Tyranids, round 10: Anvil as OC-turns per point, a fieldable pool, a final rank on frozen cuts, the list-share headline, the explainer (2026-10-02)

T-563 (Jordan, 2 Oct, via the design session): the method appraisal's experiments 1 to 4 (reports/tiers/2026-10-02-method-appraisal.md) and an explainer. Switch r10 over r8 in tools/tiertest.js (V10C), run as round 9's command plus --r10json <file>; rounds 8 and 9 come from the same run. Round 9's survival gate is gone from the code, so round 9 is shown as published (its JSON) and, where noted, rerun without the gate. A report: nothing here changes a published score or page.

The four experiments against their success numbers:

In short. The new Anvil gives back what round 9 lost on list share (pooled 0.077 → 0.160, per-army median 0.113 → 0.176, AUC 0.536 → 0.584, all close to round 8) while keeping round 9's Auspex gain (0.51) and taking the Titans out of S: no unit over 1,000 pts is S, and Anvil no longer tracks cost the way round 9's did (ρ with points 0.41 → 0.18). It misses its last success number, though: 0.18 is outside ±0.15, because turns alive still rise with a unit's size and toughness, so mid-priced monsters keep a lead over cheap infantry. Orks fall from −0.12 to −0.18 and Space Marines rise from −0.10 to −0.07, both inside one SE. Cutting the pool to fieldable units moves no panel by 0.01, as expected. The final rank plus frozen cuts puts S at exactly 8% and lifts the S-per-army check from −0.39 to −0.16. That is better, but still negative: the armies with the most near-identical datasheets (the Space Marine chapters, Deathwatch) still collect S. On the all-armies view the Tyranids now have one S (Termagants), so the within-army letters are the view to compare with Auspex: 13 units are two or more letters off there, against 14 in round 9.

1. Yardsticks: three panels and the list-share headline

ρ: Spearman's rank correlation between our score and the panel's letters (SE: Bonett and Wright's). Strict on our ruler (within-army view) / raw (all-armies letters) / within one (all-armies letters), shuffle baselines in brackets. The list-share headline (from round 10 on, the appraisal's experiment 2): over the units some winning list could field (armies with 5 or more winning lists): the median of each army's ρ with its list share (27 armies with 10 or more such units), the Tyranids' own ρ beside it (the overfit gauge: the panel we tuned on against everyone else), the AUC for units winning lists take against units they never take (0.5 is chance), the hit rate (share taken among our top quarter, against the base rate), and ρ among taken units only. Round 9 as published (with its gate); round 9 without the gate is this code's rerun.

Check: the report's own headline code on round 10's final scores gives median 0.165, AUC 0.582, taken 0.098 (the tool's, on the composite: 0.165, 0.582, 0.098; a percentile changes only ties).

Reading it. Round 10 is the first round to lift the tuned panel and list share together: Tyranids 0.49 → 0.51, per-army median 0.113 → 0.165, AUC 0.536 → 0.582, ρ among taken units 0.066 → 0.098. Space Marines move up and Orks down, both inside one SE. The all-armies raw count for the Tyranids falls (18 → 12) because the frozen cuts give S to 8% of every army pooled, while Auspex hands S to 13 of 44 Tyranids; within the army, strict on our ruler is back at 19. Removing round 9's gate alone changes little (Space Marines −0.10 → −0.14, nothing else past the second decimal), so the rerun's round 9 is a fair stand-in.

Every version on the headline. Rerun from today's code in one run (round 9 without its gate; round 9 as published from its JSON). The pooled ρ is the old headline, beside the new ones.

Order of the versions. By pooled ρ: Round 6b (CP 30) > Round 6 (CP 0) > Round 7 > V3 > Round 8 > Round 10 > Every army (T-552) > Round 5 (V5) > V4 > V2b > Round 9. By per-army median: Round 6 (CP 0) > Round 6b (CP 30) > V3 > Every army (T-552) > Round 8 > Round 7 > Round 10 > Round 5 (V5) > V4 > V2b > Round 9. Spearman between the two orders over the 11 versions: 0.91; with the AUC: 0.95.

What the headline shows. The two orders agree broadly (round 9 last on both, rounds 6 and 6b near the top) but differ in the middle. T-552 is 7th by pooled ρ and 4th by the per-army median, and round 7 slips from 3rd to 6th. So the pooled number has been ranking versions partly on army-level offsets. The AUC tells the same story as the median: rounds 6 and 6b separate taken from never-taken units best (0.59), and every round since has sat at or below that. The overfit gauge: the Tyranids' own ρ was below the median until round 5. It climbed to a +0.34 lead in round 9 and is +0.30 in round 10. The method has been learning the Tyranids faster than it generalises. Round 10 is the first round in which the median rises more than the Tyranid ρ does (+0.05 against +0.01).

2. The Tyranid table, all 44 units Auspex grades

Sorted by Auspex's tier, then name. Round 9: as published (all-armies letters on round 8's fixed cuts S ≥ 95, A ≥ 88, B ≥ 80, C ≥ 65). Round 10 all armies: the final score (the pool percentile of the composite, over the 1153 fieldable non-Legends units) on the frozen cuts S ≥ 92.02, A ≥ 77.02, B ≥ 49.96, C ≥ 20.03. Within army: round 10's composite ranked inside the Tyranids (top 10% S, 15% A, 25% B, 30% C, rest D). The four job percentiles are round 10's at the unit's best detachment (bracketed: a job it may not hold). Lists: Tyranid winning lists (of 23) taking it. What moved: round 9 → round 10, the composite before → after, and Anvil's percentile when it moved by 3 or more.

The 8 Tyranids Auspex doesn't grade: Barbgaunts D → C (within C, 28.1); Broodlord B → C (within C, 41.5); Gargoyles A → A (within S, 88.6); Genestealers C → C (within C, 26.0); Lictor S → A (within S, 87.2); Termagants S → S (within S, 93.8); Toxicrene B → C (within B, 44.9); Winged Tyranid Prime S → A (within A, 83.7).

25 of 52 Tyranid letters move round 9 → round 10. Two or more letters from Auspex: 13 in round 9 (all armies), 18 in round 10 all armies (Raveners, Hyperadapted Raveners, Von Ryan's Leapers, Zoanthropes, Maleceptor, Tyrannofex, Norn Emissary, Norn Assimilator, Hive Tyrant, Biovores, The Red Terror, Winged Hive Tyrant, Pyrovores, Hive Guard, Exocrine, Tyrant Guard, Venomthropes, Tyranid Prime with Lash Whip), 13 within the army (Raveners, Hyperadapted Raveners, Zoanthropes, Maleceptor, Tyrannofex, Hive Tyrant, Biovores, The Red Terror, Winged Hive Tyrant, Pyrovores, Hive Guard, Exocrine, Tyranid Prime with Lash Whip).

What moved, and why. Most round 9 → round 10 letter moves come from the cuts, not the score: a composite of 92 to 97 was S or A on round 8's fixed cuts, and it is A or B as a percentile. Biovores (92.4), Hive Tyrant (90.6) and Von Ryan's Leapers (93.9) barely move in score but drop a letter. The Anvil changes do the rest. The big bodies lose (the Norns 100 → 94, Tyrannofex 96 → 88, Tervigon 96 → 88, Hierophant 93 → 63), while ranged Warriors (79 → 86) and Von Ryan's Leapers (44 → 62) gain. Within the army, Hormagaunts go A → S and the Hierophant A → C (Auspex: S and D), which closes two of round 9's misses; the Tyrannofex goes S → B, which opens a new one. The rest of round 9's misses stay: Raveners and Hyperadapted Raveners (Auspex S, ours D and C), Zoanthropes, Maleceptor, the Red Terror, Biovores, Hive Tyrant, Pyrovores, Hive Guard, Exocrine and the Lash Whip Prime.

3. Anvil, before and after

Round 9: hits to destroy × OC ÷ points × Battle-shock × position × revive × the survival gate. Round 10: OC × min(turns alive, 5) ÷ points (per 100 pts) × Battle-shock × position × revive, no gate. Turns alive are turns5's T: each of the six attacker classes' hits to destroy against half the median unit's, held to 1 to 5, mixed by the meta's incoming fire (×0.5 for a lone character), so the measure is OC-turns bought per point and stops counting survival at the game's length. Raw numbers at each unit's best detachment and way in (detachment OC included). Titans: round 10 on the whole pool (before the pool cut); a pair that appears in both Titanicus armies is one row.

Reading it. The gaunts and Gargoyles (OC 11 to 20 for 45 to 80 pts, about two turns alive) are Anvil's top 2% again, as in round 8, but now for a stated reason: OC-turns per point. The Norns and the Tyrannofex keep a high Anvil (88 to 94) from five turns alive on OC 5. They are no longer at the top, because survival past five turns buys nothing. Titans fall from 91 to 100 down to 36 to 56: a Warlord's OC 30 for five turns is 3.9 OC-turns per 100 pts, against Termagants' 63.

4. The pool: fieldable units only

Rule: a unit over 2000 pts at its listed size can't be fielded in a 2,000-pt list, so it leaves the pool before any percentile (Legends were never in). 8 units drop (1161 → 1153):

Reading it. Six of the eight are the Titanicus Titans, which no winning list can take in a 2,000-pt game; with round 10's Anvil they already sat at the bottom (final 0.9 to 2.1), so taking them out shifts everyone's percentile by under 1 point. The panels don't move (Orks +0.006); the per-army median falls 0.011 because the Phantom and the Manta leave their armies' lists (two zero-share units that ranked low, which helped the Aeldari's and T'au's ρ). The rule stays for its own sake: units that can't be fielded shouldn't set anyone's percentile, and both Titanicus armies drop off the S table (they were 100% S in round 9).

5. The final score, S per army

The final score is the pool percentile of the composite (round 8's best allowed job + steps): the share of the 1153 fieldable units scoring below it, 0 to 100. A monotone transform, so no ρ moves (ties aside); it spreads the best-of-four pile-up. Units per 10-point bin:

Frozen cuts (calibrated on this run, 2 Oct, the lowest final score inside each share: S 8%, A 15%, B 27%, C 30%, D 20%): S ≥ 92.02, A ≥ 77.02, B ≥ 49.96, C ≥ 20.03, in V10C. They give S 92, A 172, B 312, C 346, D 231 of 1153. From now on they stay: a round that lifts most scores would show as more S. The within-army view: S 116, A 179, B 290, C 342, D 226.

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 9 on its cuts −0.39, on its proposed cuts −0.24; round 10 −0.16 (whole pool −0.15); within-army view −0.40. Negative means the armies we hand the most S to are the ones that place least.

Reading it. S per army is now a quota over all armies, so an army's share can be compared. The check improves from −0.39 to −0.16 but stays negative. Deathwatch (50%), Space Wolves (24%) and the plain Space Marines (17%) still lead, with 0 to 4 top-quarter lists. Necrons (2%), Tyranids (2%) and Custodes (6%) trail, with 15, 8 and 11. Two things drive it. The Marine chapters share most datasheets, so one strong datasheet counts once per chapter. And our score doesn't see what makes Necrons and Custodes win: reanimation and army rules worth more than any one unit. That is the cross-army claim the appraisal warned about. Nothing yet shows that the all-armies letters are comparable between armies.

6. The explainer: the eight units two or more letters off in round 9

node tools/tiertest.js --explain "<unit>" [--army <army>] [--out] prints one unit's round-10 numbers beside its three nearest peers (same best job, nearest in points, its own army within ×2 first), then the biggest difference from them on each raw quantity. On its own it runs the round-10 pipeline (about 25 s) and stops; in a full run, --explain-out <folder>. The eight outputs: reports/tiers/2026-10-02-explain-<unit>.md.

• Raveners (Auspex S; ours D, 10.8; best job Anvil). Its peers are the Haruspex, Toxicrene and Exocrine (Anvil, 120 to 135 pts). The biggest gap is durability: 45 hits to destroy against Special fire, where the peers need 168. That gives 2.6 turns alive against 4.1. Its damage against heavy vehicles is 0.39× theirs. Points to a missing raw quantity. Raveners are bought to arrive from Deep Strike and tunnels, take an objective or an action late, and screen. Our Runner reach is the same 0.80 as everyone's (the position track doesn't separate a burrowing arrival), and nothing values a unit that can't be shot until it arrives. The Auspex consensus is probably right.
• Neurogaunts (Auspex B; ours A, 86.9; best job Anvil). The peers are Barbgaunts, Venomthropes and Termagants. The biggest gap is damage: MEQ 0.28× the peers' median. But the Anvil raw is 4.8× theirs (OC 11 for 45 pts). A weight: OC per point is worth less than we count it. Two winning lists take Neurogaunts and Auspex calls them a tax for the Synapse rule, while Termagants, with the same Anvil and 4× the damage, are taken twice as often. Anvil as the sole job lifts any cheap high-OC body. An Anvil ceiling (the appraisal's experiment 6) is the matching knob.
• Hyperadapted Raveners (S; ours D, 12.4; best job Hammer). The peers are the Hive Crone, Tyrannofex and Harpy. The biggest gap is Banner: it lends 125 pts (60 of them a CP, the rest V2c's stand-ins) where the peers lend nothing, yet Banner is only 49. Its damage against TEQ is 1.7× theirs, but it has 0.73× their turns alive (90 hits against small arms, against their 216 to 576). A missing raw quantity and a weight: an all-rounder whose jobs are each middling. Best-job grading can't see it (Hammer 63, Anvil 45, Banner 49), as the appraisal's sum test showed (it rises to the 54th percentile under a sum). Also its tunnelling arrival, as for Raveners.
• Maleceptor (S; ours C, 43.8; Anvil 83). The peers are the Winged Hive Tyrant, Tervigon and Trygon. The biggest gap: it is 1.7× tougher against plasma, and it lends only 0.63× the peers' points. Its psychic Battle-shock and mortal wounds are the reasons players take it, and they aren't in its raw quantities beyond the Banner lend (72 pts). A missing raw quantity: what a Battle-shock test and mortal wounds against a whole enemy unit are worth to the army (the board-control kind we price at 10% of a median unit, swept as worth almost nothing).
• The Red Terror (A; ours D, 7.3; Hammer 55). The peers are the Screamer-Killer, Sporocyst and Hyperadapted Raveners. The biggest gap is lending: 0 pts against the peers' median of 21. Its damage against heavy vehicles is 0.72×, and it has 0.74× their turns alive. A wrong consensus or a missing quantity: our numbers say an average melee character for 130 pts. Its appeal (joining Raveners, its swallow-a-model rule, a tunnelling arrival) depends on its squad, and our Leader model doesn't join it to Raveners. Check the leader pairing first. If that is in, the consensus is generous.
• Zoanthropes (S; ours D overall, C within, 14.9; Hammer 65). The peers are Hive Guard, the Tyrannocyte and the Screamer-Killer. The biggest gap is durability: 54 hits against small arms against the peers' 108 to 360 (0.30×), since the 4++ and the psychic shield don't show against small arms. Damage is close to the peers' (0.82× on light vehicles). A missing raw quantity: Zoanthropes are taken for the 4+ invulnerable and the shield they give a Neurotyrant's unit (lent durability to another squad, which our aura stand-in doesn't model), and for mortal wounds that ignore saves into heavy targets. The consensus is probably right.
• Deathleaper (A; ours A all armies, S within, 88.6; Runner 98). The peers are the Neurolictor, Parasite of Mortrex and Biovores. The biggest gap: damage against heavy vehicles is 7.9× the peers' (59 against 3 to 11), plus a wider reach (1.0 against 0.73 to 1.0). Every raw quantity is at or above its peers'. A weight: Runner is a cheap job to top (Runner's points gate favours 80-pt units), so the best-job score lifts it one letter above Auspex within the army. The all-armies letter agrees with Auspex.
• Tyranid Prime with Lash Whip (C; ours A, 86.7; Banner 97). The peers are the Broodlord, the Winged Prime and the Neurotyrant. The biggest gap: it is 0.49× as tough against anti-tank, and it lends 203 pts against the peers' median of 157 (1.27×), nearly all of it the leader lift on Warriors (V2c's Banner), at 75 pts. A weight, or a wrong consensus: Banner per point rewards a cheap leader whose lift is large on paper, and 11 of 23 winning lists take it (more than most Auspex A units). Here the list data sides with us, not with Auspex's C.

Across the eight: three point to raw quantities we don't have yet: a hidden or tunnelling arrival (Raveners, Hyperadapted Raveners), a unit's protection lent to another squad (Zoanthropes), and the board-control worth of Battle-shock and mortal wounds (Maleceptor). Three point to weights: Anvil for cheap high-OC bodies (Neurogaunts), Runner's points gate (Deathleaper) and Banner per point for cheap leaders (the Lash Whip Prime). Two may be the consensus: the Lash Whip Prime, which the lists back, and the Red Terror, once its leader pairing is checked.