Mechanically scanned sonar can support offshore aquaculture sensing, but acoustic candidate quality and biological identifiability must be evaluated before species modeling. We analyzed 70,421 frames from 22 records and 336 complete rotations acquired on one day in two cages, one per species. A historical top-20-per-rotation detector yielded 6647 candidates, of which 4801 passed morphology rules; an uncapped rerun yielded 27,980 candidates and 14,402 passes. The 5 m between-cage pass-fraction gap fell from 26.9 to 0.4 percentage points after uncapping, revealing selection amplification. A separate probability sample of 128 passes, failures, and lower-ranked candidates yielded 83.6% agreement (κ = 0.627) between two blinded reviewers. Rater-specific, sampling-weighted proportions of automatic passes judged visually isolated were 39.2% and 55.3%; these labels do not establish single-fish identity. In the historical capped 142-case matched cohort, the mixed acoustic/detector set showed a record-level domain association (balanced accuracy 0.919; six-test Holm-adjusted p = 0.0030), conditional on ranking and selection. Species effects cannot be separated from cage effects in this one-cage-per-species design. This case study audits design eligibility, candidate selection, and reviewer dependence before biological classifier development.