# Candidate statistic and falsifier, job1136 Written before any statistic engine, target run or control run exists. Decision: can phase-aware conditional Hall-triad frequency reveal information absent from the complete per-prime Boolean pair-compatibility graphs? The candidate statistic is Delta(D,Q)=H_phase(D,Q)-H_graph(D,Q). For every pair e={s,t}, let B_q(s)={-s,-s-2} modq and L(e)={q:B_q(s) intersects B_q(t)}. A candidate triad consists of three disjoint pairs, all with the same exact two-prime list L(e)={q,r}. H_phase counts candidate triads whose three pair-phase sets E_q(e)=intersection_{s in e}B_q(s) are pairwise disjoint at each owner q,r. H_graph uses only Boolean pair graphs: for each two candidate pairs e,f and owner q,r, require that their four slots do NOT form a complete clique in G_owner. Count an unordered triad once. Predeclared null benchmark: condition on every G_q and retain the literal cycle-edge incidence family with all q>3. Every allowed control must have those exact graphs; independent phase-coordinate renamings provide a matched, though potentially trivial, such control. A broader sampler may not discard the cycle-edge/Helly constraint silently. The benchmark prediction is Delta=0 for every permitted control and input, at ANY scale, not merely in expectation. Falsifier: one nonzero Delta on a fully reconstructed literal q>3 input falsifies the implementation/model agreement needed for this candidate. Before allocating a producer, if the q>3 cycle-edge Helly lemma logically forces Delta=0, reject Delta as a scientific test of additional phase information. A positive value on a non-Helly abstract control is outside the null family and cannot rescue arithmetic sensitivity. Planned numerical scale if sensitivity survived: frozen D51,Q101..193 first, using return478's archived target and control measurements without regeneration. A prospective implementation would cap one missing graph-versus-phase comparison at60 CPU seconds,0.5GB RAM and0.05GB disk, one thread,15 judgment minutes. It would stop after a complete nonzero witness or exact comparison; no census/control replay. That compute is NOT authorized by this design if the identifiability gate fails. No target statistic, control draw, numeric comparison or producer was executed in job1136. The mathematical identifiability gate decides whether to run.