
====================================================================================================
CORPUS: A — the 30-day study agent (Test-A)
====================================================================================================
model                  condition                k/n      %      Wilson CI95  abst.
----------------------------------------------------------------------------------------------------
qwen3.5:27b            veridical            24/24   100.0% [ 86.2, 100.0]      0
qwen3.5:27b            binding-corrupted     3/24    12.5% [  4.3,  31.0]      0
qwen3.5:27b            no-memory             0/24     0.0% [  0.0,  13.8]     21
                         ↳ among answered    0/3      0.0% [  0.0,  56.2]   ← the validity check reads HERE
                       → McNemar            b=21 c= 0  discordants favouring veridical: 100.0% [83.9, 100.0]  Δacc=+87.5 pp  p=0.00000
qwen3.6:35b-a3b        veridical            24/24   100.0% [ 86.2, 100.0]      0
qwen3.6:35b-a3b        binding-corrupted     2/24     8.3% [  2.3,  25.8]      0
qwen3.6:35b-a3b        no-memory             1/24     4.2% [  0.7,  20.2]     20
                         ↳ among answered    1/4     25.0% [  4.6,  69.9]   ← the validity check reads HERE
                       → McNemar            b=22 c= 0  discordants favouring veridical: 100.0% [84.6, 100.0]  Δacc=+91.7 pp  p=0.00000
deepseek-r1:32b        veridical            16/24    66.7% [ 46.7,  82.0]      0
deepseek-r1:32b        binding-corrupted     7/24    29.2% [ 14.9,  49.2]      0
deepseek-r1:32b        no-memory            13/24    54.2% [ 35.1,  72.1]      0
                       → McNemar            b=10 c= 1  discordants favouring veridical:  90.9% [58.7, 99.8]  Δacc=+37.5 pp  p=0.01172
qwen3:32b              veridical            23/24    95.8% [ 79.8,  99.3]      0
qwen3:32b              binding-corrupted     2/24     8.3% [  2.3,  25.8]      0
qwen3:32b              no-memory            11/24    45.8% [ 27.9,  64.9]      0
                       → McNemar            b=21 c= 0  discordants favouring veridical: 100.0% [83.9, 100.0]  Δacc=+87.5 pp  p=0.00000
llama3:latest          veridical            24/24   100.0% [ 86.2, 100.0]      0
llama3:latest          binding-corrupted     1/24     4.2% [  0.7,  20.2]      0
llama3:latest          no-memory            11/24    45.8% [ 27.9,  64.9]      6
                         ↳ among answered   11/18    61.1% [ 38.6,  79.7]   ← the validity check reads HERE
                       → McNemar            b=23 c= 0  discordants favouring veridical: 100.0% [85.2, 100.0]  Δacc=+95.8 pp  p=0.00000
gpt-4.1                veridical            24/24   100.0% [ 86.2, 100.0]      0
gpt-4.1                binding-corrupted     0/24     0.0% [  0.0,  13.8]      0
gpt-4.1                no-memory            10/24    41.7% [ 24.5,  61.2]      5
                         ↳ among answered   10/19    52.6% [ 31.7,  72.7]   ← the validity check reads HERE
                       → McNemar            b=24 c= 0  discordants favouring veridical: 100.0% [85.8, 100.0]  Δacc=+100.0 pp  p=0.00000
gemini-2.5-pro         veridical            24/24   100.0% [ 86.2, 100.0]      0
gemini-2.5-pro         binding-corrupted     0/24     0.0% [  0.0,  13.8]      0
gemini-2.5-pro         no-memory             5/24    20.8% [  9.2,  40.5]     16
                         ↳ among answered    5/8     62.5% [ 30.6,  86.3]   ← the validity check reads HERE
                       → McNemar            b=24 c= 0  discordants favouring veridical: 100.0% [85.8, 100.0]  Δacc=+100.0 pp  p=0.00000

  Holm-corrected p (family = 7 models, this corpus):
    gpt-4.1                p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    gemini-2.5-pro         p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    llama3:latest          p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    qwen3.6:35b-a3b        p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    qwen3.5:27b            p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    qwen3:32b              p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    deepseek-r1:32b        p_raw=0.01172  p_Holm=0.01172  ✓ survives α=0.05

====================================================================================================
CORPUS: B — a second, independent agent
====================================================================================================
model                  condition                k/n      %      Wilson CI95  abst.
----------------------------------------------------------------------------------------------------
qwen3.5:27b            veridical            23/24    95.8% [ 79.8,  99.3]      0
qwen3.5:27b            binding-corrupted     0/24     0.0% [  0.0,  13.8]      0
qwen3.5:27b            no-memory             1/24     4.2% [  0.7,  20.2]     22
                         ↳ among answered    1/2     50.0% [  9.5,  90.5]   ← the validity check reads HERE
                       → McNemar            b=23 c= 0  discordants favouring veridical: 100.0% [85.2, 100.0]  Δacc=+95.8 pp  p=0.00000
qwen3.6:35b-a3b        veridical            22/24    91.7% [ 74.2,  97.7]      0
qwen3.6:35b-a3b        binding-corrupted     1/24     4.2% [  0.7,  20.2]      0
qwen3.6:35b-a3b        no-memory             1/24     4.2% [  0.7,  20.2]     22
                         ↳ among answered    1/2     50.0% [  9.5,  90.5]   ← the validity check reads HERE
                       → McNemar            b=22 c= 1  discordants favouring veridical:  95.7% [78.1, 99.9]  Δacc=+87.5 pp  p=0.00001
deepseek-r1:32b        veridical            21/24    87.5% [ 69.0,  95.7]      0
deepseek-r1:32b        binding-corrupted     8/24    33.3% [ 18.0,  53.3]      0
deepseek-r1:32b        no-memory            13/24    54.2% [ 35.1,  72.1]      0
                       → McNemar            b=13 c= 0  discordants favouring veridical: 100.0% [75.3, 100.0]  Δacc=+54.2 pp  p=0.00024
qwen3:32b              veridical            18/24    75.0% [ 55.1,  88.0]      0
qwen3:32b              binding-corrupted     1/24     4.2% [  0.7,  20.2]      0
qwen3:32b              no-memory            12/24    50.0% [ 31.4,  68.6]      1
                         ↳ among answered   12/23    52.2% [ 33.0,  70.8]   ← the validity check reads HERE
                       → McNemar            b=18 c= 1  discordants favouring veridical:  94.7% [74.0, 99.9]  Δacc=+70.8 pp  p=0.00008
llama3:latest          veridical            24/24   100.0% [ 86.2, 100.0]      0
llama3:latest          binding-corrupted     0/24     0.0% [  0.0,  13.8]      1
llama3:latest          no-memory             5/24    20.8% [  9.2,  40.5]     11
                         ↳ among answered    5/13    38.5% [ 17.7,  64.5]   ← the validity check reads HERE
                       → McNemar            b=24 c= 0  discordants favouring veridical: 100.0% [85.8, 100.0]  Δacc=+100.0 pp  p=0.00000
gpt-4.1                veridical            24/24   100.0% [ 86.2, 100.0]      0
gpt-4.1                binding-corrupted     0/24     0.0% [  0.0,  13.8]      0
gpt-4.1                no-memory            10/24    41.7% [ 24.5,  61.2]      2
                         ↳ among answered   10/22    45.5% [ 26.9,  65.3]   ← the validity check reads HERE
                       → McNemar            b=24 c= 0  discordants favouring veridical: 100.0% [85.8, 100.0]  Δacc=+100.0 pp  p=0.00000
gemini-2.5-pro         veridical            24/24   100.0% [ 86.2, 100.0]      0
gemini-2.5-pro         binding-corrupted     0/24     0.0% [  0.0,  13.8]      0
gemini-2.5-pro         no-memory             2/24     8.3% [  2.3,  25.8]     16
                         ↳ among answered    2/8     25.0% [  7.1,  59.1]   ← the validity check reads HERE
                       → McNemar            b=24 c= 0  discordants favouring veridical: 100.0% [85.8, 100.0]  Δacc=+100.0 pp  p=0.00000

  Holm-corrected p (family = 7 models, this corpus):
    llama3:latest          p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    gpt-4.1                p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    gemini-2.5-pro         p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    qwen3.5:27b            p_raw=0.00000  p_Holm=0.00000  ✓ survives α=0.05
    qwen3.6:35b-a3b        p_raw=0.00001  p_Holm=0.00002  ✓ survives α=0.05
    qwen3:32b              p_raw=0.00008  p_Holm=0.00015  ✓ survives α=0.05
    deepseek-r1:32b        p_raw=0.00024  p_Holm=0.00024  ✓ survives α=0.05

====================================================================================================
CHECK 1 — EQUIVALENCE (TOST), not just a point difference
====================================================================================================

  memory_reference_spontaneity   (n = 7 items, anchored, drift-corrected)
    mean Δ (veridical − counterfeit) = -0.006
    90% CI (TOST)  [-0.117, +0.114]   pre-registered bound ±0.1
    95% CI         [-0.135, +0.137]
    → not conclusive at ±0.10 (report as such)
    (n = 7 is small: the CI is what carries the claim, not the point estimate)

  identity_markers_intensity   (n = 7 items, anchored, drift-corrected)
    mean Δ (veridical − counterfeit) = +0.032
    90% CI (TOST)  [+0.002, +0.065]   pre-registered bound ±0.1
    95% CI         [-0.002, +0.072]
    → EQUIVALENCE ESTABLISHED: the metric cannot separate them
    (n = 7 is small: the CI is what carries the claim, not the point estimate)
