 0 in 0 of 16 armies.

Leave one army out: version 2 against version 1 against Track A

Fitted on 15 armies (shared coefficients only, λβ chosen inside the 15), the 16th army's order predicted and scored on its consensus pairs; Track A's Value on the same pairs (ties wrong). Version 1 is the same pipeline on version 1's 25 variables. Tyranids: Track A was tuned on them.

Version 2 beats Track A in 7 of 16 armies (version 1: 6), and version 1 in 8 of 16. Without the placebo column: army mean 60.0. Beside Track A's Value, version 2 beats Track A alone in 9 of 16 armies.

The scrambled-letter control

Letters permuted among each army's graded units (all its lists together), 20 seeds, through the identical leave-one-army-out pipeline, version 2's variables.

Scrambled seeds at or above the real letters' 59.8: 0 of 20.

The placebo column (random per unit): β +0.01 [−0.01, +0.04], 0 armies with a deviation clear of 0.

List share, held out (discovery only)

The same model on 11th-edition list share (pairs: one unit taken by 2 or more lists more than the other; armies with 5+ tournament lists). grimstat-corpus (https://github.com/N041M/grimstat-corpus, CC BY 4.0); counts only.

The variables (version 2: design/jobs.md priorities 1 to 25, T-707)