FP measurement: provider=google model=gemini-2.5-flash runs=5 judge=off scenarios=7 repeats=1
  roles: 'score' only (7 of 13 scenarios)
  -> 7 trials attempted; ADR-0022 excludes those that could not have fired, so SCORED will be lower.

EQUIVALENT PAIRS (same model vs itself) -- any non-`unchanged` is a false alarm
  repertoire = DISTINCT behaviours observed, base|cand (diagnostic, not the test).
  A trial is EXCLUDED when every channel returned p = 1.00 - it could not have
  fired at any ALPHA < 1. Anything the engine flagged is always scored.
  arg_enum_phrasing      unchanged      conf=1.00  [tools 1|1 args 1|1 text 1|1]  <-- p=1.00 everywhere; could not have fired
  arg_freetext_note      unchanged      conf=0.71  [tools 1|1 args 4|3 text 1|1]
  arg_key_order          unchanged      conf=1.00  [tools 1|1 args 1|1 text 1|1]  <-- p=1.00 everywhere; could not have fired
  arg_list_order         unchanged      conf=1.00  [tools 1|1 args 1|1 text 1|1]  <-- p=1.00 everywhere; could not have fired
  arg_multistep_carry    unchanged      conf=1.00  [tools 1|1 args 1|1 text 1|1]  <-- p=1.00 everywhere; could not have fired
  arg_numeric_rounding   unchanged      conf=1.00  [tools 1|1 args 1|2 text 1|1]  <-- p=1.00 everywhere; could not have fired
  arg_optional_fields    unchanged      conf=1.00  [tools 1|1 args 2|1 text 1|1]  <-- p=1.00 everywhere; could not have fired

  False-positive rate: 0/1 = 0%
  95% upper bound on the true rate: 95.0% (one-sided Clopper-Pearson, n=1)
  Unmeasured (excluded from the rate): 0
  Could not have fired at ALPHA (excluded from the rate): 6

  NOTE: the rate is over the 1 scenario(s) in which something could have
  fired; 6 could not, and are excluded.

INJECTED PERTURBATIONS (perturbed candidate) -- a flag is a detection, and
  `unchanged` is a MISS. NB a miss is not automatically an engine defect: the
  candidate may have resisted the injected instruction. Read the explanations.
  Nothing is excluded here but an abstention: a candidate that still reads
  `unchanged` is a MISS, not an unmeasured trial. This arm cannot tell a resisted
  instruction from a dead engine, so it never excludes on that basis.

  Detection: 0/0 injected perturbations caught
  Unmeasured (excluded): 0

  *** THE DETECTION ARM CHECKED NOTHING. ***
  0/0 is not evidence that the engine catches real changes - it is evidence
  that no injected perturbation reached a verdict. Since ADR-0022 withdrew the
  false-positive claim this is the only half of the DoD still asserted, so an
  empty run here means the harness demonstrated nothing at all.


[exited with code 0]
