diff --git a/plan/candidates.md b/plan/candidates.md
index 2e270f6..7f4eacf 100644
--- a/plan/candidates.md
+++ b/plan/candidates.md
@@ -793,13 +793,18 @@ vector is a feature change with its own control run, and has not been made.
 
 **Rollout**
 
-- [ ] Behind a flag, old generator default, so one match can produce both sets
-      and diff them. Test: with the flag off, behaviour is unchanged
+- [x] Behind a flag, old generator default, so one match can produce both sets
+      and diff them. Test: with the flag off, behaviour is unchanged.
+      `SDS_CANDIDATES`, `sds one --candidates`, and
+      `unit::tests::flag_off_is_the_four_verb_menu`
 - [ ] **Keep the four verbs only as a transitional instrument, and count how
       often they win.** They are not a permanent floor: if the scorer keeps
       preferring them, either the new states are bad or the scorer cannot tell
       them apart, and both are findings. Remove them once a run shows the
-      generated states winning, and record the share at which they were removed
+      generated states winning, and record the share at which they were removed.
+      Counted: `PlanTally::verb_share`, printed every round. **First real match,
+      `mirror-lance` seed 1: the four verbs won 26 of 31 movement decisions, 84%.**
+      The verbs stay until that turns over
 - [ ] `examples/pathfind.rs`: states/sec and `L x M` pairs/sec, open terrain
       against heavy forest, recorded in [PERFORMANCE.md](../docs/PERFORMANCE.md)
       in its own style
@@ -808,6 +813,34 @@ vector is a feature change with its own control run, and has not been made.
       micro-benchmark, deliberately, because a 24-game bench costs twenty
       minutes and carries +/- 25 points of noise
 
+### First run in a match
+
+`mirror-lance` seed 1, `sds one --sds-seat North --candidates`, hand-authored
+weights, against Princess. It completed: victory on round 9, 60 of 66 decisions
+answered, 0 defaulted, 6 illegal. Princess won.
+
+- **Time is not the problem.** A whole planning cycle - every unit of a lance
+  swept, scored, surfaced, and the lance reconciled - ran in 2 to 8 ms, mean 4.
+  The offline 79 ms was six views against enemy *reachable sets*; in a match
+  every enemy is one `Presence`, so `M = 1` per enemy and the sweep is `L x N`.
+  Largest single unit seen: 106 states, 848 exchanges.
+- **The menu is capped and the cap binds.** 5 to 16 candidates a unit, 11.5 mean.
+- **The four verbs won 26 of 31 movement decisions, 84%.** That is the number
+  the removal step is waiting on. Either the surfaced states are worse than a
+  straight line at the nearest enemy, or the weight set cannot tell them apart -
+  the weights were fitted on menus that never contained a generated state, which
+  makes the second the thing to rule out first.
+- **Two of the five chosen generated orders were illegal**, against four of the
+  twenty-six chosen four-verb orders. The route encoder is exact - see
+  `candidates::tests::a_route_becomes_a_step_list`, which checks every reached
+  stand - so the disagreement is between the movement model and MegaMek, not in
+  the translation.
+- **Both illegal generated orders were two different units sent to the same
+  hex**, `(11, 2)`, in the same round. The sweep blocks on where units *are* and
+  the force does not deconflict where they are *going*. The four verbs hid this
+  because two of them rarely name one hex.
+
+
 ## Heat: reported, not budgeted
 
 `VolleyOutcome` reported expected damage, `p_kill`, `p_mission_kill`, expected
diff --git a/sds/cli.py b/sds/cli.py
index f77802d..9a61c27 100644
--- a/sds/cli.py
+++ b/sds/cli.py
@@ -89,6 +89,22 @@ def _play(spec: MatchSpec, out_dir: Path, roles: dict[str, str]) -> dict:
     return result
 
 
+def _bot_env(args: argparse.Namespace) -> dict[str, str]:
+    """What the bot inherits, from the flags that switch a bot behaviour on.
+
+    Both are off unless asked for. `SDS_IMITATE` rebuilds a menu for every
+    enemy move, which is a second bot's worth of thinking; `SDS_CANDIDATES`
+    replaces the four hardcoded verbs with the reachable-state sweep, which is
+    a change to play and must be said out loud on any number it produces.
+    """
+    env: dict[str, str] = {}
+    if getattr(args, "imitate", False):
+        env["SDS_IMITATE"] = "1"
+    if getattr(args, "candidates", False):
+        env["SDS_CANDIDATES"] = "1"
+    return env
+
+
 def cmd_one(args: argparse.Namespace) -> int:
     scenario = Path(args.scenario).resolve()
     factions = _factions(scenario)
@@ -109,6 +125,7 @@ def cmd_one(args: argparse.Namespace) -> int:
         timeout_ms=args.timeout_ms,
         stall_seconds=args.stall_seconds,
         bv_destroyed_percent=args.bv_destroyed_percent,
+        env=_bot_env(args),
     )
     result = _play(spec, _run_dir(Path(args.out), "one"), roles)
     print(json.dumps(result, indent=2))
@@ -202,7 +219,7 @@ def _bench(args: argparse.Namespace, label: str, sds_bot: str | None) -> int:
                     # setting when one seat is not ours: against Princess this
                     # is the whole point, and in self-play it would only
                     # reconstruct moves the bot already logged itself.
-                    env={"SDS_IMITATE": "1"} if getattr(args, "imitate", False) else {},
+                    env=_bot_env(args),
                 ),
                 roles,
             )
@@ -613,6 +630,11 @@ def main(argv: list[str] | None = None) -> int:
     one.add_argument(
         "--sds-seat", default=None, help="faction the bot plays; omit for Princess on both sides"
     )
+    one.add_argument(
+        "--candidates",
+        action="store_true",
+        help="propose from the reachable-state sweep as well as the four verbs (SDS_CANDIDATES)",
+    )
     one.set_defaults(func=cmd_one)
 
     bench = sub.add_parser("bench", help="play many matches and aggregate")