diff --git a/plan/candidates.md b/plan/candidates.md
index fd6370e..251f762 100644
--- a/plan/candidates.md
+++ b/plan/candidates.md
@@ -216,6 +216,37 @@ rather than a test, because wall-clock assertions are flaky under load.
       the number of states or survivors fails loudly and has to be explained
       rather than absorbed
 
+## Seeing what the estimator thinks
+
+Every inspection tool here prints scalars, and a scalar cannot be judged by
+somebody who knows BattleTech. A heatmap can: put it on a board and a player can
+say "nobody would stand there" in a second, which is a class of bug no test
+catches and no win rate localises.
+
+One sheet, two sides, nobody has moved. For each stand in our `L`, the estimator
+gives two numbers aggregated over the enemy's `M`: what we deal from it, and
+what we take at it. Colour the hex by both. A hex we can hurt them from and they
+cannot answer reads one way; a hex where both sides trade heavily reads another.
+That second kind is where a match is decided and it currently has no name in any
+output we produce.
+
+- [ ] Render the two channels per hex over one board, from the scenario
+      fixtures, with no MegaMek and no match. Encode total as lightness and the
+      offence-to-defence ratio as hue: an additive blend of two colours puts
+      "high in both" and "middling in both" on the lightness axis alone, which
+      is the one the eye compares worst
+- [ ] **The same board at three power-mean exponents**, since that is the
+      argument for the operator existing. `p -> -inf` is the enemy finding our
+      worst case and should pull the good hexes toward cover; `p = 1` is the
+      enemy moving without regard to us. The scenario test asserts this
+      relationship as a number - the render is the same claim made visible
+- [ ] At least three enemies on one side, so the `N` aggregation is not
+      degenerate. Against a single enemy, offence-takes-best and
+      defence-takes-sum are the same number, and the case worth seeing - good
+      against one, exposed to three - cannot occur
+- [ ] Test what is asserted rather than what is drawn: the hex ordering the
+      render produces, not its pixels. A colour is not a regression target
+
 ## Deferred
 
 **Jump** is a second pathfinding pass, not a variant of the first: a jump cannot