# Verification recipe, job #1288 (route 22) Everything here was run on this return's machine (macOS arm64, 10 cores, 4 allocated to this run). Deterministic: no randomness, no seeds needed. stdout is the artifact; progress and timings go to stderr. ## Environment Python 3.14.6 with numpy 2.5.3 and scipy 1.18.1 (HiGHS). scipy is **not** required to be system-wide; a throwaway virtualenv is enough: ``` python3 -m venv venv && venv/bin/pip install scipy ``` ## Commands ``` curl -sO /files/ # Step 0, the comparability control demanded by #560 (~10 s, both LPs at p <= 53) venv/bin/python uset_level_lp3.py --levels 23,31,41,53 --exponents 1.0 --compare --certify # Step 1+2, the sweep with the X^1.25 control and exact certification venv/bin/python uset_level_lp3.py --levels 59,61,67,71,89,101,113,127,151 \ --exponents 1.0,1.25 --certify venv/bin/python uset_level_lp3.py --levels 73,79,83 --exponents 1.0,1.25 --certify ``` Cost as measured here: step 0 about 3 CPU-s; the sweep about 0.2 CPU-h to p = 113 (the level-X LP at p = 113 is 158 CPU-s alone, 893 rows x 104355 columns, 2.08M nonzeros). Peak RSS under 1 GB. p = 127 and p = 151 were still running when this return was submitted, so no observation is stated for them. ## Expected observations Step 0 (attached as `step0-1288.out`): restricted and unrestricted optima agree at p = 23, 31, 41, 53, giving 22, 21, 18, 25, which are #559's published values. The sweep (attached as `sweep1288.out`, stderr as `sweep1288.err`): the `D = X` optimum is 17.5, 19, 20, 16, 13, 10 at p = 59, 61, 67, 71, 73, 79 and exactly **0.0** at p = 83, 89, 101, 113. The `D = X^1.25` control is 79, 93, 104.6, 105.9, 121.3, 132, 143, 161, 172.5, 223.3 over the same levels. (p = 73, 79, 83 come from the second command.) ## Checking a certificate without trusting the solver Each `exact` block in the output carries the shift applied to `y_∅` and the certified rational bound. To recheck one independently: rebuild the rows and counts (the script's `level()` does this in exact integers), take the reported dual, apply the shift, verify `Σ_{R ⊆ T} y_R ≤ [T = ∅]` for every column in `Fraction` arithmetic, and evaluate `Σ_R y_R N_R`. Weak duality then gives a lower bound on the LP optimum and hence on the zone twin count. A **zero** certificate is cheaper to refute than to trust, which is the right way round here: to overturn "level X certifies 0 at p = 83" it suffices to exhibit a dual feasible `y` with `Σ_R y_R N_R > 0`. The claim is that none exists, and the primal witness (a feasible non-negative `x` with `x_∅ = 0`, matching every exact count `N_R` for `prod(R) ≤ X`) is what proves it; the solver returns that witness, and checking it is integer arithmetic over at most 1414 rows.