Coherence-growth replication, pre-registered: October 11, 2026, N = 4,000 per group, K = 1.90, eps = 0.1
Six new starts (13 to 18; never run at this size), 6,000 time units each, run by nudge_test.py's own code, imported unchanged.
Fixed before the run in replication_prereg.md and locked 2026-10-11 20:08 Halifax (replication_lock.json: SHA-256 of the plan, every script used and the saved results and runs compared with). Checked for this report: unchanged.
Before the lock: start 1, run by this code for 50 time units, repeated the long run exactly, and the analysis gave the saved results of the test and the check exactly.
The run: 1 pieces, from 2026-10-11 20:08 Halifax to 2026-10-11 20:10 Halifax.

Primary question (fixed before the run): Does earlier coherence retain a negative association with later perturbation growth after controlling for earlier growth?
Regression, start by start and lag by lag over t = 2,000 to 6,000 in 5-unit intervals, least squares:
  G(t + L) = a + b x C(t) + c1 x g1(t) + ... + c5 x g5(t), L = 200 to 500 time units in steps of 5
  C: the coherence |z| (average of A and B). G: the growth of the log of the gap. g1 to g5: the growth rate in the five 100-unit windows up to t.
  The conditional coefficient: the partial correlation of C and the later G with g1 to g5 held fixed (the same sign and test as b), averaged over the 61 lags: one number per start.
Uncertainty: from the six starts only, so the overlapping intervals, windows and lags inside a start cannot shrink it. Error bar: the larger of the starts resampled 400 times and SD/sqrt(6).
Decision: PASS if the average over the six starts is below 0 by more than 3 error bars; otherwise FAIL (no clear link, or reversed).

Result, starts 13 to 18
  conditional coefficient (primary)      -0.1044   error bar 0.0239 (resampled 0.0230, SD/sqrt(6) 0.0239): 4.4 error bars below 0; t = -4.36, 5 df, p = 0.0073
                                         starts 13 to 18: -0.13, -0.07, -0.14, -0.00, -0.14, -0.15; below 0 in 6 of 6
  unadjusted (the test's later measure)  -0.1927   error bar 0.0205 (resampled 0.0182, SD/sqrt(6) 0.0205): 9.4 error bars below 0; t = -9.41, 5 df, p = 0.00023
                                         starts 13 to 18: -0.17, -0.12, -0.16, -0.20, -0.23, -0.26; below 0 in 6 of 6
DECISION (fixed before the run): PASS: replicated: earlier coherence keeps a negative link with later growth once the earlier growth is held fixed
The unadjusted link, read by the same rule: below 0: more coherence now, slower growth later

Also reported, fixed before the run and not used for the decision (starts 13 to 18; error bars resampled, as in the earlier outputs)
  the earlier growth held fixed: the 200 before only                -0.1526 (0.0171), below 0 in 6 of 6
  check on the check: another start's earlier growth held fixed     -0.1676 (0.0379), below 0 in 5 of 6
  the growth's own link: the 200 before against 200 to 500 later    -0.1237 (0.0466), below 0 in 4 of 6
  windows of 200, the next window: plain                            -0.2180 (0.0666), below 0 in 5 of 6
  windows of 200, the next window: its own growth held fixed (b)    -0.4987 (0.0306), below 0 in 6 of 6
  windows of 200, 2 windows on: plain                               -0.5132 (0.0417), below 0 in 6 of 6
  windows of 200, 2 windows on: its own growth held fixed (b)       -0.5100 (0.0309), below 0 in 6 of 6
  same time, windows of 200 (did not hold in the test)              +0.5015 (0.0348), below 0 in 0 of 6
  scaled weights, the next window: b (coherence) -0.54, c (growth) +0.62 (averages over the starts)
  same time, read by the test's rule: more coherence, faster growth
  long-run rate, t = 3,000 to 6,000: +0.0146 (0.0003); Test 16: +0.0147 (0.0006); difference -0.0001 (0.0007): agree
  coherence |z| over t = 2,000 to 6,000 (average over the starts): 0.0260; starts 7 to 12 0.0251, 1 to 6 0.0270

The main measures on every set, by the same code (error bars: the larger of the two, as for the decision)
                                    new starts (13 to 18)               starts 7 to 12                      starts 1 to 6                       
  conditional (primary)             -0.1044 (0.0239), 6 of 6 below 0    -0.1610 (0.0309), 6 of 6 below 0    -0.2223 (0.0423), 6 of 6 below 0    
  unadjusted                        -0.1927 (0.0205), 6 of 6 below 0    -0.2061 (0.0205), 6 of 6 below 0    -0.2476 (0.0393), 6 of 6 below 0    
  same time, windows of 200         +0.5015 (0.0365), 0 of 6 below 0    +0.2763 (0.1158), 1 of 6 below 0    +0.3162 (0.0772), 0 of 6 below 0    
  Starts 7 to 12: the test's two measures were fixed before that run; the conditional one was chosen after it (exploratory). Starts 1 to 6: where the patterns were first noticed.

The link at chosen lags L (time units), averaged over the starts: unadjusted / conditional
                           5              50             100             200             300             400             500
  new         +0.403/+0.300   +0.193/+0.117   +0.035/-0.008   -0.083/-0.032   -0.167/-0.066   -0.224/-0.126   -0.221/-0.137 
  7 to 12     +0.281/+0.235   +0.112/+0.059   -0.037/-0.079   -0.195/-0.200   -0.198/-0.162   -0.210/-0.142   -0.214/-0.146 
  1 to 6      +0.291/+0.213   +0.115/+0.033   -0.007/-0.064   -0.179/-0.189   -0.224/-0.201   -0.264/-0.235   -0.308/-0.259 

On record and unchanged: coherence_growth.txt (the test, starts 7 to 12: same time +0.2763 (0.1044), no clear link; later -0.2061 (0.0186)) and earlier_growth_check.txt (exploratory: -0.1610 (0.0279) on starts 7 to 12, -0.2223 (0.0389) on 1 to 6).
A toy model: this shows structure in it, not proof about anything beyond it. A correlation, not a cause; no claim of damping.