Overnight, 4–5 Oct: the morning summary

For Jordan's colleague and Jordan. Every number is from the repo's own tools on the frozen inputs; every confirmed setting is a file in tools/ledger-benchmarks/; every step is a row in reports/tiers/ledger-steps.md.

The short answer

• The model improved on every honest test overnight. Seven reviewer lists 72.3% → 75.0% pairwise; Tyranid units the model never saw 73.9% → 78.7%; the 14 never-fitted armies 57.8% → 59.1% (59.7% before step 34; inside noise).
• Confirmed, in order: 29 the mortal-wound multiplier back to 1; 30 healing, buff and detachment multipliers halved toward 1; 33 Jordan's Biovore dial applied to Biovores only (it had been lifting the Harpy); 34 Battle-shock pressure priced in points (Neurolictor 28 → 12, Neurotyrant 37 → 30); 34b eleven reaction-move rules no longer counted as command points (Von Ryan's Leapers 2 → 11). Baseline: tools/ledger-benchmarks/2026-10-05-step34b-baseline.json.
• Not kept: 31 (other multipliers halved), 32 (charge-landing factor), 28 and 28b (every mission card per unit), 35a (support as a credit at fixed strength); the fit to tournament inclusion (T-651) moved nothing.
• Provisional, not confirmed: 35b, support as a credit with its strength re-fitted. It moves every aimed Tyranid unit toward the reviewers (Winged Hive Tyrant 3 → 15, Tervigon 22 → 33, Tyrannofex 27 → 20, Exocrine 20 → 12) and lifts the seven lists to 76.1%, but that is in-sample: the 14 never-fitted armies fall 59.1% → 57.7% and held-out Tyranid units re-fitted in folds fall 76.0% → 71.4%. Saved as tools/ledger-benchmarks/2026-10-05-provisional-35b.json; unwound unless later steps confirm it.

What the reviews found (reports/tiers/2026-10-04-review.md, 2026-10-05-critique.md)

• The ledger is in effect Kill per point (~91% of a unit's Value). About 20 numbers had been tuned on data that supports 7–9; several were fitted on 3–6 units. Shrinking them is where most of tonight's gain came from.
• The reviewers largely see what players take: tournament list share orders the Tyranid reviewers 86% right. A rules-only model probably tops out in the high 70s to about 80% on Tyranids, not at the reviewers' 90%.
• The support multipliers (Synapse, command points, buff, heal) multiply the carrier's own Kill. That explains most remaining misses: the units rated too high carry big multipliers; most rated too low carry none.

Open for you and Jordan

1. Is a tier "best value per point" or "worth taking in a list"? Grading by total worth is the only change that lifts the 14 never-fitted armies (+1.5) but it drops Tyranids (−2.9).
2. The thin multipliers (Fights First, Lone Operative, Synapse, command points) were each fitted on a handful of units: Jordan's own call on each would beat any fit.
3. Space Marines lack role-scarcity and Soak-effect data (an estimated +1 to +3 points); adding it changes the frozen inputs, which needs Jordan's OK.
4. Step 35b: keep exploring support-as-credit (it fixes the Tyranid order) with a version that doesn't cost the other armies, or unwind it.

The Tyranid table now (confirmed baseline, step 34b)

Now = our rank of 52; Their rank = the reviewers' mean (five lists); Gap = ours − theirs (negative: we rate it too high); ⚠ = 10+ places; Last night = step 26. 28 of 52 in the reviewers' tier.