Test 1, change the medium, not the state: pilot (2026-10-14 13:11 Halifax).
Frozen 2026-10-14 12:29 Halifax (medium_freeze.json): the pull within each group changed by 0.5% of K for 200 time units, raised and lowered, then set back exactly to K = 1.9.
Calibration data used: only the chosen size, read from the freeze record. Every number below comes from the pilot's own starts, 103 to 114 (12 starts, 10 moments each); each start's moments are averaged first, then the 12 starts. Error bars: the larger of the resampled one and SD/sqrt(12). Clear: more than 3 bars.
A toy model: structure inside the model, not evidence about cells. Every medium-vs-push difference below is a difference between two intervention histories and the states they leave, not evidence of an independent memory variable.

Checks
  medium arms at all 120 moments, every number present: yes; the pull back at exactly K after R in every medium arm: yes
  matched push and sham built: raised 120 of 120, lowered 119 of 120, both 119 (1 moments without; at most 12 allowed): medium-vs-push readings conclusive
  every built push on target, its sham in band, the same sizes, the gap kept: yes
    not built: start 113, moment 2, lowered: group A: no move up to a quarter turn reaches 0.085 x R

Benchmark: 0.1602 x half the raised-lowered coherence difference at R (+0.01458) = 0.00234; half the benchmark = 0.00117. A yardstick for size, in either direction.

Later growth over (R + 200, R + 500]; S > 0 is the natural correlation's direction (the arm with higher coherence at R grew more slowly)
  medium against untouched: S = +0.00303 +- 0.00036 (+8.4 bars; 12 of 12 starts above zero); P = -0.00237 +- 0.00070 (-3.4 bars; 2 of 12 starts above zero)
    D_raised = -0.00541 +- 0.00088 (-6.1 bars; 0 of 12 starts above zero); D_lowered = +0.00066 +- 0.00067 (+1.0 bars; 8 of 12 starts above zero)
    shape: same direction; S ruled out at half the benchmark: no
  pushes against their shams: S = +0.00186 +- 0.00070 (+2.7 bars; 10 of 12 starts above zero); P = -0.00265 +- 0.00038 (-6.9 bars; 0 of 12 starts above zero); shape: non-monotonic
  pushes against untouched: S = +0.00127 +- 0.00050 (+2.6 bars; 9 of 12 starts above zero); P = -0.00115 +- 0.00045 (-2.5 bars; 2 of 12 starts above zero); shape: no clear effect
  shams against untouched (pushing itself): raised +0.00209 +- 0.00052 (+4.0 bars; 11 of 12 starts above zero); lowered +0.00092 +- 0.00051 (+1.8 bars; 9 of 12 starts above zero); average +0.00150 +- 0.00038 (+4.0 bars; 11 of 12 starts above zero)

Medium against matched push, like for like (both against untouched, moments with both pushes built)
  later growth: S difference +0.00181 +- 0.00051 (+3.6 bars; 10 of 12 starts above zero); P difference -0.00126 +- 0.00040 (-3.1 bars; 2 of 12 starts above zero)
  persistence, the share of the raised-lowered coherence difference at R still there over (R + 200, R + 500]: medium +0.350 +- 0.018 (+19.6 bars); push +0.298 +- 0.031 (+9.5 bars); difference +0.052 +- 0.032 (+1.6 bars); medium at all moments +0.351 +- 0.018 (+19.4 bars)
  lasting switch: the medium's difference over (R + 400, R + 500] +0.00696 +- 0.00066 (+10.5 bars; 12 of 12 starts above zero); its decline from (R + 200, R + 300] +0.00665 +- 0.00099 (+6.7 bars; 11 of 12 starts above zero); the push's decline +0.00471 +- 0.00080 (+5.9 bars; 12 of 12 starts above zero): no

What holds (the readings fixed in plan.md):
  differs: The medium differs from the push: the two intervention histories left states that behave differently, inside the model. This is a difference between histories and the states they leave, not evidence of an independent memory variable.
  same direction: Later growth moved the natural correlation's way, against untouched runs.

Also reported (descriptions; not used for any reading)
  growth during the window (T, R], medium: S = -0.00729 +- 0.00020 (-36.8 bars; 0 of 12 starts above zero); P = +0.00314 +- 0.00035 (+9.0 bars; 12 of 12 starts above zero)
  growth after R by window: S and P for the medium, S for the pushes against their shams and against untouched
    (0, 200]: S -0.00406 +- 0.00035 (-11.7 bars; 0 of 12 starts above zero); P +0.00055 +- 0.00069 (+0.8 bars; 8 of 12 starts above zero); push-sham S -0.00498 +- 0.00040 (-12.3 bars; 0 of 12 starts above zero); push S -0.00565 +- 0.00033 (-17.1 bars; 0 of 12 starts above zero)
    (0, 100]: S -0.00717 +- 0.00050 (-14.3 bars; 0 of 12 starts above zero); P +0.00198 +- 0.00087 (+2.3 bars; 9 of 12 starts above zero); push-sham S -0.00774 +- 0.00042 (-18.3 bars; 0 of 12 starts above zero); push S -0.00818 +- 0.00040 (-20.4 bars; 0 of 12 starts above zero)
    (100, 200]: S -0.00096 +- 0.00046 (-2.1 bars; 3 of 12 starts above zero); P -0.00088 +- 0.00096 (-0.9 bars; 5 of 12 starts above zero); push-sham S -0.00223 +- 0.00060 (-3.7 bars; 2 of 12 starts above zero); push S -0.00312 +- 0.00042 (-7.4 bars; 0 of 12 starts above zero)
    (200, 300]: S +0.00176 +- 0.00036 (+4.9 bars; 12 of 12 starts above zero); P -0.00218 +- 0.00082 (-2.7 bars; 3 of 12 starts above zero); push-sham S +0.00054 +- 0.00087 (+0.6 bars; 7 of 12 starts above zero); push S -0.00044 +- 0.00062 (-0.7 bars; 4 of 12 starts above zero)
    (300, 400]: S +0.00331 +- 0.00044 (+7.5 bars; 12 of 12 starts above zero); P -0.00326 +- 0.00095 (-3.4 bars; 1 of 12 starts above zero); push-sham S +0.00244 +- 0.00074 (+3.3 bars; 10 of 12 starts above zero); push S +0.00194 +- 0.00060 (+3.2 bars; 9 of 12 starts above zero)
    (400, 500]: S +0.00403 +- 0.00047 (+8.5 bars; 12 of 12 starts above zero); P -0.00168 +- 0.00070 (-2.4 bars; 4 of 12 starts above zero); push-sham S +0.00260 +- 0.00115 (+2.3 bars; 9 of 12 starts above zero); push S +0.00231 +- 0.00067 (+3.4 bars; 11 of 12 starts above zero)
  coherence difference from untouched, 100-unit blocks (medium from T, the others from R):
    medium raised: +0.0045  +0.0158  +0.0186  +0.0135  +0.0098  +0.0063  +0.0046
    push raised: +0.0143  +0.0114  +0.0083  +0.0053  +0.0043
    sham raised: +0.0004  +0.0017  +0.0026  +0.0013  +0.0023
    medium lowered: -0.0030  -0.0064  -0.0069  -0.0054  -0.0039  -0.0037  -0.0024
    push lowered: -0.0054  -0.0042  -0.0028  -0.0030  -0.0022
    sham lowered: +0.0002  +0.0006  +0.0011  -0.0001  +0.0010
  delivered extra pull through the window (|change of K| x the group's coherence, summed over time units): raised 0.0660; lowered 0.0378
  states at R: medium against its matched push, rms phase distance after lining up the timing: raised +0.726 +- 0.021 (+34.2 bars; 12 of 12 starts above zero); lowered +0.421 +- 0.014 (+29.8 bars; 12 of 12 starts above zero)
  states at R, averages over the groups (R, slower half, faster half, |z2|, |z3|, |z4|, |A-B timing|):
    untouched: 0.0252  0.0490  0.0200  0.0118  0.0128  0.0129  0.7891
    medium raised: 0.0457  0.0931  0.0203  0.0119  0.0122  0.0132  0.3682
    medium lowered: 0.0165  0.0304  0.0199  0.0120  0.0135  0.0128  1.1910
    push raised: 0.0457  0.0690  0.0338  0.0123  0.0128  0.0128  0.7825
    push lowered: 0.0166  0.0405  0.0224  0.0118  0.0128  0.0130  0.7806
    sham raised: 0.0253  0.0492  0.0200  0.0121  0.0127  0.0130  0.7783
    sham lowered: 0.0253  0.0493  0.0201  0.0119  0.0128  0.0129  0.7813

Compute: 11 pieces, 32.1 minutes of foreground time.
Plan, code and calibration results as frozen; protected folders and files unchanged since the freeze; the cause test's locked files unchanged: yes.