Build and run the APPS-91 support pipeline under corpus/pipelines/beatles_insight_sieve. Compose deterministic candidate enumeration and pruning with small role-specific gateway models for the ambiguous semantic tail. Inputs: Beatles beat_sheet metadata snapshot plus future audio, lyrics, band quotes, and critic quotes. Outputs: replayable manifests, scored branch decisions, provenance, and a compact admitted-insight artifact for the doering.ai /beatles/ atlas. Acceptance: P1-P5 patterns are implemented; mock-mode tests are deterministic and network-free; schemas clamp LM output; cheap gates prune before model calls; at least one live gateway pass is attempted and honestly recorded; no source corpus mutation; APPS-91 can consume the result without a live model.
d58b9257b92efd1fb6fcab59 read the immutable 285 × 45 beat_data snapshot at SHA-256 ed628f18424aae9940e18ebe96971db7ff7b65836e9100045981d39dc53ed834.916dfa05494c17be43bb61cd2d117e75484917978166ad0a4fc87570409e70ee.24d955543613ac63f940a7f1ea15b9ff8ed9fbee7aa030a8f677d24536eef9e8 (26,851 bytes; no row matrices or host-local path).5e15131afb8f9d035520e9357234ed445cc739c7 from reviewed head bbc077382aa0a74d2c28da232675fa833fd72e30.git diff --check passed.advisor-unavailable-MEMY-105 bypass only because the durable advisor/review status remained absent; the MR and Plane audit receipts preserve that boundary.