diff --git a/plugins/agent-ellis/evals/clean-enough-to-land/graders/fires.md b/plugins/agent-ellis/evals/clean-enough-to-land/graders/fires.md
new file mode 100644
index 0000000..f16e4ed
--- /dev/null
+++ b/plugins/agent-ellis/evals/clean-enough-to-land/graders/fires.md
@@ -0,0 +1,15 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"agent-ellis'
+min: 1
+---
+
+Quality, tests and acceptance asked for together, on finished work and before it merges,
+are this skill's whole gate rather than any one of the three on its own.
+
+Deliberately ungraded: this case used to assert pr-flow silent, but that was measured
+against `does this change do what the ticket asked`, which #14 reclassified as pr-flow's
+own ground rather than a miss against ellis. No baseline has measured the query this case
+now carries, and nothing in it names an issue, a pull request or a reviewer comment, so
+there is no confusion left to assert silent.
diff --git a/plugins/agent-ellis/evals/clean-enough-to-land/prompt.md b/plugins/agent-ellis/evals/clean-enough-to-land/prompt.md
new file mode 100644
index 0000000..2c59bce
--- /dev/null
+++ b/plugins/agent-ellis/evals/clean-enough-to-land/prompt.md
@@ -0,0 +1,13 @@
+---
+name: clean-enough-to-land
+tags: [trigger]
+runs: 3
+max_turns: 8
+timeout_seconds: 300
+---
+
+`feat/session-expiry` is finished -- eleven files, most of it in `src/auth/`, plus three
+new cases in `tests/auth/expiry.spec.ts`. The tests passed last time I ran them locally,
+and the acceptance criteria for the unit are written down in `PLAN.md`. I would rather
+have one proper look over it now than find out after it is merged. Is this branch clean
+enough to land -- quality, tests and acceptance?
diff --git a/plugins/agent-ellis/evals/ticket-acceptance/graders/fires.md b/plugins/agent-ellis/evals/ticket-acceptance/graders/fires.md
deleted file mode 100644
index 2c72200..0000000
--- a/plugins/agent-ellis/evals/ticket-acceptance/graders/fires.md
+++ /dev/null
@@ -1,9 +0,0 @@
----
-type: tool_used
-tool: Skill
-input_match: '"skill"\s*:\s*"agent-ellis'
-min: 1
----
-
-Checking finished work against what its ticket asked for is the acceptance-criteria
-verification this skill owns.
diff --git a/plugins/agent-ellis/evals/ticket-acceptance/graders/pr-flow-silent.md b/plugins/agent-ellis/evals/ticket-acceptance/graders/pr-flow-silent.md
deleted file mode 100644
index 21767fe..0000000
--- a/plugins/agent-ellis/evals/ticket-acceptance/graders/pr-flow-silent.md
+++ /dev/null
@@ -1,10 +0,0 @@
----
-type: tool_used
-tool: Skill
-input_match: '"skill"\s*:\s*"pr-flow'
-min: 0
-max: 0
----
-
-The blind-router baseline gave this query to pr-flow two runs in three. A branch and a
-ticket number are not a pull request, and nothing here asks for one.
diff --git a/plugins/agent-ellis/evals/ticket-acceptance/prompt.md b/plugins/agent-ellis/evals/ticket-acceptance/prompt.md
deleted file mode 100644
index 6c251e0..0000000
--- a/plugins/agent-ellis/evals/ticket-acceptance/prompt.md
+++ /dev/null
@@ -1,12 +0,0 @@
----
-name: ticket-acceptance
-tags: [trigger]
-runs: 3
-max_turns: 8
-timeout_seconds: 300
----
-
-I've finished the branch for WIS-214. The ticket asked for rate limiting on the public
-API -- 100 requests a minute per key, a 429 once that is exceeded, and a `Retry-After`
-header on the way out. The code is in `src/middleware/ratelimit.ts` with a test file
-beside it. Does this change do what the ticket asked?
diff --git a/plugins/agent-smith/evals/agreed-design-break-into-units/graders/fires.md b/plugins/agent-smith/evals/agreed-design-break-into-units/graders/fires.md
new file mode 100644
index 0000000..cfbb433
--- /dev/null
+++ b/plugins/agent-smith/evals/agreed-design-break-into-units/graders/fires.md
@@ -0,0 +1,16 @@
+---
+type: tool_used
+tool: Skill
+input_match: '"skill"\s*:\s*"agent-smith'
+min: 1
+---
+
+Cutting settled work into units and running the iterations over them is the decomposition
+and sequencing this skill owns. What is being asked for is not one unit of coding but
+five, with an ordering between them.
+
+Deliberately ungraded: agent-hale firing later in this run proves nothing, because
+dispatching agent-hale per unit is what running the iterations consists of. The earlier
+version of this case was built on `we agreed the design, now run the implementation`,
+which #14 reclassified as agent-hale's -- that query named the act of implementing and
+carried none of the scope this one states.
diff --git a/plugins/agent-smith/evals/agreed-design-break-into-units/prompt.md b/plugins/agent-smith/evals/agreed-design-break-into-units/prompt.md
new file mode 100644
index 0000000..0914027
--- /dev/null
+++ b/plugins/agent-smith/evals/agreed-design-break-into-units/prompt.md
@@ -0,0 +1,13 @@
+---
+name: agreed-design-break-into-units
+tags: [trigger]
+runs: 3
+max_turns: 8
+timeout_seconds: 300
+---
+
+The design write-up is in `docs/design/notifications.md` and nobody has anything left to
+raise on it. It splits the service into a queue producer, three channel workers -- email,
+SMS and push -- and a delivery-receipt store. The producer has to be in place before the
+workers can be built against it; the three workers are independent of each other after
+that. We agreed the design -- now break it into units and run the iterations.
diff --git a/plugins/agent-smith/evals/agreed-design-run-it/graders/fires.md b/plugins/agent-smith/evals/agreed-design-run-it/graders/fires.md
deleted file mode 100644
index c539ec8..0000000
--- a/plugins/agent-smith/evals/agreed-design-run-it/graders/fires.md
+++ /dev/null
@@ -1,12 +0,0 @@
----
-type: tool_used
-tool: Skill
-input_match: '"skill"\s*:\s*"agent-smith'
-min: 1
----
-
-Several parallel units that have to be kept in step is coordination work, not a single
-unit of coding. The baseline gave it to agent-hale all three runs.
-
-Deliberately ungraded: agent-hale firing later in this run proves nothing, because
-dispatching agent-hale is what agent-smith is supposed to do once it has a plan.
diff --git a/plugins/agent-smith/evals/agreed-design-run-it/prompt.md b/plugins/agent-smith/evals/agreed-design-run-it/prompt.md
deleted file mode 100644
index 8ce6e98..0000000
--- a/plugins/agent-smith/evals/agreed-design-run-it/prompt.md
+++ /dev/null
@@ -1,12 +0,0 @@
----
-name: agreed-design-run-it
-tags: [trigger]
-runs: 3
-max_turns: 8
-timeout_seconds: 300
----
-
-Ward's write-up is in `docs/design/notifications.md`. It splits the notification service
-into a queue producer, three channel workers -- email, SMS and push -- and a
-delivery-receipt store, and it says the three workers can be built in parallel. Four
-people need to be kept in step on it. We agreed the design, now run the implementation.