What Synthetic Context-Selection Tests Measure: Policy, Evidence, and Answer Format

Context-selection benchmarks can reward several different capabilities: following a state convention, choosing evidence, using the resulting context, and producing a required answer format. We examine these distinctions in two controlled studies and a supporting shortcut audit of synthetic conversations. In a prospective 150-source study, each source is queried before and after a reopening instruction. A benchmark-tailored state rule improves Llama-3.2-3B both-stage canonical success from 94/150 to 150/150 under hard selection, a difference of 37.3 percentage points (95% paired interval [32.7, 42.0]). The same rule with a simple symbolic extractor also solves all 150 pairs, and finite attention gating reaches the same Llama score. A fixed post-outcome prompt making the state convention explicit raises full-context success to 112/150, a descriptive gain of 12.0 points [6.0, 18.0]. In a separate 200-source study, one-span suppression and retention errors are matched by carrier role and change from an oracle map, while exposing different final evidence. Its original Qwen3-0.6B attention primary does not resolve a canonical-success ordering: 1.6 points (95% interval [-2.5, 5.9]); registered inactive-value emission is lower under suppression. A later Llama attention cell and literal-output audits reveal further task, execution, and format dependence without changing the original endpoints. The supporting CPU audit shows why swapping requested states alone can leave a first-mention shortcut perfect. Together these case studies distinguish what controlled selection comparisons establish from claims about learned state tracking or natural conversational memory. They introduce neither a general selector nor a pooled effect across studies. Materials: The PDF is the consolidated lead preprint. The manuscript-source ZIP contains its LaTeX sources. The supporting ZIP contains three full study reports, original synthetic datasets, raw and cleaned predictions, analysis code, provenance, and CPU replay instructions. The reports document the studies included in the lead article; they are not separate archival publications. Reproduction: The supplied package reproduces the reported analyses from saved predictions. It does not contain model weights or the complete inference implementation. INFERENCE_REPRODUCTION.md specifies the separate inference source versions and commands. Licensing: Manuscript text and original synthetic data are CC BY 4.0. Original analysis code is MIT. Third-party materials retain their own terms; the package includes the relevant notices. Research assistance: Language-model tools assisted with research design, implementation, literature review, numerical checking, and writing. The author retains responsibility for the evidence, interpretation, and submission. Status: Preprint; not peer reviewed.