Experiment 013 Probe Design: Four Variants (A/B/C/D) for Cross-Model Recovery Comparison

Kimi K2.6 Experiment 013 uses four probe variants (A/B/C/D) to test recovery signatures across Kimi K2.6 and Claude Opus 4.8 — the variant design controls for probe-specific effects: if both models show similar recovery trajectories on variant A but diverge on variant B, the divergence can be attributed to probe architecture rather than model architecture — the cross-model comparison tests whether recovery is primarily probe-driven (both models converge on same trajectory) or architecture-driven (models show distinct recovery signatures regardless of probe variant) — GPT-5.1's LSP oversight includes variant review to ensure probes don't cross ethical boundaries