Tyranids, round 9: Anvil divides by points once, a survival gate, three panels (2026-10-02)

T-562 (Jordan, 2 Oct, via the design session). Round 8's run (step grading, fixed cuts, CP worth 30) with Anvil changed: one division by points instead of two, and an absolute survival gate. Switch r9 over r8 (tools/tiertest.js, round 8's command plus --r9json <file>, and --sweep for the one-at-a-time sensitivity sweep); round 8's columns come from the same run, and round 8's JSON reproduces byte for byte. Answers points 4, 5 and 9 of the independent critique (reports/tiers/2026-10-02-brief-critique.md). A report: nothing here changes a published score or page.

1. Three panels (from round 9 on, every report leads with these)

ρ: Spearman's rank correlation between our score and the panel's letters (SE: Bonett and Wright's). Strict: the same letter as the panel re-cut on our ruler (within-army view; round 7 had one view); raw: the panel's own letters, all-armies view; within one: raw letters, all-armies view. Shuffle baselines in brackets (2,000 permutations). The Space Marine and Ork panels have never been used to tune anything.

Reading it. Round 9 moves the Tyranid ρ 0.47 → 0.49 and the Space Marine ρ −0.18 → −0.10, while Orks stay at −0.12; every change is inside one SE. The two clean panels are still at or below zero. Every unit's list share falls 0.162 → 0.077: the cheap bodies that winning lists take a lot of (and that round 8 put at Anvil 99 to 100) lose their automatic place at the top.

2. The Tyranid table, all 44 units Auspex grades

Sorted by Auspex's tier, then name. Round 8 and Round 9 (current cuts): all-armies letters on the fixed cuts (S ≥ 95, A ≥ 88, B ≥ 80, C ≥ 65). Proposed cuts: section 6 (S ≥ 98.79, A ≥ 94.49, B ≥ 86.13, C ≥ 71.15). Within army: round 9's scores ranked inside the Tyranids (top 10% S, 15% A, 25% B, 30% C, rest D). Score and the four job percentiles are round 9's, each unit at its best detachment. Lists: Tyranid winning lists (of 23) taking it. What moved: round 8 → round 9 at the current cuts, as the sum that makes the score, with Anvil's percentile before → after.

The 8 Tyranids Auspex doesn't grade:

13 of 52 Tyranid letters move round 8 → round 9 at the current cuts; the proposed cuts move 25 more against round 9's current letters.

3. Anvil, before and after

Round 8: Defense per 100 pts × OC per 100 pts × Battle-shock × revives × position, i.e. hits to destroy × OC ÷ points², so halving a unit's cost quadruples its Anvil. Round 9: hits to destroy × OC ÷ points (Defense per 100 pts × the unit's OC), one division, × the same Battle-shock, revive and position factors, × the survival gate: the turns the unit lives under a typical enemy turn of fire (Hammer's own model, turns alive Tu: absolute hits to destroy against half the median unit's, the six attacker classes weighted by the meta's incoming fire), removed in one turn × V9C.GATE1 = 0.5, in two × V9C.GATE2 = 0.8, three or more × 1. OC 0 still scores 0.

Why not the geometric mean. The critique offered √(Defense per 100 pts × OC per 100 pts). That is √(hits × OC) ÷ points, the square root of round 8's product, and a percentile only reads the order, so it ranks every unit exactly as round 8 did (only the other factors' leverage changes). The single-division product is the one that changes anything, and it is one line in the code (hold9 in raw5).

Raw numbers are each unit's at its best detachment and way in, so they include detachment OC (Crusher Stampede's +2 for the Tyrannofex) and position. The gate: over all 1161 units, 693 live three turns or more (×1), 224 two (×0.8), 244 one (×0.5); Tyranids 35, 12 and 5. A minimum-size horde of ten lives just over two of these turns (Termagants 2.09), so the gate doesn't bite the gaunts; what moves them is the single division. Bodies still score well (Termagants are S on Anvil 100 in round 8 and 96 now), but Neurogaunts fall 99 → 89 and the big, hard units rise (the Norns 96 → 100, the Tyrannofex 92 → 96). A unit's minimum size now matters: the measure grows with the models in it (twice the bodies, twice the hits and twice the OC for twice the points), so large minimum units gain.

4. Sensitivity: each hand-set constant, one at a time

Round 9 with one constant changed at a time, everything else as round 9. Letters: units of all 1161 whose all-armies letter (current cuts) differs from round 9's, low / high (Tyranids among them in brackets). ρ: the Tyranid ρ with Auspex at low / high (round 9: 0.494). Values: ×0.5 and ×1.5 unless the range says otherwise. The cuts only re-letter round 9's scores. Sorted by letters moved, both runs together. Restoring every constant gives round 9 back (0 letters differ).

Not swept: V3C.ROUNDS (five battle rounds is the game), V3C.GAP (the table), V5C.MW_D and V4C.HV_W (read once when the pool is built, before any round runs), and the transport constants the round-9 path doesn't read (V2C.RIDE_COST, V3C.SHIELD, V3C.CAP_PER: V2c's transport credit is worked out before round 5 and kept). V4's delivery profile (which feeds the debuffs' rounds in range) is also V4's, not re-run per value. The data-file shares (Crusher Stampede's 0.9, the condition shares in detCtx other than Synapse and below-half) are in the effects files, not constants.

The five constants that move most (the letter cuts aside: they are the ruler, section 6): Runner's points gate (75 and 260 pts together) (399); Advance (3.5") (288); The objectives' distances (15", 21", 27" together) (286); Strategic Reserves' arrival distance (18") (206); Runner's half-weight parts (0.5) (183). Most of the leverage sits in geometry and Runner (the points gate, how far a unit moves, where the objectives are), not in the worths we argued over: a CP's worth, Battle-shock's, the debuff share and Synapse coverage move 3 letters or none. Those are read only for the Tyranids (tierfx.json), so they can move at most 52 units, and they move scores without crossing a cut (the Tyranid ρ still moves: CP 15 gives 0.481). The one-turn gate moves nothing because the units it catches (Tu ≤ 1: lone characters, spores) have OC 0 or 1 and never hold as their best job; the two-turn gate at 0.4 lifts the Tyranid ρ to 0.530 (it lowers Warriors and the infiltrators' Anvil) and is the only constant here worth a look on that score, inside one SE. No Tyranid ρ in the sweep leaves 0.464 to 0.530.

5. Every version against all three panels

Every version rerun from today's code and data in one run (none had to be taken from its report). ρ (SE) for each panel, and every unit's ρ with its list share (units in 5 or more winning lists of their army). The reruns match the reports' Tyranid ρ for every round; V2b to V4 read today's rankings (T-546's Toughness), so their reports' numbers can differ in the second decimal.

What it shows. From V2b to round 9 the Tyranid ρ climbs 0.07 → 0.49; Space Marines go −0.24 → −0.10 and Orks 0.06 → −0.12. The critique's prediction (the tuned panel climbs, the clean ones stay flat or fall) holds: the climb is on the panel we steered by. Round 9 is the first change since round 7 that lifts a clean panel, and only within one SE.

6. Cut calibration (a proposal for Jordan; the default is unchanged)

Round 8's fixed cuts (S ≥ 95, A ≥ 88, B ≥ 80, C ≥ 65) give, under round 9's scores: S 247, A 268, B 243, C 248, D 155. A best-of-four percentile score puts many units at 95 or more, so S is about a fifth of the field. Proposed (set once and freeze as absolutes): the scores that give S 8%, A 15%, B 27%, C 30%, D 20% of all 1161 units under round 9: S ≥ 98.79, A ≥ 94.49, B ≥ 86.13, C ≥ 71.15, D below; they give S 93, A 173, B 315, C 347, D 233. After freezing, a later round that lifts every score would show as more S, not be re-cut away. The Tyranid table (section 2) shows both.

Against the panels (all-armies letters, proposed cuts): Tyranids strict on our ruler 21, raw 13 (chance 7.4), within one 28 (chance 21.0); Space Marines raw 10, within one 22; Orks raw 10, within one 27. On the current cuts: Tyranids raw 18, within one 31.

S per army. ρ between an army's S share and its top-quarter lists (June to September, data/metatrack.json; the 29 armies with 10 or more units): round 8 −0.37 (round 8's report: −0.37), round 9 on the current cuts −0.39, on the proposed cuts −0.24. Negative means the armies we hand the most S to are the ones that place least.