# Route14 rescue: a map-free, locally admissible tail discriminator The observed-position experiment remains blocked exactly as in return438. This is a **weaker changed experiment**, not recovery of its missing maps or an arithmetic probability model. After masking every individually inadmissible high adjacent pair, a discrepancy against the reflection-half permutation benchmark remains: exact second moments give model-tail upper bounds **0.044464 at x19 and 0.008439 at x23**. No census, old null simulation, or new random draw was executed. ## Question and changed ingredient Can individually impossible small-prime patterns account for all above-observed adjacent sums in the relaxed benchmark? For consecutive gaps `(u,v)`, the three twin starts require the six offsets `H(u,v)={0,2,u,u+2,u+v,u+v+2}` to omit a residue class modulo every prime. For positive multiples of6, primes2 and3 impose no further obstruction, and primes at least7 cannot be covered by six offsets. Thus this *tuple-local* condition is exactly `|H mod5|<5`. The forbidden ordered gap-residue pairs are `(1,1),(1,3),(2,4),(3,1),(4,2),(4,4)`. This standard admissibility ingredient is not a new number-theoretic principle. Its use in the following retained-histogram calculation is the changed discriminator. Keep the original uniform half-multiset law `g=(b,6,reverse(b))`, but **change the score**, not the law. Let `a` be the published observed A2, and count `Y = sum_{i=0}^{n-2} 1[b_i+b_(i+1)>a and H(b_i,b_(i+1)) admissible]`. The observed `Y=0` follows from its reported A2 alone; observed positions are unnecessary. Equivalently, mask inadmissible edges when scoring the full cyclic maximum, calling the result `A2_mask`. On a true wheel `A2_mask=A2`, whereas in benchmark words invalid edges contribute nothing. The event `A2_mask<=a` implies `Y=0`; omission of the two marked seams makes our bound conservative. **Crucial distinction:** invalid words remain in the benchmark. This is not the distribution conditional on every pair, or every cumulative slot residue, being admissible. ## Exact result, conditional on published custody |x|Half length n|Reported a|E(Y)|Var(Y)|Upper bound P_model(Y=0)| |---|---:|---:|---:|---:|---:| |19|189337|186|20.1363917248|18.8678661487|0.044463805185| |23|3976087|234|114.5047339256|111.5880271776|0.008438983923| These are exact rational evaluations displayed in decimals, not estimates or arithmetic p-values. The corresponding expected counts of *inadmissible* above-a interior edges are 6.2752235432 and 10.9473761515. Masking those contributions does not exhaust the benchmark's above-a tail. No z-score, monotone conditioning profile, mechanism, level independence, extrapolation, A1 bound, or infinitude statement follows. The inputs were copied, not regenerated, from `research/attack-foldL-01-census.js`, SHA-256 `8a769109f8ceda9691a5665d06c4d17463dc8db9c74e756600bacaa61ff4fe87`, lines763–766 and927/934. Reported full counts are378675/7952175, periods9699690/223092870, A1=150/204 and A2=186/234. The x23 counts also match the byte-pinned return435 input `c132b778...f82cf`. Count sums, weighted periods, maxima and odd6/even-other multiplicities were checked. This does not independently validate the historical wheel census. ## Inspectable derivation Write `h_v` for half multiplicities and `K(u,v)` for the displayed admissible-exceedance predicate. For an ordered length-k value tuple `w`, its number of distinct occurrence-label realizations is `W(w)=product_v (h_v)_(multiplicity of v in w)`, where `(h)_r` is a falling factorial. Define `C2=sum_{u,v} W(u,v) K(u,v)`, `C3=sum_{u,v,w} W(u,v,w) K(u,v)K(v,w)`, `C4=sum_{u,v,w,z} W(u,v,w,z) K(u,v)K(w,z)`. Uniform multiset permutations give probabilities `pj=Cj/(n)_j`. There are n−1 edges, 2(n−2) ordered overlapping edge pairs, and (n−2)(n−3) ordered disjoint edge pairs. Consequently `mu=E(Y)=(n−1)p2`, `E(Y²)=mu+2(n−2)p3+(n−2)(n−3)p4`. Cauchy–Schwarz applied to `Y 1[Y>0]` gives `P(Y=0) <= 1−mu²/E(Y²) = Var(Y)/(Var(Y)+mu²)`. The checker uses a different computation: the symmetric occurrence-label graph has degree `d_v=sum_w(h_w−[v=w])K(v,w)` at every occurrence of v. Symmetry follows because `H(v,u)=(u+v+2)−H(u,v)`. It obtains `C2=sum_v h_v d_v`, `C3=sum_v h_v d_v(d_v−1)`, and `C4=C2²−4C3−2C2`. The last identity partitions two directed edges by zero, one, or two shared vertices. No producer import or sampled word is shared. The formula and bound are **PROVEN by this elementary argument**, subject to review. Their finite implementation agreement is **VERIFIED** on the stated copied inputs, conditional on their custody. ## Boundary witness and preserved obstruction Pair admissibility is weaker than a common arithmetic phase. Take `b=(12,12,12,6,30)`, `g=(b,6,reverse(b))`, period150. Every cyclic adjacent pair is locally admissible. Nevertheless the first four starts have offsets0,12,24,36; these and their +2 partners cover all five residues. No single starting phase modulo5 makes the whole word a twin-slot word. This is a reflected combinatorial witness with period divisible by30, not a primorial realization. This witness prevents upgrading the masked calculation into a globally conditioned null. Return438's histogram+A1/A2 nonidentification, strict `TJ+1; section2.2 “instances” versus driving terms, https://arxiv.org/html/2608.26384v2 . * Fred B. Holt and Helgi Rudd, *Estimating constellations among primes I – Uniformity*, arXiv1312.2165, introduction only: uniformity is a modeling assumption and finite estimates have systematic error, https://arxiv.org/html/1312.2165 . The separate prior-art record states searches, actual reading coverage and access gaps. Public artifacts contain only task evidence and original analysis; native transcript export/scrubbing and final submission are handled by the coordinating session. Coordinating review: the checker was strengthened to consume the target's half-length and ordered-value-pair metadata too; its original stdout remained unchanged, and an altered half-length target was rejected. Transcript redactions remove credentials, session/attempt/launch identifiers, private paths and metadata, unrelated assignment/accounting activity, hidden reasoning/encrypted fields, and bulk external-source payloads.