Runs 076–077 found that, at fixed fixed-point separation δ, the one-block composition deviation |D| depends on cut class C₁ (partial +0.60 over all pairs; +0.72 over unequal-class pairs with the amplitude (ln N₁)/C₁ also held) and not on the amplitude. PN-004's continuation House set two readings against each other. Order: C₁ is the order of vanishing of the crossing map at the wall, and a mismatch in it obstructs composition directly. Curvature: equal-C₁ patterns have similar curvature s = R″/(λ(λ − 1)) at matched p*, so the first-order term, proportional to s_A − s_B, cancels by arithmetic and C₁ only proxies the curvature mismatch. Which is it?

Sealed 2026-09-30, after PN-004's Continuation 1 House and before any statistic of |s_A − s_B| by class, or any partial controlling |Δy₉| or |s_A − s_B| for |ΔC₁|, was computed (codifier). Known at sealing: RESULTS076 and RESULTS077 in full; the per-pattern s table (so the class offset in s is visible: where cut 2 and cut 4 fixed points overlap, near p* 0.80–0.83, cut-2 s is about −5 and cut-4 s about −3.5 to −4); the 077 reported partial of |D| with |ΔC₁| given |δ| and |Δy₉|-as-amplitude (+0.367, over unequal pairs only).

The panel's forecasts this run scores (seat record, continuation). The House's F1: the partial of |D| with |ΔC₁| controlling |δ| and |Δy₉| (not s) lies in 0.30 to 0.60, sign positive with p = 0.68. Its F4: for pairs with |δ| < 0.01 and |ΔC₁| = 2, |D| is 0.0003 to 0.0020 and at least 3× the equal-C₁ baseline in that bin. Both are computed and reported here; they are scored in the seat record, not as clauses.

Why the codifier leans to curvature. s tracks p* almost monotonically (Kendall −0.91) but with a class offset, so at matched separation an unequal-class pair carries a curvature mismatch an equal-class pair does not. The first-order term is proportional to s_A − s_B, and although it overestimates D beyond δ ≈ 0.01, the higher-order terms are built from the same curvatures. The order reading has no derived term yet (the House's own ASSERTED 1). What B can show: controlling |s_A − s_B| removes the path C₁ → s → D, so B's partial is the direct effect of cut class. B below 0.15 shows that curvature at p* is a sufficient statistic for the class effect; it does not show that order is irrelevant, since an order-of-vanishing difference may itself set the s offset. Telling those apart needs a derived order term, which this run does not have.

Reported, not scored. Per |δ| bin, the median |s_A − s_B| of equal against unequal pairs (known at sealing from the s table to favour equal pairs, so not scored; the top bin has only 10 equal pairs). The partial with |δ| and |Δy₉| only (the House's F1); with |δ| and |s_A − s_B| only; Spearman of |s_A − s_B| with |ΔC₁| given |δ|; the ten equal-C₁ pairs of the top |δ| bin listed with δ, D, s_A, s_B (the House's test 3); every pair with |δ| < 0.01 and |ΔC₁| = 2 with |D|, against the equal-C₁ median in that bin (the House's F4 and test 4); bootstrap intervals for A's bin medians and for B.

The reductio this seal passed first. By construction, caught: a first draft scored whether equal pairs have smaller |s_A − s_B| at matched δ; the class offset in the s table nearly guaranteed it, so it is reported, not scored. The new A matches on |s_A − s_B| and tests |D|, which was unseen. B's partial is new: the 077 partial with |Δy₉| was over unequal pairs and did not control curvature. Where |s_A − s_B| nearly determines |ΔC₁|, a low B would hold by construction, hence gate G2. Collinearity: gates G1 and G2. Negation: the order world shows equal and unequal pairs with similar curvature mismatch (A fails) and a full partial that stays high (B fails). Cost: seconds, stored data.

Reviewed before commit by a second reader (OPS.md); its edits are folded in: clause A rebuilt to match on curvature mismatch and test |D| (the first draft's A was nearly guaranteed by the s table); gates for collinearity and for |s_A − s_B| determining |ΔC₁|; B's signed band; what B can and cannot show; the 077 figures named with their populations.

Scored 2026-09-30 — unanticipated

Scored by the codifier from RESULTS078.md (7290c06), as written. The lab used run 076's exact |δ| cuts, as the statistic line requires ("exactly as run 076"); the rounded cuts printed in the seal give the same A.

Gates — B scorable. G1 = +0.857 (limit 0.95); G2 = 0.735 (limit 0.80).

(A) Curvature accounts for the class gap — confirmed. Median r over the 11 eligible cells is 1.355, inside [0.5, 2.0] (function bootstrap 1.00 to 2.25). Matched on curvature mismatch as well as separation, the 3–10× equal/unequal gap of run 076 shrinks to about 1.4×. Cells range from 0.86 to 11.6, so the matching is not uniform.

(B) Curvature carries the class effect — unanticipated. The partial of |D| with |ΔC₁| given |δ|, |Δy₉| and |s_A − s_B| is +0.185 (bootstrap +0.10 to +0.27): below the falsifier's 0.30, above the prediction's 0.15. From 076's +0.60, controlling the exponent gap and the curvature mismatch removes about two thirds; a small direct cut-class term remains, its interval excluding zero.

Outcome. Unanticipated. Curvature mismatch at p* is the main channel by which cut class reaches the one-block composition (A), and the class effect is mostly, not wholly, carried by it (B). The order reading keeps a residue; the curvature reading carries most of the weight. Note that curvature is itself tied to class (the partial of |s_A − s_B| with |ΔC₁| given |δ| is +0.57), so "curvature carries it" does not say order is irrelevant: the order of contact may be what sets the curvature offset.

The panel's forecasts (PN-004, Continuation 1 House). F1, the partial given |δ| and |Δy₉| only, in 0.30 to 0.60 with positive sign: measured +0.195, sign hit, interval miss. F4, pairs with |δ| < 0.01 and |ΔC₁| = 2 at least 3× the equal-C₁ baseline with |D| in 0.0003 to 0.0020: 11 of 12 pairs at 3× or more (2.3× to 33.9×), 9 of 12 inside the interval, median |D| 6.6e-4; hit.