==============================================================================
C2b-0a — is gpt-oss-safeguard-120b a fixed point of its own quantizer?
==============================================================================
  preserve       lossless  72/72 tensors byte-identical, max_abs_err=0, mismatched_bytes=0, mismatched_scale_blocks=0 (10248s)
  recompute-up   lossless  72/72 tensors byte-identical, max_abs_err=0, mismatched_bytes=0, mismatched_scale_blocks=0 (10247s)
  blocks=3,583,180,800  elements=114,661,785,600  negative_zero_codes=7,200,168,322  nan_scale_blocks=0

  VERDICT: safeguard is A FIXED POINT under both policies.
  abort A5 (NOT triggered): blend work may proceed on the quantization side.

==============================================================================
C2b-0b (i) — the shared-exponent census
==============================================================================
  blocks compared              1,194,393,600
  blocks byte-identical        1,145,057,192 (95.869334%)
  blocks whose E8M0 differs    397,501 (0.033281%)
  int(scale_r) - int(scale_c):  -1:197,366  +0:1,193,996,099  +1:200,135

==============================================================================
C2b-0b (ii) — the code-distance census (ladder steps)
==============================================================================
  elements                             38,220,595,200
  elements in exponent-matched blocks  38,207,875,168 (99.966719%)
  codes differing between parents      65,430,005 (0.171190%)
  values differing between parents     50,522,990 (0.132188%)
  of which sign-of-zero only           4,411,478

  |ladder distance| histogram, all elements:
     0:     38,159,579,979  (99.840360% of elements)
     1:         49,489,844  (0.129485% of elements)   [81.1107% of moved]
     2:         11,514,700  (0.030127% of elements)   [18.8718% of moved]
     3:             10,677  (0.000028% of elements)   [0.0175% of moved]
    one step as a share of everything that moved: 81.1107%   -> P-C2b-Q clause (a) HOLDS for all elements

  |ladder distance| histogram, exponent-matched only:
     0:     38,157,778,162  (99.868883% of elements)
     1:         45,858,387  (0.120023% of elements)   [91.5392% of moved]
     2:          4,238,619  (0.011094% of elements)   [8.4608% of moved]
    one step as a share of everything that moved: 91.5392%   -> P-C2b-Q clause (a) HOLDS for exponent-matched only

==============================================================================
C2b-0b (iii) — the alpha grid (P-C2b-Q)
==============================================================================
   alpha    values changed vs candidate         frac    eps_rel  signal_rel  digest
    0.00                              0   0.000000%  0.000e+00   0.000e+00  e265f10e0e03
    0.10                              0   0.000000%  6.909e-05   6.909e-05  555f1301e2ba  <- eps DOMINATES signal
    0.30                      4,238,619   0.011090%  1.937e-04   2.073e-04  8e9171aff39a
    0.50                     27,562,160   0.072113%  3.036e-04   3.454e-04  1a92af93a546
    0.70                     50,522,990   0.132188%  1.937e-04   4.836e-04  b0521916c50b
    0.90                     50,522,990   0.132188%  6.909e-05   6.218e-04  ae4a353006a6
    1.00                     50,522,990   0.132188%  0.000e+00   6.909e-04  5456d3a8a4f5

  pairwise elements whose CODE differs between alpha outputs (fraction of all elements; digests above are the distinctness authority):
               0.00       0.10       0.30       0.50       0.70       0.90       1.00
   0.00    0.00e+00   1.44e-04   2.55e-04   8.60e-04   1.46e-03   1.46e-03   1.60e-03
   0.10    1.44e-04   0.00e+00   1.11e-04   7.21e-04   1.32e-03   1.32e-03   1.46e-03
   0.30    2.55e-04   1.11e-04   0.00e+00   6.10e-04   1.21e-03   1.32e-03   1.46e-03
   0.50    8.60e-04   7.21e-04   6.10e-04   0.00e+00   6.01e-04   7.12e-04   8.48e-04
   0.70    1.46e-03   1.32e-03   1.21e-03   6.01e-04   0.00e+00   1.11e-04   2.53e-04
   0.90    1.46e-03   1.32e-03   1.32e-03   7.12e-04   1.11e-04   0.00e+00   1.42e-04
   1.00    1.60e-03   1.46e-03   1.46e-03   8.48e-04   2.53e-04   1.42e-04   0.00e+00

  distinct byte outputs over the interior grid [0.1, 0.3, 0.5, 0.7, 0.9]: 5 of 5
  distinct byte outputs over the full grid [0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0]: 7 of 7
    1a92af93a546  <- alphas [0.5]
    5456d3a8a4f5  <- alphas [1.0]
    555f1301e2ba  <- alphas [0.1]
    8e9171aff39a  <- alphas [0.3]
    ae4a353006a6  <- alphas [0.9]
    b0521916c50b  <- alphas [0.7]
    e265f10e0e03  <- alphas [0.0]

  P-C2b-Q (literal falsifier: three or more distinct outputs over the interior grid)
  VERDICT: FALSIFIED — the grid produced 5 distinct outputs, not two.
  quantitative companion: the largest adjacent-alpha separation anywhere on the interior grid is 23,323,541 elements (0.0610% of the checkpoint).

==============================================================================
The C2b gate (c2-spec.json c2b.expert.arms_if_licensed._gate)
==============================================================================
  'run ONLY if C2b-0b shows a materially non-degenerate staircase — i.e. the
   alpha grid produces more than two distinct byte outputs.'
  -> distinct outputs on the interior grid = 5: expert blends are LICENSED
