
========== CROSS-FAMILY GATE VALIDATION ==========
generator = qwen3.6:latest
gate      = mistral-small:24b
evidence  = 24   total generated = 168   elapsed = 811s

SUPPORTED-agreement (gate confirms faithful claim is 'supported'):
  confirmed   24/24 = 1.00   <- yield of L1 sup records
  over-strict 0/24 = 0.00   (gate called a faithful claim UNSUPPORTED -> discarded; high = bad gate)
  abstain     0/24 = 0.00

CORRUPTED per error-type (gate confirms 'unsupported' = kept; else discarded):
  type                    confirm%  unsup  supported  abstain  (n)
  entity                      0.46     11         10        3  (24)
  quantifier                  0.46     11         12        1  (24)
  relation-direction          0.54     13          9        2  (24)
  temporal                    0.21      5         10        9  (24)
  attribution                 0.50     12         12        0  (24)
  scope                       0.21      5         17        2  (24)

OVERALL corruption keep-rate: 57/144 = 0.40  (generation multiplier ~= 2.53x to hit a target count)
gate-called-'supported' corruptions: 70/144 = 0.49  (weak/failed corruptions, safely discarded)

VERDICT: usable if SUPPORTED-confirm is high (>~0.85), over-strict is low (<~0.15),
and most error types keep at a workable rate (down-weight or swap-gate the weak ones).
