diff --git a/baselines/hand-authored.json b/baselines/hand-authored.json
new file mode 100644
index 0000000..f629622
--- /dev/null
+++ b/baselines/hand-authored.json
@@ -0,0 +1,48 @@
+{
+  "name": "hand-authored",
+  "bot": "/work/target/release/sds-bot --weights /work/weights/hand-authored.json",
+  "scenario": "suite:suite:bfe96866022f",
+  "games": 24,
+  "decided": 19,
+  "wins": 4,
+  "outcomes": {
+    "victory": 19,
+    "round-limit": 3,
+    "stalled": 2
+  },
+  "bv_left": 0.21960347553646897,
+  "bv_left_opponent": 0.5672029341088792,
+  "bv_destroyed_percent": 70,
+  "max_rounds": 40,
+  "opponent": "princess",
+  "epoch": 3,
+  "commit": "5639825",
+  "recorded": "2026-08-20T01:00:52Z",
+  "notes": "hand-authored weights vs Princess, epoch 3, with imitation corpus",
+  "seeds": [
+    1,
+    2,
+    3,
+    4,
+    5,
+    6,
+    7,
+    8,
+    9,
+    10,
+    11,
+    12,
+    13,
+    14,
+    15,
+    16,
+    17,
+    18,
+    19,
+    20,
+    21,
+    22,
+    23,
+    24
+  ]
+}
diff --git a/crates/sds-core/src/features/mod.rs b/crates/sds-core/src/features/mod.rs
index fd690eb..19e96bf 100644
--- a/crates/sds-core/src/features/mod.rs
+++ b/crates/sds-core/src/features/mod.rs
@@ -309,14 +309,19 @@ impl Weights {
             // the enemy's guns costs about what the same share of damage dealt
             // is worth, and the bot has been paying it without a term for it.
             .with::<positional::Exposure>(-3.0)
-            // Being seen at all is the precondition for being shot. Small next
-            // to exposure, which prices *whose* guns are doing the seeing.
-            .with::<positional::LosIn>(-1.0)
+            // Being seen at all is the precondition for being shot. Equal and
+            // opposite to `los_out` on purpose: MegaMek's trace is only
+            // asymmetric when the two units differ in height or elevation, so
+            // on flat ground the pair cancels and `exposure` prices the danger
+            // on its own. Where a ridge does make them differ, this pair is the
+            // only term that says so.
+            .with::<positional::LosIn>(-1.5)
             // Terrain that spoils their aim is worth a good share of the damage
             // it prevents, and unlike range it costs us nothing to stand in.
             .with::<positional::CoverQuality>(2.0)
             // We cannot shoot what we cannot see, and the projected volley a
-            // move is scored on assumes a line that may not exist.
+            // move is scored on assumes a line that may not exist. See
+            // `los_in` for why the two are the same size.
             .with::<positional::LosOut>(1.5)
             // Standing where our own guns work is the largest single thing a
             // move can buy, so it outranks every cost on this list.
diff --git a/docs/HIERARCHY.md b/docs/HIERARCHY.md
index a1ac741..da11fe8 100644
--- a/docs/HIERARCHY.md
+++ b/docs/HIERARCHY.md
@@ -108,10 +108,11 @@ self-report, so a unit that is wrong about itself is never contradicted.
 `sds_core::features` is the replacement. A feature is a named number computed
 from the observation, and it declares how it is normalised — `bounded`, `rank`
 or `local` — in the type system, so an unnormalised feature does not compile and
-a `local` one cannot be given a learned weight. Firing, target, latch and
-force-level features exist and are tested; the positional group waits on line of
-sight. Nothing in the bot reads them yet, and `serves` is still what decides a
-match.
+a `local` one cannot be given a learned weight. Firing, target, latch,
+force-level and positional features exist and are tested. The positional group
+runs off `sds_core::los` through one survey per movement candidate, which is
+what lets a hex be scored on what could shoot it and not only on what it could
+shoot.
 
 ## What is not done yet
 
diff --git a/docs/PERFORMANCE.md b/docs/PERFORMANCE.md
index 86409a2..f4ef806 100644
--- a/docs/PERFORMANCE.md
+++ b/docs/PERFORMANCE.md
@@ -45,9 +45,12 @@ host: still several times the bot's whole budget.
 
 ## What will dominate once the thinking lands
 
-**Line of sight**, and it turned out cheaper than feared. `cover_quality`,
-`arc_exposure`, `crossfire`, `break_los` and a terrain-aware threat map all need
-it, and it is O(hexes along the line) per pair: ~1280 queries a round at 8v8.
+**Line of sight**, and it turned out cheaper than feared. `exposure`,
+`cover_quality`, `los_in`, `los_out`, `rear_arc_gain` and a terrain-aware threat
+map all need it, and it is O(hexes along the line) per pair: ~1280 queries a
+round at 8v8. The nine positional features that shipped ask through one
+`Posture::survey` per candidate, so the count is two traces per enemy per
+candidate however many features read the result.
 This file predicted 5-20 us each and 6-25 ms/round. Measured, one query is
 **0.7-1.5 us** and a whole round is **0.3-0.7 ms** — better than a fortieth of
 the guess even taking the high numbers, and about a tenth of what the host
diff --git a/weights/hand-authored.json b/weights/hand-authored.json
index f58c74a..bf154a1 100644
--- a/weights/hand-authored.json
+++ b/weights/hand-authored.json
@@ -13,7 +13,7 @@
     "has_been_fired_upon": 0.5,
     "heat_incurred": -2.0,
     "honour_broken": 1.0,
-    "los_in": -1.0,
+    "los_in": -1.5,
     "los_out": 1.5,
     "overkill": -3.0,
     "p_kill": 8.0,