# Weighted partial convexification on one prime pair **Design only.** No arithmetic incidence construction, phase enumeration, random draw, LP solve or published computation was run in job952. Generic block convexification is known; this is a scoped arithmetic extension, not a new optimization theorem. All dependencies retain their existing review status. ## Object and certificates Use the finite support D and phase sets K(q,b) of route4. A block partition P has disjoint prime blocks whose union is Q. For block B and its complete phase tuple beta, set U(B,beta)=union over q in B of K(q,beta_q). Every prime has exactly one phase. With nonnegative weights w on slots and sum_D w=1, define M_B(w) = max_beta sum_{s in U(B,beta)} w_s C_P(w) = sum_{B in P} M_B(w) margin_P(w) = 1 - C_P(w). A strict positive margin proves noncoverability on this D: every actual full phase vector has weighted union at most C_P(w), whereas covering all D has weight1. A supplied rational cap z_B must dominate EVERY block column, not a sampled or greedy maximum, and sum_B z_B<1 must hold exactly. That check is the entire positive certificate. The normalized primal/dual pair is tau_P = max tau mu[B,beta]>=0; sum_beta mu[B,beta]=1 for every B; sum_{B,beta: s in U(B,beta)} mu[B,beta] >= tau for every s. tau_P = min sum_B z_B w_s>=0; sum_s w_s=1; z_B >= sum_{s in U(B,beta)} w_s for every B,beta. These are finite LP duals; variables are rationally checkable. A primal rational block distribution with every slot coverage>=1 certifies that this particular partition has no strict nonnegative-weight certificate. It does not exhibit an integer covering phase vector. Lower/upper exact bounds l_P<=tau_P<=u_P come from the two feasible certificates. u_P<1 is success; l_P>=1 is failure of this partition/certificate class. Otherwise unresolved, even if floating output appears decisive. ## Relation to the singleton class The singleton partition is exactly #362's normalized certificate. M_{q,r}(w)<=M_q(w)+M_r(w), hence merging blocks can only increase the certificate margin for the same w. Equivalently, projecting a block distribution to single-prime marginals yields singleton expected hit counts at least its union probabilities. Therefore tau_P<=tau_single. Strict improvement is possible, but not guaranteed. The #362 four-slot example distinguishes the classes without arithmetic claims: group A has columns{1,2},{3,4}; group B has{1,3},{2,4}. Singleton distributions half/half give coverage1 on all four slots, and uniform slot weights give cap sum1, so tau_single=1. Merging the two groups produces four union columns, each missing a different single slot. Uniform slot weights give cap3/4; uniform probability on these four columns gives each slot coverage3/4. Thus tau_pair=3/4 and margin1/4, although singleton certificates are all silent. This is an arbitrary incidence example, not a prime interval or a newly executed enumeration. Unweighted pair-block grouping is already research/history/staging/attack-ab-coupling.md section2. Its section8 also names a fractional cover of PRIME SUBSETS with costs maxcov(B). That optimization weights blocks. Here the new finite quantity optimizes SLOT weights for one fixed joint phase domain. The two coefficient spaces should not be conflated. The project's unweighted hierarchy's rising-depth/cost obstruction, and the generic information barriers, remain intact. ## Frozen experiment and prerequisite D is the first66 slots of #357 input SHA ea81d82b582acc99a8d96411eb1dc28589421cb332ac5609915049cc8bffa10d. Reuse #363's exact half-open interval[9409,13722), old T97, Q the19 primes101..193. The parsed slots and F1<0 must be confirmed in a selected bounded input/witness check; no new census or moving the endpoint is authorized by this proposal. Freeze partition P={{101,103}} plus singleton blocks for the other17 primes. Do not choose a pair after seeing outcomes. **Eligibility first:** read/reuse an existing route4 rational singleton fractional-cover witness on this identical N66 support, with slot coverage>=1. Ask2 already requests the existing result. Do not duplicate route4's singleton search for this experiment. If the source is unpublished, triage is blocked on it; if a strict singleton positive margin is published, this particular escape test is ineligible and stops. A baseline fractional cover supplies a rigorous class obstruction, not an accepted arithmetic result unless independently reviewed. Once eligible and compute is separately offered, construct all pair union columns on the frozen input and optimize only the fixed pair-block LP. There are exactly101*103=10403 raw pair tuples. At most17*193=3281 remaining singleton columns, at most13684 total raw columns, 66 normalized weight variables and18 caps. These are design size bounds, not measured run costs. Duplicate masks may be merged with their original phase representatives retained. Preserve every raw phase membership for the certificate checker; no argmax-only restriction. A future run should first check the symbolic four-slot control, then solve the arithmetic dual with one solver thread and extract rational weights/caps. Exhaustive integer/rational verification of all columns is mandatory. If needed, obtain a rational primal distribution for an exact negative certificate. Rational reconstruction or solver status alone is not enough. Publish original solver output, reconstructed witness and the exact checks, including any failed reconstruction. A found positive witness is checked against literal modular K(q,b) by a checker that shares no incidence-building code with the producer. No full optimality proof is needed for u_P<1. Stop after one fixed partition and one fixed support. u_P<1 together with baseline coverage>=1 is the finite separating result. l_P>=1 refutes only this partition's strict weighted escape. An actual integer cover also refutes noncoverability and invalidates a purported positive certificate. Budget expiry, numerical ambiguity or missing source gives unresolved/blocked, not a negative mathematical answer. Do not sweep all pairs, triples, additional starts or prefix lengths as a rescue within this proposal. ## Cost and theorem gap Future pilot:0.5 agent h,0.1 CPU h total including incidence construction, solver, witness reconstruction and exact check; one thread, RAM1GB, disk0.1GB. Stage the first solve at60 CPU seconds, included in the360 CPU-second total. Abort cleanly on memory/time limit and report the partial observation. No benchmark justifies expecting success inside that estimate. Independent review, if a finite positive witness appears, needs its own exact scope/execution and judgment budgets. To reach the exponent/infinitude consumer, one would still need uniformly, over all relevant growing arithmetic supports, a positive block-weight margin at a target polynomial interval scale, together with controlled block size and construction/proof cost. A finite N66 certificate cannot establish that. The cheapest experiment attacks only the weakest first premise: whether one joint two-prime ingredient ever escapes a certified singleton obstruction on the frozen arithmetic support. No fixed block-size uniform theorem, eventual positivity, H_alpha or twin-prime infinitude is assumed.