slowdini · slowdini · Aug 24, 2026 · Aug 24, 2026
diff --git a/README.md b/README.md
@@ -110,6 +110,10 @@ eval-magic run --mode revision
 
 The command help and generated runbook describe baseline selection and the rest of the workflow.
 
+An eval can treat coordinated skills as one treatment by setting `skill_name` to an ordered list.
+Pass one listed member with `--skill`; it remains the eval owner and supplies fixtures. See
+`eval-magic docs isolation` for the complete configuration, Mode A/B behavior, and provenance.
+
 ## How it works
 
 Each eval case runs once per condition and repetition in its own clean Git repository. The two arms

diff --git a/docs/guides/isolation.md b/docs/guides/isolation.md
@@ -144,40 +144,63 @@ then use `isolates_live_sources` to record the operator assertion.
 dispatch's setting-source selection. A plugin can appear there and remain absent from the dispatch,
 or the reverse. Use the dispatch's init event.
 
-## The skill under test is a copy
-
-Every skill an eval stages is copied into the eval home before any dispatch runs, and
-each condition stages from that copy. Nothing the agent can reach is read from your own
-skill directory, so editing a skill mid-campaign cannot change what a prepared iteration
-measures.
+## Treatment skills are copies
+
+`skill_name` in the `evals/evals.json` file accepts either one skill name or an ordered,
+non-empty list:
+
+```json
+{
+  "skill_name": ["review-workflow", "review-verification"],
+  "evals": [
+    {
+      "id": "review-change",
+      "prompt": "Review this change.",
+      "expected_output": "A prioritized review."
+    }
+  ]
+}
+```
 
-The copy is the working tree as it sits on disk, not a checkout of a commit —
-evaluating an uncommitted revision is the ordinary case, and in a `--mode revision` run
-the edit under test is uncommitted by definition. What the run measured is recorded rather than inferred, in
-`conditions.json`, each `run.json`, `benchmark.json`, and the `BASELINE.md` written by
+With a list, `--skill` selects the eval owner: the member whose `evals/` directory supplies the
+definitions and fixtures, and whose name owns the workspace and promotion destination. The owner
+must appear in the list. `--stage-name` is unavailable because one override cannot name several
+staged skills.
+
+Every treatment member is copied into the eval home before any dispatch runs, and each condition
+stages from those copies. Mode A stages all treatment members in `with_skill` and none in
+`without_skill`. Other siblings from `--skill-dir` remain ambient in both arms. Mode B snapshots
+and stages the complete set in both revisions. A scalar `skill_name` retains the existing
+single-skill paths and artifacts.
+
+Each copy is the working tree as it sits on disk, not a checkout of a commit. Evaluating an
+uncommitted revision is the ordinary case, and in a `--mode revision` run the edit under test is
+uncommitted by definition. What the run measured is recorded rather than inferred in
+`conditions.json`, each `run.json`, `benchmark.json`, and the `BASELINE.md` file written by
 `promote-baseline`:
 
 ```sh
 jq '.skill_source' conditions.json
 ```
 
-`dirty: true` means the recorded revision alone does not identify what ran. Commit the
-skill before a run whose result you intend to publish.
+`dirty: true` means the recorded revision alone does not identify what ran. Commit the treatment
+skills before a run whose result you intend to publish.
 
-Sibling skills staged by `--skill-dir` are copied the same way, and the roster is
-captured once when the run resolves. The `siblings` field names exactly what every
-environment received.
+Ambient skills staged by `--skill-dir` are copied the same way, and the roster is captured once
+when the run resolves. For a multi-skill treatment, `skill_source.eval_owner` names the owner and
+`skill_source.skills` records every treatment member's resolved source and revision. The
+`siblings` field, when present, names ambient skills staged in both arms.
 
 The eval home sits outside the skill's own repository: under `$XDG_DATA_HOME/eval-magic`
 (or `~/.local/share/eval-magic`), in a directory named for the skill directory it serves.
 `run` prints the path it chose, and every command it suggests carries `--workspace-dir`,
 so there is nothing to remember. `EVAL_MAGIC_WORKSPACE_DIR` moves the default;
 `--workspace-dir` overrides both.
 
-Copying does not remove the live directory from the machine, so a dispatch can still read
-it by absolute path. `detect-stray-writes` reports that as a live-source read, and
-`aggregate` carries it into `validity_warnings` for the same reason a discoverable
-plugin copy is carried there: the arm may not be comparing what it claims to.
+Copying does not remove the live directories from the machine, so a dispatch can still read one by
+absolute path. `detect-stray-writes` checks every treatment source and reports that as a live-source
+read. `aggregate` carries it into `validity_warnings` for the same reason a discoverable plugin
+copy is carried there: the arm may not be comparing what it claims to.
 
 ## The task repository is a separate boundary
 

diff --git a/docs/guides/judging.md b/docs/guides/judging.md
@@ -73,7 +73,15 @@ An authored `llm_judge` assertion can request several independent verdicts for t
 
 Use `run --judge-samples N` to set a campaign-wide default. An assertion's `samples` field takes
 precedence over that default. The effective count must be at least one. The framework-injected
-`__skill_invoked` meta-check is not substantive grading and remains single-shot.
+`__skill_invoked` checks are not substantive grading and remain single-shot per treatment member.
+
+For a multi-skill treatment, deterministic transcript grading checks every staged slug separately.
+The response files use `__skill_invoked__skill-N.json`, and each meta result names its
+`skill_name`, so partial and complete invocation are distinguishable. Harness descriptors supply
+the tool and argument signature; a harness without deterministic invocation events receives one
+LLM fallback task per member. The suite-level `meta_summary.skill_invoked` value is true when any
+treatment member was invoked. The `benchmark.json` file retains the suite rate and adds per-skill
+counts and rates. A scalar treatment keeps the `__skill_invoked.json` filename and artifact shape.
 
 Each sample is a separate judge task and response, but every sample for a run receives the exact
 same bounded `judge-evidence.md`. The agent is not rerun, and eval-magic does not rebuild or expand

diff --git a/schema/benchmark.schema.json b/schema/benchmark.schema.json
@@ -166,9 +166,36 @@
           "type": "array",
           "items": { "type": "string" },
           "description": "Sibling skills staged alongside the skill under test, as the roster was captured at resolution."
+        },
+        "eval_owner": {
+          "type": "string",
+          "description": "CLI-selected skill that owns the eval definitions and workspace namespace. Present for multi-skill treatments."
+        },
+        "skills": {
+          "type": "array",
+          "minItems": 1,
+          "items": { "$ref": "#/definitions/skillSourceEntry" },
+          "description": "Ordered resolved source and revision for every multi-skill treatment member."
         }
       }
     },
+    "skillSourceEntry": {
+      "type": "object",
+      "required": ["name", "kind", "source", "branch"],
+      "additionalProperties": false,
+      "properties": {
+        "name": { "type": "string" },
+        "kind": { "type": "string", "enum": ["git", "path"] },
+        "source": { "type": "string" },
+        "resolved_path": { "type": "string" },
+        "ref": { "type": "string" },
+        "revision": { "type": "string" },
+        "origin_url": { "type": "string" },
+        "branch": { "type": "string" },
+        "host_local": { "type": "boolean" },
+        "dirty": { "type": "boolean" }
+      }
+    },
     "assertionCount": {
       "type": "object",
       "required": ["passed", "n"],
@@ -234,7 +261,20 @@
         "duration_ms": { "$ref": "#/definitions/stats" },
         "total_tokens": { "$ref": "#/definitions/stats" },
         "skill_invocation_n": { "type": "integer" },
-        "skill_invocation_rate": { "type": ["number", "null"] }
+        "skill_invocation_rate": { "type": ["number", "null"] },
+        "skill_invocations": {
+          "type": "object",
+          "description": "Per-treatment-member invocation counts and rates for multi-skill evals.",
+          "additionalProperties": {
+            "type": "object",
+            "required": ["n", "rate"],
+            "additionalProperties": false,
+            "properties": {
+              "n": { "type": "integer", "minimum": 0 },
+              "rate": { "type": "number", "minimum": 0, "maximum": 1 }
+            }
+          }
+        }
       }
     },
     "diffScopeRun": {

diff --git a/schema/evals.schema.json b/schema/evals.schema.json
@@ -8,8 +8,16 @@
   "additionalProperties": false,
   "properties": {
     "skill_name": {
-      "type": "string",
-      "description": "Name of the skill being evaluated. Should match the skill directory name."
+      "description": "Name of the skill being evaluated, or an ordered non-empty set of coordinated skills. The CLI-selected eval owner must be a member of a set.",
+      "oneOf": [
+        { "type": "string", "minLength": 1 },
+        {
+          "type": "array",
+          "minItems": 1,
+          "uniqueItems": true,
+          "items": { "type": "string", "minLength": 1 }
+        }
+      ]
     },
     "codebase": {
       "$ref": "#/definitions/codebase",

diff --git a/schema/grading.schema.json b/schema/grading.schema.json
@@ -17,7 +17,7 @@
     "meta_results": {
       "type": "array",
       "description": "Framework-injected meta-assertions (e.g. skill-invocation check). Reserved id prefix: __ (double underscore). Tracked separately from substantive assertion_results so they do not pollute the skill effectiveness pass_rate.",
-      "items": { "$ref": "#/definitions/binaryAssertionResult" }
+      "items": { "$ref": "#/definitions/metaResult" }
     },
     "meta_summary": {
       "type": "object",
@@ -27,13 +27,29 @@
         "failed": { "type": "integer", "minimum": 0 },
         "total": { "type": "integer", "minimum": 0 },
         "skill_invoked": {
-          "description": "True when the skill-invocation meta-check passed; false when the judge found no evidence the skill influenced behavior; null when no skill was loaded for this run.",
+          "description": "True when at least one treatment member's invocation meta-check passed; false when none did; null when no treatment skill was loaded for this run.",
           "type": ["boolean", "null"]
         }
       }
     }
   },
   "definitions": {
+    "metaResult": {
+      "type": "object",
+      "required": ["id", "passed", "evidence"],
+      "additionalProperties": false,
+      "properties": {
+        "id": { "type": "string" },
+        "skill_name": { "type": "string", "description": "Treatment member checked; absent for scalar legacy artifacts." },
+        "passed": { "type": "boolean" },
+        "evidence": { "type": "string" },
+        "confidence": { "type": "number", "minimum": 0, "maximum": 1 },
+        "grader": {
+          "type": "string",
+          "enum": ["transcript_check", "llm_judge", "command_check", "diff_scope"]
+        }
+      }
+    },
     "assertionResult": {
       "oneOf": [
         { "$ref": "#/definitions/binaryAssertionResult" },

diff --git a/schema/judge-tasks.schema.json b/schema/judge-tasks.schema.json
@@ -54,6 +54,10 @@
           "description": "1-based run index within a multi-run (eval, condition) cell; absent for single-run cells."
         },
         "assertion_id": { "type": "string" },
+        "skill_name": {
+          "type": "string",
+          "description": "Treatment member checked by a multi-skill invocation meta task."
+        },
         "sample_index": {
           "type": "integer",
           "minimum": 1,

diff --git a/schema/run-record.schema.json b/schema/run-record.schema.json
@@ -27,6 +27,11 @@
       "type": ["string", "null"],
       "description": "Absolute path to the SKILL.md the subagent could load, or null if no skill was provided (without_skill condition)."
     },
+    "skills": {
+      "type": "array",
+      "description": "Ordered treatment roster for list-authored evals. Empty in the control arm and absent in scalar legacy records.",
+      "items": { "$ref": "#/definitions/conditionSkill" }
+    },
     "prompt": {
       "type": "string",
       "description": "The user prompt as dispatched to the subagent."
@@ -95,6 +100,16 @@
     }
   },
   "definitions": {
+    "conditionSkill": {
+      "type": "object",
+      "required": ["name", "skill_path", "staged_skill_slug"],
+      "additionalProperties": false,
+      "properties": {
+        "name": { "type": "string" },
+        "skill_path": { "type": "string" },
+        "staged_skill_slug": { "type": ["string", "null"] }
+      }
+    },
     "responderOutcome": {
       "type": "object",
       "required": ["ending"],
@@ -321,9 +336,36 @@
           "type": "array",
           "items": { "type": "string" },
           "description": "Sibling skills staged alongside the skill under test, as the roster was captured at resolution."
+        },
+        "eval_owner": {
+          "type": "string",
+          "description": "CLI-selected skill that owns the eval definitions and workspace namespace. Present for multi-skill treatments."
+        },
+        "skills": {
+          "type": "array",
+          "minItems": 1,
+          "items": { "$ref": "#/definitions/skillSourceEntry" },
+          "description": "Ordered resolved source and revision for every multi-skill treatment member."
         }
       }
     },
+    "skillSourceEntry": {
+      "type": "object",
+      "required": ["name", "kind", "source", "branch"],
+      "additionalProperties": false,
+      "properties": {
+        "name": { "type": "string" },
+        "kind": { "type": "string", "enum": ["git", "path"] },
+        "source": { "type": "string" },
+        "resolved_path": { "type": "string" },
+        "ref": { "type": "string" },
+        "revision": { "type": "string" },
+        "origin_url": { "type": "string" },
+        "branch": { "type": "string" },
+        "host_local": { "type": "boolean" },
+        "dirty": { "type": "boolean" }
+      }
+    },
     "conversationTool": {
       "type": "object",
       "required": ["type", "ordinal", "round", "name"],

diff --git a/src/adapters/skill_shadow/grouping.rs b/src/adapters/skill_shadow/grouping.rs
@@ -48,10 +48,10 @@ impl PluginShadowReport {
         subject_skill_name: &str,
         expected_cells: &[(String, String)],
     ) -> Self {
-        Self::from_observed_sources_with_class(
+        Self::from_observed_sources_for_subjects_with_class(
             config_dir,
             sources,
-            subject_skill_name,
+            &[subject_skill_name],
             expected_cells,
             ShadowFindingClass::OperatorEnvironment,
         )
@@ -63,6 +63,22 @@ impl PluginShadowReport {
         subject_skill_name: &str,
         expected_cells: &[(String, String)],
         class: ShadowFindingClass,
+    ) -> Self {
+        Self::from_observed_sources_for_subjects_with_class(
+            config_dir,
+            sources,
+            &[subject_skill_name],
+            expected_cells,
+            class,
+        )
+    }
+
+    pub(crate) fn from_observed_sources_for_subjects_with_class(
+        config_dir: impl Into<String>,
+        sources: Vec<ShadowSource>,
+        subject_skill_names: &[&str],
+        expected_cells: &[(String, String)],
+        class: ShadowFindingClass,
     ) -> Self {
         let mut merged = Vec::<ShadowSource>::new();
         for mut source in sources {
@@ -88,7 +104,7 @@ impl PluginShadowReport {
         }
         for finding in &mut report.findings {
             finding.class = class;
-            finding.role = if finding.skill_name == subject_skill_name {
+            finding.role = if subject_skill_names.contains(&finding.skill_name.as_str()) {
                 ShadowSkillRole::Subject
             } else {
                 ShadowSkillRole::Sibling