The unit tier system: a brief for a critical reader (2026-10-02)

A description, not a recommendation. The full spec is design/scoring.md. Nothing here is on the live site, whose letter is an older rule.

1. Objective

• What a letter should mean: how much a unit is worth taking, S (best) to D, in its best legal build.
• The visitor: a Warhammer 40,000 (11th edition) player choosing units for a 2,000-point tournament list.
• The stated aim: a tier list made from maths alone that agrees with expert consensus where consensus is right, and says why where it differs.
• What "right" would look like: high rank agreement with expert lists and with what winning lists take, and every disagreement traceable to a named number.

2. What we measure

Units are scored at their smallest size. Two raw numbers stay published under any version: Offense (expected wounds removed per 100 points into five stand-in targets) and Defense (hits needed to destroy the unit per 100 points from five stand-in attackers). The four jobs are built on them.

• Hammer: enemy points destroyed per 100 points, before the unit dies. Damage per turn comes from the dice engine (exact averages) into five targets (light infantry, Marine-like, Terminator-like, light vehicle, heavy vehicle), each wound priced at its stand-in's points per wound. Damage counts only on turns the unit is in range (a delivery model) and is weighted by the chance it is still alive. Hammer = 0.75 × the peak target + 0.25 × the meta-weighted mean.
• Anvil: Defense per 100 points × objective control per 100 points × a Battle-shock factor (only the share of the game spent below half strength) × regeneration (the heal closed form, or ×1.15 for revives without numbers) × a position factor. Points enter twice.
• Runner: (reach + turns alive + Leadership + 0.5 × (units per 100 points + small footprint)) ÷ 4, each a percentile, then × a points gate. Joining characters, units with no OC and one-use units score 0.
• Banner: what the unit lends others, per point: a leader's gain on its best squad against spending the same points on more bodies; buff auras run through the engine on the two squads most often listed with it; transport capacity; enemy debuffs; forced Battle-shock; Command points; spawned units.

3. How a letter is made

• Percentiles: each job is ranked over all 1,161 non-Legends units of every army (a datasheet shared by armies is a row in each).
• Best job plus a bonus: the score is the unit's best allowed job. In round 7 each other allowed job above 60 added 0.25 a point, capped at 100. It is now being set to +5 for a second job at 90 or more, +2 for 80 to 89, nothing below, no cap. A weak job never lowers a unit.
• Letters: fixed cuts in the all-armies view; by rank position in the within-army view. (Round 7 as run used rank quotas, S 7%, A 15%, B 28%, C 30%, D the rest, with S and A split across jobs by their share of winning lists' points.)
• Context rule: each unit is scored in each of its army's detachments and keeps the best; its best loadout from a legal-wargear solver.
• Transport bundle: a unit may be scored with its best Dedicated Transport, priced together, only if the pair beats both the passenger and the transport alone per point.
• Delivery and position: five rounds on a 44" × 60" table, units 30" apart. Ways in: walk (with Scouts or Infiltrators), Deep Strike (turn 2, 9" away), Strategic Reserves (turn 2, 18"), tunnels (7"). A position track follows each way in; objectives sit 15" (midfield), 21" (enemy half) and 27" (enemy zone) out.
• What counts as on: army rules on; Oath of Moment at full share; a CP worth 30 points; Battle-shock 10% of a median unit; enemy debuffs at half.

4. What we validate against, and the numbers

• One expert list (Tyranids): Auspex Tactics' YouTube tier list, 44 graded units. Round 7: Spearman ρ 0.50 (SE about 0.12). Same letter on our quota ruler 20 of 44 (shuffle baseline 10.3); same raw letter 12; within one letter 29 (baseline 20.4).
• History of ρ with that list: 0.07 (V2b), 0.21 (V2c), 0.19 (V4), 0.38 (V5), 0.37 (every army scored), 0.34 (round 6), 0.39 (6b), 0.50 (7). Today's live letter, which includes list share, scores 0.65.
• List share: the share of top-quarter tournament lists (grimstat-corpus, July to September) that take a unit, among lists that could. Tyranids ρ 0.39 (52 units, 23 lists); every unit 0.164.
• Circularity: list share is popularity, not strength, and it also feeds the system: target weights, incoming fire weights, the S/A job split, which squads an aura is tested on, the Battle-shock fail rate and, in V1, Banner's fitted weights.
• Check panels, never used to steer: Space Marines (Tista Minis, 30 units) ρ −0.05 (SE 0.19); Orks (Sprues & Brews, 45 units) ρ −0.13 (SE 0.15). Their letters are our mapping of each author's sections.

5. Known gaps

• Engine mechanics not built: fighting after being slain; Fights First from abilities; Precision and Anti as modifiers; damage re-rolls; keyword exclusions; repeated arrival from reserves (one arrival is played); spawning beyond Spore Mines; the Norns' Singular Purpose is approximated from a hand file.
• Sources: Crucible characters are unread in every army (absent from the 11th-edition pages we read).
• Stratagems: not modelled; a CP is a flat 30 points until each unit's usable stratagems are priced.
• No board: no terrain, line of sight always assumed, no opponent choices.

6. Assumptions

• Every job is valued per point.
• Four jobs cover what a unit is bought for.
• A unit's best job sets its band; other jobs only add.
• A unit is scored alone, at its smallest size, in its best detachment.
• Leaders do not lift their squad's letter; the leader is credited instead.
• Five stand-in targets and six attackers represent the enemy.
• The winning lists' mix is the right weighting of targets and incoming fire.
• Five turns on a 44" × 60" table, objectives at 15", 21" and 27".
• Every shot has line of sight; no terrain.
• Turns alive do not depend on the unit's cost.
• The median unit (120 points) is the yardstick for Battle-shock and aura worth.
• A CP is worth 30 points for every unit and army.
• Army rules and Oath of Moment are always on.
• Conditions the engine cannot check hold for a fixed share of the game.
• An aura reaches the two squads most often listed with its bearer.
• Auspex's list is a fair stand-in for expert consensus.
• List share measures worth, not only fashion.
• Every unit in the game is the right comparison field.
• Letters are absolute across armies.
• Reserves have no list-level cap.
• One panel's ρ difference under its standard error is noise.

7. Levers we have

• The constants (table in section 2).
• The job definitions and which jobs each kind of unit may hold.
• The bonus rule for second jobs.
• The letter cuts or quotas, global and within-army.
• The target and attacker set and their weights.
• The position track: distances, objectives, floor.
• The panel: which expert lists, and how many.

How to use this brief

Return at most ten critiques, ranked by how much each would change the letters if acted on. For each, say what we would observe in the data if the critique is right (a unit, a group of units or a yardstick that should move, and which way) and the cheapest test that would show it: a constant to change, a switch to turn off, a subset to compare, or a second source to read. Gaps this brief omits count too.