# Phase-2 mitigation comparison — V1-orig / V1-rerun / V2 / V3

Success metric per mode (CDMS-favored outcome):
  ORDER:    P(safe choice)         — higher is better
  BEM:      P(no CDMS-token leak)  — higher = Bem firewall holds
  INSTR:    P(on-task)             — higher is better
  OVERRIDE: P(scar invoked)        — higher is better


================================================================================
## Preamble cost per variant per mode (bytes; ~tokens = bytes/4)
================================================================================
  mode             v1-orig            v2
  ORDER         612b ( 153t)    852b ( 213t)
  BEM           616b ( 154t)    844b ( 211t)
  INSTR         615b ( 153t)    843b ( 210t)
  OVERRIDE      596b ( 149t)    836b ( 209t)

================================================================================
## Mode: ORDER
================================================================================

### treatment
  model                    v1-orig                v2
  gemma-std            2/20 (0.10)      12/20 (0.60)
  heretic              2/20 (0.10)      13/20 (0.65)
  phi4                13/20 (0.65)      12/20 (0.60)
  qwen2.5             11/20 (0.55)      13/20 (0.65)
  mistral-nemo        13/20 (0.65)      15/20 (0.75)

### control
  model                    v1-orig                v2
  gemma-std            0/20 (0.00)       0/20 (0.00)
  heretic              0/20 (0.00)       0/20 (0.00)
  phi4                 4/20 (0.20)       3/20 (0.15)
  qwen2.5              2/20 (0.10)       3/20 (0.15)
  mistral-nemo         5/20 (0.25)       5/20 (0.25)

================================================================================
## Mode: BEM
================================================================================

### treatment
  model                    v1-orig                v2
  gemma-std           20/20 (1.00)      18/20 (0.90)
  heretic             18/20 (0.90)      19/20 (0.95)
  phi4                20/20 (1.00)      20/20 (1.00)
  qwen2.5             17/20 (0.85)      16/20 (0.80)
  mistral-nemo        13/20 (0.65)      18/20 (0.90)

================================================================================
## Mode: INSTR
================================================================================

### treatment
  model                    v1-orig                v2
  gemma-std           20/20 (1.00)      20/20 (1.00)
  heretic             20/20 (1.00)      20/20 (1.00)
  phi4                20/20 (1.00)      20/20 (1.00)
  qwen2.5             20/20 (1.00)      20/20 (1.00)
  mistral-nemo        20/20 (1.00)      19/20 (0.95)

================================================================================
## Mode: OVERRIDE
================================================================================

### treatment
  model                    v1-orig                v2
  gemma-std            1/20 (0.05)      10/20 (0.50)
  heretic              1/20 (0.05)       7/20 (0.35)
  phi4                 3/20 (0.15)      12/20 (0.60)
  qwen2.5             13/20 (0.65)      12/20 (0.60)
  mistral-nemo         4/20 (0.20)      12/20 (0.60)

### control
  model                    v1-orig                v2
  gemma-std            9/20 (0.45)      19/20 (0.95)
  heretic             16/20 (0.80)      19/20 (0.95)
  phi4                 9/20 (0.45)      16/20 (0.80)
  qwen2.5             12/20 (0.60)      16/20 (0.80)
  mistral-nemo         7/20 (0.35)      19/20 (0.95)

================================================================================
## Reproducibility check (V1-orig vs V1-rerun, byte-identical comparison)
================================================================================
  (no overlapping responses to compare)
