diff --git a/plan/training.md b/plan/training.md
index 5c1e32d..b9ecac1 100644
--- a/plan/training.md
+++ b/plan/training.md
@@ -547,6 +547,48 @@ that row order does not move the answer at all.
 
 `sds corpus save` scans the same way, for the same reason.
 
+## The displaced rows move the fit more than the damage model does
+
+`--drop-displaced` leaves out decisions where the weights' first choice was
+taken by another unit before the force reconciled, so `chosen` is what was left
+rather than what was preferred. This epic already records that dropping them
+moved the offence-to-defence balance sharply toward defence on the 300-match
+corpus. **It reproduces on an independent corpus, recorded three days later
+against a corrected damage model.**
+
+| fit | corpus | offence:defence |
+|---|---|---|
+| `fitted-r11` | 300 matches, old damage model | 2.72:1 |
+| `fitted-r12-nodisp` | same corpus, displaced dropped | **0.59:1** |
+| `fitted-r13` | 240 matches, corrected grouping | 1.84:1 |
+| `fitted-r13-nodisp` | same corpus, displaced dropped | **0.82:1** |
+
+Ratios on `compare-weights.py`'s definition - only terms still pointing their
+family's way, wrong-way terms named and excluded - which is not the computation
+that produced the 1.69:1 and 0.40:1 recorded earlier in this epic. Read the
+columns against each other, not across.
+
+**The comparison worth making is which change moves more.** The corrected damage
+model took the balance from 2.72:1 to 1.84:1. Dropping the displaced rows took
+it from 2.72:1 to 0.59:1 on one corpus and 1.84:1 to 0.82:1 on the other. **A
+bookkeeping decision about which rows are evidence outweighs the damage model
+correction, on both corpora, by roughly a factor of two.**
+
+One concrete instance of the same thing: `incoming_damage` comes out **+0.0087**
+in `fitted-r13`, which is the wrong sign and is warned about - and **-0.0336**
+in `fitted-r13-nodisp`, which is the right one and is not. The rows where the
+bot did not get the hex it asked for are, on their own, enough to make defence
+price backwards.
+
+- [ ] This is now two corpora agreeing. `--drop-displaced` should stop being a
+      flag somebody remembers and become either the default or a documented
+      refusal to fit without a decision. A 20% subset of rows that inverts the
+      central balance of the vector is not a nuisance correction
+- [ ] The two arms still are not rankable - `r_squared` 0.2577 against 0.2408
+      and agreement 0.4830 against 0.4426 are measured over 3.85M rows against
+      2.79M. The direction of the weight change is the finding; the fit
+      statistics cannot say which vector is better
+
 ## For jmm: which features should declare an expected sign
 
 `EXPECTED_SIGNS` covers the **firing family only** - `expected_damage`,