[cyberuniversal]

Closes #198

Summary

• publishes the completed issue-196 corpus-free fixed-checkpoint summary and audit;
• adds fixed epoch 1, fixed epoch 2, and dev-selected results to the research status documents;
• updates the manuscript abstract, methods, results, conclusion, and limitations with a compact checkpoint-policy table;
• states the supported interpretation: checkpoint selection amplifies the effect magnitude but does not create the mixed-over-synthetic direction.

Key result

• fixed epoch 1: F2 21.68%, F3 27.79%, gap +6.11 pp;
• fixed epoch 2: F2 25.40%, F3 31.98%, gap +6.58 pp;
• dev-selected: F2 21.68%, F3 31.98%, gap +10.29 pp;
• F3 is higher in all five seeds under both fixed policies.

Validation

• full repository suite: 316 passed, 2 optional data skips;
• compileall, JSON parsing/consistency, anonymous identifier/credential scan, and diff checks passed;
• Tectonic 0.16.9 build passed;
• final PDF: 8 A4 pages, visually inspected page by page with no clipping, overlap, broken tables, or illegible content;
• official MRL 2026 submission call rechecked: anonymized EMNLP 2025 format, 4/8 pages excluding references.

Safeguards

Publication-only work used the reviewed corpus-free aggregate. No private prediction read, Kubernetes object, GPU, model load, inference, training, metric rerun, checkpoint selection/reselection, QALB test, diagnostics, prompt/parser change, or XG occurred.