Written by the supervisor at Jordan's request ("pause after runs and actually analyse"), from the step-1, round-12 and round-13 reports, the matrix, field and allocation reports, and the unit-by-unit tables. No new run. Job names are the site's: Tank (Anvil), DPS (Hammer), Support (Banner), Action (Runner).
1. The numbers that matter
| Space Marines, 2 reviewers (never tuned on) | Tyranids, 2 new reviewers | Tyranids, Auspex (tuned on) | |
|---|---|---|---|
| Reviewers agree with each other (ceiling) | 0.69 → about 0.83 | about 0.80 for all three | |
| List share alone (how often tournament lists take it, any placing) | 0.23 | 0.72 | 0.81 |
| Our engine (B1, the current baseline) | 0.29 | 0.37 | 0.70 |
Read across the rows:
- We are at about 0.3 to 0.4 on reviewers we never tuned on; the reviewers reach about 0.8 with each other. The 0.70 on Auspex's Tyranids is mostly what eleven rounds of tuning on that one list bought; it does not carry to other reviewers of the same army.
- Reviewers largely rank by what wins. On Tyranids, list share alone gets 0.72 to 0.81, twice our engine. On Space Marines list share is weak (0.23), because Marine lists are split across chapters and few, and there our engine edges it.
- Our yardsticks cannot see small changes. Each panel's margin of error is ±0.14 to ±0.19; pooled over the four unseen panels (162 units) the interval on a change is about ±0.10. Every compose-layer change tried in rounds 12 and 13 moved agreement by 0.05 or less. Those changes are not good or bad; they are invisible. Tuning knobs against these panels is over as a method: only changes that move agreement by 0.1 or more can be judged, so only big changes are worth making. The way to make the yardstick see 0.05 is more graded units: a panel for each of three more armies would roughly halve the interval.
2. The misses are not random: one signature, both armies
Lining our letters up against every reviewer, unit by unit (Tyranid table, round 13; the same check on the Space Marines):
We rate too high: cheap bodies and cheap utility. Tyranids: Ripper Swarms (ours S, reviewers B/C/B), Pyrovores (S vs A), Neurogaunts (A vs B/C/C), Termagants and Gargoyles (A vs B), Carnifexes (B vs C/C/C). Space Marines: the Impulsor (S vs A/B), Outriders (S vs B/C), the Librarian and Captain in Phobos Armour (S and A vs B), the Tactical Squad (A vs B). Mean cost about 65 pts.
We rate too low: the expensive damage-dealers, the big monsters and the leaders that make an army work. Tyranids: Hive Guard (ours D, reviewers A/B/B), Exocrine (C vs A/A/A), Zoanthropes (C vs S/A/B), Maleceptor (C vs S/A), Genestealers and the Broodlord (C vs A/A), the Swarmlord and the Hive Tyrant (B vs S/S/A and S/A/S), both Norns (B vs S/B/S), the Tyrannofex (B vs S/A/A). Space Marines: Infernus Squad (D vs S), Bladeguard Veterans (C vs S/A), the Repulsor, Land Speeder and Terminator Assault Squad (C vs S), Hellblasters and Inceptors (C/D vs A), the Apothecary Biologis (D vs A). Mean cost about 140 pts.
Three causes, and the job numbers separate them (Tyranid job percentiles: DPS / Tank / Action / Support):
- Price. The final score leans on cost (−0.41 with points on frozen round 11; −0.30 on B1). Taking it to zero (round 13's trial, α 0.6) lifts the leaders (the Swarmlord, the Hive Tyrant and Raveners go up a letter) but leaves the damage-dealers where they were: Exocrine C, Zoanthropes C, Maleceptor C, Hive Guard D. So price is part of it, not the whole.
- The composite swallows one strong job. The final score is the pool percentile of each unit's best job. Cheap units reach 95 to 100 in Tank or Support per point (Termagants Tank 99, Gargoyles Tank 98, Ripper Swarms Support 82 to 100, Neurogaunts Tank 98), so a unit whose best job is 85 to 94 lands mid-table. Held down here, on B1: the Tyrannofex (DPS 88, final B), both Norns (Tank 94, final B), the Swarmlord (Support 96, final B). Replacing the five targets cannot lift these; only a different composite can. This is the design review's critique 2, and T-585 tests it.
- Damage measured against the wrong targets. DPS (Hammer) scores each gun into five invented targets with hand-set weights, so a unit built to delete one kind of target is averaged over targets it would never shoot. Held down here, on B1: the Exocrine (DPS 65), the Maleceptor (55), Hive Guard (65), Zoanthropes (65), and on the Marine side probably Infernus Squad and Hellblasters. This is what the matrix and the field replace.
One oddity to settle along the way: at minimum size (B1), Hive Guard's DPS falls from 86 to 65 and Zoanthropes' from 85 to 65, while the Tyrannofex's rises. Damage per point should not depend much on squad size, so something size-dependent (a lend, a per-squad buff, or how a unit's best squad is picked) sits inside DPS. It needs one explainer run before the field's numbers are read.
A fourth, smaller cause shows on Space Marines: lone characters. We score a Phobos Librarian alone (and over-rate it) and score a support character like the Apothecary Biologis by a lend that barely registers (and under-rate it). Reviewers grade characters by what they do for the squad they join.
3. What is solid, and what is not
Solid (tested, reproducible): the comparison battery and the lock; the matrix (every real gun against every defender, agreeing with today's Offense numbers at 0.997 or better on all five targets, rebuilt incrementally); gun targeting (it picks the game-plan sheet's targets 54 of 54 times; anti-tank fire stays off chaff); the field (447 lists, 99.7% matched); compose reproducing round 11 exactly, so any change is now one measured change.
Not solid: the four jobs as a frame (untested: T-585); the price treatment (half the lean explained); survival in the reference battle (95% of the field dies in five rounds until the exposure share is calibrated to your "two-thirds or more"); 4,301 conditional abilities still counted as always on; 34 gun keywords not fully read (T-582); stratagems (T-558); the meta (one state: the lists are a continuum). And two limits on every yardstick: every panel but Tista Minis is a summary of a video (transcript unread, dates approximate), and the corpus is small (112 top-quarter lists; 24 of 31 armies have fewer than five).
4. Options
A. The field, as one measured change, with predictions written first. Finish the field (T-569 step 4: the calibrated battle, each unit's DPS and survival against the real field) and measure it on B1, with the trial's price fix beside it. Predictions go on the job number first and the letter second, since the composite can swallow a letter: before the run, write each named unit's raw quantity and its expected direction (for example, "the Exocrine's DPS percentile against the field rises from 65", "Ripper Swarms' Support falls"). Cause 3's units should rise in DPS; cause 2's units should not move much (only a composite change moves them). About two agent-rounds.
B. The frame: T-585. Factor analysis and a free fit on every raw variable, beside the four-job composite, on B1's table. It needs only the table that exists, about one agent-round, and answers whether the composite is what holds cause 2's units down. The free fit uses today's DPS numbers and reruns on the matrix's later; the factor structure doesn't depend on that.
C. Accept that tier lists measure popularity. Use list share for the letter and our maths for the reasons. That is the cheapest way to agree with reviewers, and it breaks your rule that the letter is our maths, not the meta. Not recommended; noted because the data says reviewers mostly do this.
Recommendation: A and B together, in parallel on B1, read as two separate measured changes: B answers cause 2, A answers cause 3. First, one explainer run on the size oddity. From now on, judge every change by named unit predictions on job numbers plus the pooled number, never by small moves in ρ. And if more tier lists turn up for any army, they are the cheapest way to sharpen every yardstick.