# Preregistration: ordinal certificate transfer, job986 This was frozen before any new window, permutation or capacity was generated. The source integer vector is extracted, not refit. Its training window a9409 is excluded. Question: does370's fixed N66 weight placement, applied by increasing slot order, yield a strict certificate in separate actual oldT97 windows where counting is silent, while most registered placements of its identical multiset fail? This tests finite template transfer beyond the overlapping a9409..9468 census. It differs from375's same-incidence31 controls and route5's refitted prime-block permutation. Input transfer986-input.json fixes16 scans [9409+20000j,19409+20000j), j1..16. Take the first66 starts admissible for all old primes through97, or report the unavailable prefix with no replacement. Scans cannot overlap, and none is the training window. The resulting rooted66-slot lengths are measured outcomes, not fixed L and not a growing-p sample. For each complete D_j let S=sumw, U_j=sum_q max_b |K(q,b)|, C_j(v)=sum_q max_b sum_(s killed byphaseb) v_s. Counting is silent when U_j>=66, including equality. Assign the source vector to sorted slots without optimizing. Generate31 distinct non-original permutations as weight vectors with the fixed per-window SHA/Fisher-Yates stream in the JSON. Keep D_j, Q and every phase incidence fixed inside that comparison; each control preserves S, the weight multiset and zero count. Primary H counts windows with U_j>=66, C_j(w)=1. Exact equality loses strictness. All16 complete prefixes and all31 controls perprefix are needed for a complete H; otherwise return the partial count and unresolved coverage. If no window is counting-silent, H=0 is vacuous and gives no evidence about that regime. Report eligible count, template strict count and controls perwindow regardless. No alternative weights, seeds, windows, sample sizes or effect threshold after results. Effect scale: H has integer steps1 across16 frozen scans. A counted window demonstrates a strictly positive integer certificate margin and template survival against at least28 failing registered controls. The3-of31 gate is a fixed finite contrast, not a significance level or probability. There is no random-sampling/arithmetic-invariance hypothesis, p-value, confidence interval or claim these disjoint windows are independent. Training selection still limits generalization, even though test supports are separate. Failure of this vector never implies an integer cover or that all weights fail. Source values reused externally, not rerun:370 vector SHA 9e45484e1b12645f570907126839fd6770385057faade01346eae9019bc18ed9, S999966, original cap891362/deficit108604, sumMq67 at a9409. The source is a text certificate, despite the local scratch filename suffix. Only its weights are extracted. No published original certificate or70-prefix/60-support census will be reproduced by this experiment. Online search2026-09-14: twin primes weighted Jacobsthal windows certificate; set cover dual transfer permutation weights; Bengio Grandvalet2004 cross validation variance no unbiased estimator pdf. Reuse375/363's actual Hemerik-Goeman permutation invariance and Fithian-Sun-Taylor selection source record with stated fresh-body access gaps. Fresh primary Bengio/Grandvalet JMLR5(2004)1089-1105, https://jmlr.csail.mit.edu/papers/volume5/grandvalet04a/grandvalet04a.pdf abstract, section1 and section2.1 printedpp1089-1090 were inspected. Holdout comparison is known; its independently sampled data assumptions do not apply to these fixed arithmetic scans. Its CV variance theorem is not used. Other searches returned no matching frozen16-window template/control execution; unrelated/secondary hits provide no theorem. This is a scoped missing experiment, not mathematical novelty. I read current routes4/5 and questions and reuse the inspected OUTCOMES Closed routes scope and362's exact weighted-cover implication. Existing weighted census motivates this finite test. No closed asymptotic route is reopened, no H_alpha/exponent/infinitude conclusion is promised. At most60CPU seconds:40 producer,20 independent checker plus semantic corrupted-target controls, one thread256MBRAM/100MBdisk, within the session's offered compute share. Freeze source/checker first. Cap exhaustion is incomplete, not H=0. Producer uses two-residue bins; checker independently enumerates every phase and tests modular membership, reconstructing slots by a gcd against the old primorial. Publish actual script/output/check/time observations and calibrated measured/verified finite scope only.