Step 1 closed: every panel, the baselines, two ceilings and the outcome signal (2026-10-03)

T-576 and the close of step 1 in design/tier-plan.md's queue (section 10). Command: node tools/tierbaseline.js --json reports/tiers/2026-10-03-step1.json (about a second; 2,000 bootstrap resamples, 2,000 shuffles, 200 random tie-breaks, fixed seeds). Inputs: docs/data/rankings.json, round 11's published JSON, reports/tiers/panel/. Round 11 is read, not re-run. Lists counted from grimstat-corpus (CC BY 4.0). A report only: nothing on the site changes.

In short. Five panels are now scored, on three armies. Pooled over the four panels no round was tuned on (166 units), list share alone gets 0.41 [0.28, 0.53] and round 11 gets 0.19 [0.04, 0.34]. Points alone get about 0 whichever way round they are read. Round 11 beats list share on one panel, the new Space Marine panel, and loses on the other four. On both armies that have two reviewers, the reviewers agree with each other far more than round 11 agrees with either: ceiling √ρ 0.83 for the Space Marines and 0.86 for the Tyranids. The verdict of T-570 stands: redesign the jobs (T-568, T-569) before any fit. The bar for the fitter is now list share's held-out 0.41, not 0.21.

1. The headline: pooled held-out ρ

"Held out" means no round was tuned on the panel. The headline pools Tista Minis' and Auspex's Space Marines, Sprues & Brews' Orks and Hivemind Hobbies' Tyranids, weighting each panel by its unit count; the interval resamples units within each panel. "All panels" adds Auspex's Tyranids, which steered rounds 5 to 11.

Two caveats on the label. First, Hivemind's Tyranids grade the army we tuned on, through another reviewer. So round 11's 0.35 there is less of a surprise than its 0.41 on Auspex's Marines. Second, the fitting panels are held out only until the fitter runs. After that, the only truly unseen panel is the locked Astra Militarum list, and leave one army out is the honest score.

2. Every panel, with lift over its shuffle

Each cell gives ρ [95%] with its lift over that panel's shuffle in brackets, then within one on the reviewer's own letter shape, with its lift. The lift is the number minus the mean of 2,000 random orders of the same units.

• Hivemind's four patch drops. Pyrovores, Neurotyrant, Tyrannocyte and Toxicrene are left out because their points moved by more than 10% between 2 Jul and today. The battery now honours a panel's patchDrop (T-576).
• Auspex's Space Marine shape. The battery scores letters. The reviewer's seven-step grade (ord) is stored for the fitter's own cut-points.
• List share's 0.81 on both Tyranid panels says these two reviewers grade close to what wins lists take. Plan section 2 treats that as a reference point for the army, not a cap.
• Points' sign still flips by panel: cheaper is better for both Marine panels, dearer is better for the Orks and both Tyranid panels. Pooled with one fixed sign, points are worth nothing.

3. Two ceilings

The ceiling is ρ between two reviewers on the units both grade, letters against letters, with ties averaged. √ρ is the most an engine can expect against either of them.

Round 11 reaches 0.14 and 0.41 against the two Marine reviewers, where the ceiling is 0.83. On the Tyranids it reaches 0.35 against the untuned reviewer, where the ceiling is 0.86. List share alone, at 0.81, sits near the Tyranid ceiling, which suggests both Tyranid reviewers lean on what the lists take.

4. Round 11 against list share, per army

The median per-army ρ with list share is 0.173 over 27 armies. Five armies are negative and seven are within ±0.1 (unchanged; no new data).

5. The outcome signal (T-571, reports/tiers/2026-10-03-outcomes.md)

The outcome data covers 382 placed lists from 23 events across 33 armies, July to September 2026. It holds final placings only, with no game records. A unit's lift is the mean finishing percentile of its army's lists that take it, minus the mean of those that don't. Of 418 units with 3 or more lists on each side, 33 have a 95% interval clear of zero, where chance alone gives about 21. That is a real signal, but too weak to grade one unit. Its use is as an aggregate yardstick: the correlation of our letters with lift, weighted by the smaller side. No faction-versus-faction table may be read by a machine today. It is not yet compared with the panels or round 11 (next step).

6. Panel inventory

One locked panel, not two. The plan asked for two locked panels. 11th edition came out in June 2026, which makes both Votann lists 10th-edition, and Jordan has no more lists. So the second lock is waived (the supervisor's call; Jordan: "use what you can"). The final check rests on Astra Militarum alone, 61 units, which is wide enough for one ρ with an interval of about ±0.25.

Patch alignment. Points were read from BSData's own files at the reviewers' commits at run time. Nothing of theirs is kept in the repo, and the newest read matches rankings.json exactly. Hivemind: 2 Jul to 30 Sep checked, four units dropped. Auspex's Space Marines: 2 Sep to 30 Sep checked, no move, so T-575's gap is closed.

Addendum, 3 Oct: the panels re-sorted after the design review (T-580)

Command: node tools/tierbaseline.js --sanity --json reports/tiers/2026-10-03-battery-t580.json (same seeds, same data: BSData 374f505, corpus 570b8c6, before the 1 Oct dataslate). Jordan's answers to the design review, passed on by the supervisor, change four things:

• The Orks are retired (reports/tiers/panel/retired/): a codex review written before points, not a tier list. The battery reads the file only with --sanity, as a sanity list: lift over shuffle, outside every pool, ceiling and gate. tools/tiertest.js's hand list no longer reads it (the old rounds' Orks columns print "–").
• A third Tyranid reviewer. reports/tiers/panel/2026-10-03-tyranids-second.json is a second creator's list (Jordan: two different creators, both current, used as is; which of the two is Hivemind Hobbies is not recalled and not chased). Both new Tyranid files are fitting panels. The second file was dated to about early July, so the same three graded units the other file drops for points moved more than 10% since then (Neurotyrant, Tyrannocyte, Toxicrene) are dropped here too (patchDrop; the builder's call, flagged).
• "Held out" narrowed to panels of armies no round was tuned on: the two Space Marine panels. The new Tyranid lists are "same army, new reviewer"; together the two sets are "unseen".
• The gate. Lift over shuffle on every fitting panel, none worse than its frozen predecessor; operationally, the pooled lift over the shuffle on the unseen panels rises above round 11's (with a paired bootstrap interval) and no fitting panel falls more than one SE below round 11. "38 of 44 within one of Auspex" is a bench sanity number. And "list share" in every number here is the share of entrant lists (any placing), not of top-quarter lists.

Pooled

Per panel (ρ [95%] (lift over the shuffle))

Ceilings

The Tyranid ceiling, three reviewers: √ of the mean pairwise ρ, √0.65 ≈ 0.80. Before the patch drops the two new lists agree at 0.57 on 43 shared units (16 the same letter); the three dropped units take that to 0.54 on 40. The two new lists agree with each other less than either agrees with Auspex, and their tier names, shapes (4/12/19/11/3 against 6/11/16/9/2) and graded units differ: nothing in the numbers suggests one video summarised twice, which fits Jordan's ruling.

What it says. On the armies we never tuned on, round 11 (0.30) edges entrants' list share (0.23), inside the margin. On the Tyranids, list share is far ahead (0.72 against 0.38) and both new reviewers lean on it. Pooled over the unseen panels list share leads, 0.49 against 0.34. The bar for the fitter is list share's held-out ρ: 0.23 on the held-out set as now defined, 0.49 over everything unseen.