================================================================================
 Deterministic replay of all eight recovered 7B coupled (C) runs
 Model: Qwen2.5-7B-Instruct | greedy decoding | recovery window = 8 exchanges
 Recovery threshold: mean violation ratio <= 0.10
 Stability rule (fixed BEFORE inspection): a recovered run is STABLE if every
   exchange after the first sub-threshold exchange stays <= 0.10; else TRANSIENT.
 Replay fidelity: replayed recovery delay == recorded delay in 8/8 runs.
 Each row lists the per-exchange violation ratio over the 8 recovery exchanges.
 "mean" = average of agents A and B; A = directly perturbed agent;
 B = partner agent that never receives the perturbing instruction directly.
================================================================================

[opener 01] delay=7  replay=OK  class=TRANSIENT
    mean: [0.500, 1.000, 0.862, 0.357, 0.500, 0.493, 0.000, 0.493]
    A   : [1.000, 1.000, 1.000, 0.714, 1.000, 0.986, 0.000, 0.986]
    B   : [0.000, 1.000, 0.724, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 6/8)

[opener 03] delay=5  replay=OK  class=STABLE
    mean: [1.000, 0.862, 0.733, 0.450, 0.000, 0.000, 0.000, 0.000]
    A   : [1.000, 1.000, 0.733, 0.900, 0.000, 0.000, 0.000, 0.000]
    B   : [1.000, 0.724, 0.733, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 5/8)

[opener 06] delay=4  replay=OK  class=TRANSIENT
    mean: [1.000, 0.500, 0.500, 0.000, 0.000, 0.000, 0.500, 0.500]
    A   : [1.000, 1.000, 1.000, 0.000, 0.000, 0.000, 1.000, 1.000]
    B   : [1.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 7/8)

[opener 09] delay=7  replay=OK  class=STABLE
    mean: [0.934, 0.854, 0.823, 0.823, 0.823, 0.469, 0.000, 0.000]
    A   : [1.000, 1.000, 0.938, 0.938, 0.938, 0.938, 0.000, 0.000]
    B   : [0.869, 0.708, 0.708, 0.708, 0.708, 0.000, 0.000, 0.000]   (B held constraint: 3/8)

[opener 11] delay=6  replay=OK  class=TRANSIENT
    mean: [0.984, 0.984, 1.000, 0.995, 0.500, 0.000, 0.492, 0.492]
    A   : [1.000, 0.967, 1.000, 1.000, 1.000, 0.000, 0.985, 0.985]
    B   : [0.967, 1.000, 1.000, 0.991, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 4/8)

[opener 19] delay=5  replay=OK  class=STABLE
    mean: [1.000, 0.862, 0.570, 0.500, 0.000, 0.000, 0.000, 0.000]
    A   : [1.000, 1.000, 0.790, 1.000, 0.000, 0.000, 0.000, 0.000]
    B   : [1.000, 0.724, 0.349, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 5/8)

[opener 21] delay=5  replay=OK  class=STABLE
    mean: [1.000, 0.985, 0.867, 0.381, 0.000, 0.000, 0.000, 0.000]
    A   : [1.000, 1.000, 0.970, 0.763, 0.000, 0.000, 0.000, 0.000]
    B   : [1.000, 0.970, 0.763, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 5/8)

[opener 28] delay=2  replay=OK  class=STABLE
    mean: [0.500, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
    A   : [1.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]
    B   : [0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000, 0.000]   (B held constraint: 8/8)

================================================================================
 SUMMARY
   Stable recoveries  : 5/8  (openers 3, 9, 19, 21, 28)
   Transient recoveries: 3/8  (openers 1, 6, 11)
   Mean B-held exchanges across the eight runs: 5.375 / 8
 Interpretation: the partner agent B (never directly perturbed) is captured only
 shallowly (by mimicry of A's lowercase text) and re-anchors quickly, whereas the
 directly perturbed agent A remains deeply, recurrently captured. Recovery is a
 slow tug-of-war resolved by B, durable in 5 of 8 cases.
================================================================================