diff --git a/apps/cli/src/commands/results/serve.ts b/apps/cli/src/commands/results/serve.ts
index 121d9eac1..72371a7ab 100644
--- a/apps/cli/src/commands/results/serve.ts
+++ b/apps/cli/src/commands/results/serve.ts
@@ -1,5 +1,5 @@
 /**
- * `agentv studio` — starts the AgentV Studio server, a React SPA for
+ * `agentv studio` / `agentv serve` — starts the AgentV Studio server, a React SPA for
  * reviewing evaluation results.
  *
  * The server uses Hono for routing and @hono/node-server to listen.
@@ -114,6 +114,21 @@ export function loadResults(content: string): EvaluationResult[] {
   return results;
 }
 
+export function resolveDashboardMode(
+  projectCount: number,
+  options: { multi?: boolean; single?: boolean },
+): { isMultiProject: boolean; showMultiWarning: boolean } {
+  if (options.single === true) {
+    return { isMultiProject: false, showMultiWarning: options.multi === true };
+  }
+
+  if (options.multi === true) {
+    return { isMultiProject: true, showMultiWarning: true };
+  }
+
+  return { isMultiProject: projectCount > 1, showMultiWarning: false };
+}
+
 // ── Feedback persistence ─────────────────────────────────────────────────
 
 interface FeedbackReview {
@@ -669,12 +684,13 @@ async function handleTargets(c: C, { searchDir, agentvDir }: DataContext) {
 function handleConfig(
   c: C,
   { agentvDir, searchDir }: DataContext,
-  options?: { readOnly?: boolean },
+  options?: { readOnly?: boolean; multiProjectDashboard?: boolean },
 ) {
   return c.json({
     ...loadStudioConfig(agentvDir),
     read_only: options?.readOnly === true,
     project_name: path.basename(searchDir),
+    multi_project_dashboard: options?.multiProjectDashboard === true,
   });
 }
 
@@ -694,7 +710,7 @@ export function createApp(
   resultDir: string,
   cwd?: string,
   sourceFile?: string,
-  options?: { studioDir?: string; readOnly?: boolean },
+  options?: { studioDir?: string; readOnly?: boolean; multiProjectDashboard?: boolean },
 ): Hono {
   const searchDir = cwd ?? resultDir;
   const agentvDir = path.join(searchDir, '.agentv');
@@ -906,7 +922,12 @@ export function createApp(
 
   // ── Data routes (unscoped) ────────────────────────────────────────────
 
-  app.get('/api/config', (c) => handleConfig(c, defaultCtx, { readOnly }));
+  app.get('/api/config', (c) =>
+    handleConfig(c, defaultCtx, {
+      readOnly,
+      multiProjectDashboard: options?.multiProjectDashboard,
+    }),
+  );
   app.get('/api/remote/status', async (c) => c.json(await getRemoteResultsStatus(searchDir)));
   app.post('/api/remote/sync', async (c) => c.json(await syncRemoteResults(searchDir)));
   app.get('/api/runs', (c) => handleRuns(c, defaultCtx));
@@ -1006,7 +1027,12 @@ export function createApp(
   // Same handlers as above, with project-resolved DataContext via withProject.
 
   app.get('/api/projects/:projectId/config', (c) =>
-    withProject(c, (ctx, dataCtx) => handleConfig(ctx, dataCtx, { readOnly })),
+    withProject(c, (ctx, dataCtx) =>
+      handleConfig(ctx, dataCtx, {
+        readOnly,
+        multiProjectDashboard: options?.multiProjectDashboard,
+      }),
+    ),
   );
   app.get('/api/projects/:projectId/remote/status', (c) =>
     withProject(c, async (ctx, dataCtx) =>
@@ -1159,7 +1185,12 @@ export const resultsServeCommand = command({
     }),
     multi: flag({
       long: 'multi',
-      description: 'Launch in multi-project dashboard mode',
+      description:
+        'Launch in multi-project dashboard mode (deprecated; use auto-detect or --single)',
+    }),
+    single: flag({
+      long: 'single',
+      description: 'Force single-project dashboard mode',
     }),
     add: option({
       type: optional(string),
@@ -1181,7 +1212,7 @@ export const resultsServeCommand = command({
       description: 'Disable write operations and launch Studio in read-only leaderboard mode',
     }),
   },
-  handler: async ({ source, port, dir, multi, add, remove, discover, readOnly }) => {
+  handler: async ({ source, port, dir, multi, single, add, remove, discover, readOnly }) => {
     const cwd = dir ?? process.cwd();
     const listenPort = port ?? (process.env.PORT ? Number(process.env.PORT) : 3117);
 
@@ -1224,7 +1255,10 @@ export const resultsServeCommand = command({
 
     // ── Determine multi-project mode ────────────────────────────────
     const registry = loadProjectRegistry();
-    const isMultiProject = multi || registry.projects.length > 0;
+    const { isMultiProject, showMultiWarning } = resolveDashboardMode(registry.projects.length, {
+      multi,
+      single,
+    });
 
     try {
       let results: EvaluationResult[] = [];
@@ -1254,7 +1288,16 @@ export const resultsServeCommand = command({
 
       // Use the run directory for feedback storage (matches #764 behavior)
       const resultDir = sourceFile ? path.dirname(path.resolve(sourceFile)) : cwd;
-      const app = createApp(results, resultDir, cwd, sourceFile, { readOnly });
+      const app = createApp(results, resultDir, cwd, sourceFile, {
+        readOnly,
+        multiProjectDashboard: isMultiProject,
+      });
+
+      if (showMultiWarning) {
+        console.warn(
+          'Warning: --multi is deprecated. Studio now auto-detects multi-project mode when multiple projects are registered. Use --single to force the single-project view.',
+        );
+      }
 
       if (isMultiProject) {
         console.log(`Multi-project mode: ${registry.projects.length} project(s) registered`);
diff --git a/apps/cli/test/commands/results/serve.test.ts b/apps/cli/test/commands/results/serve.test.ts
index 4a4ea4085..e1f788b8e 100644
--- a/apps/cli/test/commands/results/serve.test.ts
+++ b/apps/cli/test/commands/results/serve.test.ts
@@ -3,7 +3,12 @@ import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync
 import { tmpdir } from 'node:os';
 import path from 'node:path';
 
-import { createApp, loadResults, resolveSourceFile } from '../../../src/commands/results/serve.js';
+import {
+  createApp,
+  loadResults,
+  resolveDashboardMode,
+  resolveSourceFile,
+} from '../../../src/commands/results/serve.js';
 
 // ── Sample JSONL content (snake_case, matching on-disk format) ──────────
 
@@ -94,6 +99,52 @@ describe('loadResults', () => {
   });
 });
 
+// ── resolveDashboardMode ───────────────────────────────────────────────
+
+describe('resolveDashboardMode', () => {
+  it('defaults to single-project mode when no projects are registered', () => {
+    expect(resolveDashboardMode(0, {})).toEqual({
+      isMultiProject: false,
+      showMultiWarning: false,
+    });
+  });
+
+  it('defaults to single-project mode when exactly one project is registered', () => {
+    expect(resolveDashboardMode(1, {})).toEqual({
+      isMultiProject: false,
+      showMultiWarning: false,
+    });
+  });
+
+  it('defaults to multi-project mode when multiple projects are registered', () => {
+    expect(resolveDashboardMode(2, {})).toEqual({
+      isMultiProject: true,
+      showMultiWarning: false,
+    });
+  });
+
+  it('forces multi-project mode with a deprecation warning when --multi is used', () => {
+    expect(resolveDashboardMode(1, { multi: true })).toEqual({
+      isMultiProject: true,
+      showMultiWarning: true,
+    });
+  });
+
+  it('forces single-project mode when --single is used', () => {
+    expect(resolveDashboardMode(3, { single: true })).toEqual({
+      isMultiProject: false,
+      showMultiWarning: false,
+    });
+  });
+
+  it('lets --single override --multi', () => {
+    expect(resolveDashboardMode(3, { multi: true, single: true })).toEqual({
+      isMultiProject: false,
+      showMultiWarning: true,
+    });
+  });
+});
+
 // ── Mock studio dist ─────────────────────────────────────────────────────
 
 const MOCK_STUDIO_HTML = `<!doctype html>
@@ -319,12 +370,17 @@ describe('serve app', () => {
       const app = createApp(results, tempDir, undefined, undefined, {
         studioDir,
         readOnly: true,
+        multiProjectDashboard: true,
       });
 
       const res = await app.request('/api/config');
       expect(res.status).toBe(200);
-      const data = (await res.json()) as { read_only?: boolean };
+      const data = (await res.json()) as {
+        read_only?: boolean;
+        multi_project_dashboard?: boolean;
+      };
       expect(data.read_only).toBe(true);
+      expect(data.multi_project_dashboard).toBe(true);
     });
   });
 
diff --git a/apps/studio/src/components/CompareTab.tsx b/apps/studio/src/components/CompareTab.tsx
index a448214d3..a633d70ea 100644
--- a/apps/studio/src/components/CompareTab.tsx
+++ b/apps/studio/src/components/CompareTab.tsx
@@ -67,37 +67,20 @@ export function CompareTab({ data, isLoading, isError, error }: CompareTabProps)
     cellMap.set(JSON.stringify([cell.experiment, cell.target]), cell);
   }
 
-  // Find best pass rate per row (target) for highlighting
-  const bestByTarget = new Map<string, number>();
-  const worstByTarget = new Map<string, number>();
-  for (const target of targets) {
-    let best = -1;
-    let worst = 2;
-    for (const experiment of experiments) {
-      const cell = cellMap.get(JSON.stringify([experiment, target]));
-      if (cell) {
-        if (cell.pass_rate > best) best = cell.pass_rate;
-        if (cell.pass_rate < worst) worst = cell.pass_rate;
-      }
-    }
-    bestByTarget.set(target, best);
-    worstByTarget.set(target, worst);
-  }
-
   return (
     <div className="space-y-4">
       <div className="flex items-center gap-4 text-sm text-gray-400">
         <span className="flex items-center gap-1.5">
-          <span className="inline-block h-3 w-3 rounded-sm bg-emerald-900/60 ring-1 ring-emerald-700/40" />
-          &gt;80%
+          <span className="inline-block h-3 w-3 rounded-sm bg-gray-800/60 ring-1 ring-emerald-500/60" />
+          <span className="text-emerald-400">80%+</span>
         </span>
         <span className="flex items-center gap-1.5">
-          <span className="inline-block h-3 w-3 rounded-sm bg-amber-900/40 ring-1 ring-amber-700/40" />
-          50-80%
+          <span className="inline-block h-3 w-3 rounded-sm bg-gray-800/60 ring-1 ring-amber-500/60" />
+          <span className="text-amber-400">50–80%</span>
         </span>
         <span className="flex items-center gap-1.5">
-          <span className="inline-block h-3 w-3 rounded-sm bg-red-900/40 ring-1 ring-red-700/40" />
-          &lt;50%
+          <span className="inline-block h-3 w-3 rounded-sm bg-gray-800/60 ring-1 ring-red-500/60" />
+          <span className="text-red-400">&lt;50%</span>
         </span>
         <span className="flex items-center gap-1.5">
           <span className="inline-block h-3 w-3 rounded-sm border border-dashed border-gray-700" />
@@ -124,8 +107,6 @@ export function CompareTab({ data, isLoading, isError, error }: CompareTabProps)
                 target={target}
                 experiments={experiments}
                 cellMap={cellMap}
-                bestRate={bestByTarget.get(target) ?? 0}
-                worstRate={worstByTarget.get(target) ?? 0}
               />
             ))}
           </tbody>
@@ -139,14 +120,10 @@ function CompareRow({
   target,
   experiments,
   cellMap,
-  bestRate,
-  worstRate,
 }: {
   target: string;
   experiments: string[];
   cellMap: Map<string, CompareCell>;
-  bestRate: number;
-  worstRate: number;
 }) {
   return (
     <tr className="transition-colors hover:bg-gray-900/30">
@@ -156,15 +133,7 @@ function CompareRow({
         return (
           <td key={exp} className="px-2 py-2">
             {cell ? (
-              <CompareMatrixCell
-                cell={cell}
-                isBest={
-                  experiments.length > 1 && cell.pass_rate === bestRate && bestRate !== worstRate
-                }
-                isWorst={
-                  experiments.length > 1 && cell.pass_rate === worstRate && bestRate !== worstRate
-                }
-              />
+              <CompareMatrixCell cell={cell} />
             ) : (
               <div className="flex items-center justify-center rounded-lg border border-dashed border-gray-700 px-3 py-4 text-gray-600">
                 --
@@ -177,10 +146,10 @@ function CompareRow({
   );
 }
 
-function passRateColorClass(rate: number): string {
-  if (rate >= 0.8) return 'bg-emerald-900/60 ring-emerald-700/40';
-  if (rate >= 0.5) return 'bg-amber-900/40 ring-amber-700/40';
-  return 'bg-red-900/40 ring-red-700/40';
+function passRateRingClass(rate: number): string {
+  if (rate >= 0.8) return 'ring-emerald-500/60';
+  if (rate >= 0.5) return 'ring-amber-500/60';
+  return 'ring-red-500/60';
 }
 
 function passRateTextClass(rate: number): string {
@@ -189,15 +158,7 @@ function passRateTextClass(rate: number): string {
   return 'text-red-400';
 }
 
-function CompareMatrixCell({
-  cell,
-  isBest,
-  isWorst,
-}: {
-  cell: CompareCell;
-  isBest: boolean;
-  isWorst: boolean;
-}) {
+function CompareMatrixCell({ cell }: { cell: CompareCell }) {
   const [expanded, setExpanded] = useState(false);
   const pct = Math.round(cell.pass_rate * 100);
   const avgPct = Math.round(cell.avg_score * 100);
@@ -208,26 +169,14 @@ function CompareMatrixCell({
         type="button"
         onClick={() => setExpanded(!expanded)}
         aria-expanded={expanded}
-        className={`w-full rounded-lg px-3 py-3 text-center ring-1 transition-colors ${passRateColorClass(cell.pass_rate)} hover:brightness-110 ${
-          isBest ? 'ring-2 ring-emerald-500/60' : isWorst ? 'ring-2 ring-red-500/40' : ''
-        }`}
+        className={`w-full rounded-lg bg-gray-800/60 px-3 py-3 text-center ring-1 transition-colors hover:bg-gray-700/60 ${passRateRingClass(cell.pass_rate)}`}
       >
-        <div className="flex items-center justify-center gap-1">
+        <div className="flex items-center justify-center">
           <span
             className={`text-lg font-semibold tabular-nums ${passRateTextClass(cell.pass_rate)}`}
           >
             {pct}%
           </span>
-          {isBest && (
-            <span className="text-xs text-emerald-400" title="Best performer">
-              &#9650;
-            </span>
-          )}
-          {isWorst && (
-            <span className="text-xs text-red-400" title="Worst performer">
-              &#9660;
-            </span>
-          )}
         </div>
         <div className="mt-0.5 text-xs text-gray-400">
           {cell.passed_count}/{cell.eval_count} pass | avg {avgPct}%
diff --git a/apps/studio/src/components/Sidebar.tsx b/apps/studio/src/components/Sidebar.tsx
index 474fcb426..62ec615a7 100644
--- a/apps/studio/src/components/Sidebar.tsx
+++ b/apps/studio/src/components/Sidebar.tsx
@@ -366,7 +366,7 @@ function ProjectRunDetailSidebar({
 
       <div className="border-b border-gray-800 px-4 py-2">
         <Link to="/" className="text-xs text-gray-400 hover:text-cyan-400">
-          &larr; All Projects
+          &larr; All Benchmarks
         </Link>
         <p className="mt-1 truncate text-sm font-medium text-gray-300">{projectId}</p>
       </div>
diff --git a/apps/studio/src/lib/types.ts b/apps/studio/src/lib/types.ts
index 19f7e96fc..93752599d 100644
--- a/apps/studio/src/lib/types.ts
+++ b/apps/studio/src/lib/types.ts
@@ -202,6 +202,7 @@ export interface StudioConfigResponse {
   pass_threshold?: number;
   read_only?: boolean;
   project_name?: string;
+  multi_project_dashboard?: boolean;
 }
 
 export interface RemoteStatusResponse {
diff --git a/apps/studio/src/routes/index.tsx b/apps/studio/src/routes/index.tsx
index f3059d44a..082518ad6 100644
--- a/apps/studio/src/routes/index.tsx
+++ b/apps/studio/src/routes/index.tsx
@@ -1,7 +1,7 @@
 /**
- * Home route: shows Projects Dashboard when projects are registered,
- * or the existing tabbed landing page (Runs, Experiments, Targets)
- * when in single-project mode.
+ * Home route: shows the multi-project dashboard when the server enables it,
+ * or the existing tabbed landing page (Runs, Experiments, Compare, Targets)
+ * in single-project mode.
  *
  * Uses URL search param `?tab=` for tab persistence.
  */
@@ -43,13 +43,15 @@ export const Route = createFileRoute('/')({
 
 function HomePage() {
   const { data: projectData, isLoading: projectsLoading } = useProjectList();
+  const { data: config, isLoading: configLoading } = useStudioConfig();
   const hasProjects = (projectData?.projects.length ?? 0) > 0;
+  const multiProjectDashboard = config?.multi_project_dashboard;
 
-  if (projectsLoading) {
+  if (projectsLoading || configLoading) {
     return <LoadingSkeleton />;
   }
 
-  if (hasProjects) {
+  if (multiProjectDashboard === true || (multiProjectDashboard === undefined && hasProjects)) {
     return <ProjectsDashboard />;
   }
 
@@ -104,7 +106,7 @@ function ProjectsDashboard() {
   return (
     <div className="space-y-6">
       <div className="flex items-center justify-between">
-        <h1 className="text-2xl font-semibold text-white">Projects</h1>
+        <h1 className="text-2xl font-semibold text-white">Benchmarks</h1>
         <div className="flex gap-2">
           {!isReadOnly && (
             <>
@@ -120,7 +122,7 @@ function ProjectsDashboard() {
                 onClick={() => setShowAddForm(!showAddForm)}
                 className="rounded-md bg-cyan-600 px-3 py-1.5 text-sm font-medium text-white hover:bg-cyan-500"
               >
-                {showAddForm ? 'Cancel' : 'Add Project'}
+                {showAddForm ? 'Cancel' : 'Add Benchmark'}
               </button>
             </>
           )}
@@ -140,7 +142,7 @@ function ProjectsDashboard() {
               type="text"
               value={addPath}
               onChange={(e) => setAddPath(e.target.value)}
-              placeholder="Project path (e.g., /home/user/projects/my-app)"
+              placeholder="Benchmark path (e.g., /home/user/projects/my-evals)"
               className="flex-1 rounded-md border border-gray-700 bg-gray-800 px-3 py-1.5 text-sm text-white placeholder-gray-500 focus:border-cyan-600 focus:outline-none"
             />
             <button
@@ -155,7 +157,7 @@ function ProjectsDashboard() {
               type="text"
               value={discoverPath}
               onChange={(e) => setDiscoverPath(e.target.value)}
-              placeholder="Discover projects in directory..."
+              placeholder="Discover benchmarks in directory..."
               className="flex-1 rounded-md border border-gray-700 bg-gray-800 px-3 py-1.5 text-sm text-white placeholder-gray-500 focus:border-cyan-600 focus:outline-none"
             />
             <button
diff --git a/apps/web/src/assets/screenshots/studio-compare.png b/apps/web/src/assets/screenshots/studio-compare.png
new file mode 100644
index 000000000..78c8690de
Binary files /dev/null and b/apps/web/src/assets/screenshots/studio-compare.png differ
diff --git a/apps/web/src/assets/screenshots/studio-projects-multi.png b/apps/web/src/assets/screenshots/studio-projects-multi.png
new file mode 100644
index 000000000..0255eb85f
Binary files /dev/null and b/apps/web/src/assets/screenshots/studio-projects-multi.png differ
diff --git a/apps/web/src/assets/screenshots/studio-runs-bench.png b/apps/web/src/assets/screenshots/studio-runs-bench.png
new file mode 100644
index 000000000..22c6a51ef
Binary files /dev/null and b/apps/web/src/assets/screenshots/studio-runs-bench.png differ
diff --git a/apps/web/src/content/docs/docs/evaluation/eval-cases.mdx b/apps/web/src/content/docs/docs/evaluation/eval-cases.mdx
index 2f69246c2..5ef6b832c 100644
--- a/apps/web/src/content/docs/docs/evaluation/eval-cases.mdx
+++ b/apps/web/src/content/docs/docs/evaluation/eval-cases.mdx
@@ -158,6 +158,7 @@ tests:
 ```
 
 The `metadata` field is included in the stdin JSON passed to lifecycle commands as `case_metadata`.
+Operational checkout state belongs under `workspace.repos[].checkout.base_commit`; `metadata.base_commit` is informational only. `workspace.docker.base_commit` is retained as a deprecated compatibility bridge for legacy Docker-backed evals.
 
 ## Per-Test Assertions
 
diff --git a/apps/web/src/content/docs/docs/evaluation/eval-files.mdx b/apps/web/src/content/docs/docs/evaluation/eval-files.mdx
index 521de9814..cf98867cb 100644
--- a/apps/web/src/content/docs/docs/evaluation/eval-files.mdx
+++ b/apps/web/src/content/docs/docs/evaluation/eval-files.mdx
@@ -94,6 +94,22 @@ tests:
 
 `assertions` supports all evaluator types, including deterministic assertion types (`contains`, `regex`, `is_json`, `equals`) and `rubrics`. See [Tests](/docs/evaluation/eval-cases/#per-test-assertions) for per-test assertions usage.
 
+### Assertion Includes
+
+Reusable assertion sets can be factored into template files and referenced from any `assertions` array:
+
+```yaml
+assertions:
+  - include: safe-response
+  - include: ./shared/format.yaml
+```
+
+Resolution rules:
+- `include: name` resolves to `.agentv/templates/{name}.yaml` with the closest matching directory winning
+- Relative paths resolve from the eval file location, so `include: ./shared/format.yaml` works as expected
+- Nested includes are allowed up to depth 3 to keep cycles and runaway recursion bounded
+- Suite-level includes follow the same merge behavior as other suite-level assertions and still respect `execution.skip_defaults: true`
+
 ### Suite-level Input
 
 The `input` field defines messages that are **prepended** to every test's input. This avoids repeating the same prompt or system context in each test case — following the same pattern as suite-level `assertions`.
diff --git a/apps/web/src/content/docs/docs/tools/import.mdx b/apps/web/src/content/docs/docs/tools/import.mdx
index 2521a6d26..4b3d4e153 100644
--- a/apps/web/src/content/docs/docs/tools/import.mdx
+++ b/apps/web/src/content/docs/docs/tools/import.mdx
@@ -147,3 +147,36 @@ agentv eval evals/my-eval.yaml --transcript .agentv/transcripts/claude-4c4f9e4e.
 ```
 
 See `examples/features/import-claude/` for a complete working example.
+
+## HuggingFace Datasets (SWE-bench)
+
+Use `scripts/import-huggingface.py` to convert HuggingFace benchmark datasets into AgentV eval files. Currently supports SWE-bench-style datasets.
+
+```bash
+uv run scripts/import-huggingface.py \
+  --repo SWE-bench/SWE-bench_Verified \
+  --split test \
+  --limit 10 \
+  --output evals/swebench/
+```
+
+Each instance becomes an EVAL.yaml with:
+- `input` — the problem statement
+- `workspace.docker.image` — the pre-built SWE-bench Docker image (`ghcr.io/epoch-research/swe-bench.eval.x86_64.<instance_id>:latest`)
+- `workspace.docker.base_commit` — the commit to reset to before the agent runs
+- `assertions` — `code-grader` tasks that run `FAIL_TO_PASS` and `PASS_TO_PASS` pytest suites inside the container
+
+Run an imported SWE-bench eval against any coding agent target:
+
+```bash
+# Import one instance
+uv run scripts/import-huggingface.py \
+  --repo SWE-bench/SWE-bench_Verified \
+  --limit 1 \
+  --output /tmp/swebench-eval/
+
+# Run with a coding agent target
+agentv eval /tmp/swebench-eval/*.EVAL.yaml --target codex
+```
+
+The Docker workspace spins up the pre-built SWE-bench image, checks out `base_commit`, runs the agent to apply a patch, then grades by running the test suite inside the container.
diff --git a/apps/web/src/content/docs/docs/tools/studio.mdx b/apps/web/src/content/docs/docs/tools/studio.mdx
index 233ae4c5d..57dd61ff1 100644
--- a/apps/web/src/content/docs/docs/tools/studio.mdx
+++ b/apps/web/src/content/docs/docs/tools/studio.mdx
@@ -10,10 +10,13 @@ import studioRuns from '../../../../assets/screenshots/studio-runs.png';
 import studioRunDetail from '../../../../assets/screenshots/studio-run-detail.png';
 import studioExperiments from '../../../../assets/screenshots/studio-experiments.png';
 import studioProjects from '../../../../assets/screenshots/studio-projects.png';
+import studioProjectsMulti from '../../../../assets/screenshots/studio-projects-multi.png';
+import studioCompare from '../../../../assets/screenshots/studio-compare.png';
+import studioRunsBench from '../../../../assets/screenshots/studio-runs-bench.png';
 
 The `studio` command launches a web-based dashboard for browsing evaluation runs, inspecting individual test results, and reviewing scores. It shows both local runs and runs synced from a remote results repository.
 
-<Image src={studioRuns} alt="AgentV Studio showing evaluation runs with source badges, pass rates, and scores" />
+<Image src={studioRunsBench} alt="AgentV Studio showing evaluation runs with pass rates, targets, and experiment names" />
 
 ## Usage
 
@@ -37,7 +40,8 @@ agentv studio .agentv/results/runs/2026-03-30T11-45-56-989Z
 |--------|-------------|
 | `--port`, `-p` | Port to listen on (flag > `PORT` env var > 3117) |
 | `--dir`, `-d` | Working directory (default: current directory) |
-| `--multi` | Launch in multi-project dashboard mode |
+| `--multi` | Launch in multi-project dashboard mode (deprecated; use auto-detect or `--single`) |
+| `--single` | Force single-project dashboard mode |
 | `--add <path>` | Register a project by path |
 | `--remove <id>` | Unregister a project by ID |
 | `--discover <path>` | Scan a directory tree for repos with `.agentv/` |
@@ -49,6 +53,7 @@ agentv studio .agentv/results/runs/2026-03-30T11-45-56-989Z
 - **Targets** — group runs by target (model/agent)
 - **Run Detail** — drill into a run to see per-test results, scores, and evaluator output
 - **Human Review** — add feedback annotations to individual test results
+- **Comparison Matrix** — experiment × target matrix showing pass rates across dimensions
 - **Remote Results** — sync and browse runs pushed from other machines or CI (see [Remote Results](#remote-results))
 
 ## Run Detail
@@ -63,24 +68,43 @@ The Experiments tab groups runs by experiment name so you can compare the impact
 
 <Image src={studioExperiments} alt="AgentV Studio experiments tab comparing with_skills (100%) vs without_skills (60%) pass rates" />
 
-## Multi-Project Dashboard
+## Comparison Matrix
 
-By default, Studio shows results for the current directory. The multi-project mode lets you view results across multiple repositories from a single dashboard.
+The **Compare** tab shows a cross-model, cross-experiment performance matrix. Cells are color-coded by pass rate: green (80%+), yellow (50–80%), red (below 50%). The best performer per row has an emerald ring; the worst has a red ring. Click any cell to expand per-test-case results.
 
-### Registering Projects
+<Image src={studioCompare} alt="AgentV Studio comparison matrix showing experiment vs target pass rates with color coding" />
 
-Register projects one at a time:
+Run the same eval against multiple providers or experiment variants, then open the Compare tab:
 
 ```bash
-agentv studio --add /path/to/project-a
-agentv studio --add /path/to/project-b
+agentv eval my.EVAL.yaml --target azure --experiment baseline
+agentv eval my.EVAL.yaml --target azure --experiment with-caching
+agentv eval my.EVAL.yaml --target gemini --experiment baseline
+agentv eval my.EVAL.yaml --target gemini --experiment with-caching
+agentv studio  # Compare tab shows 2x2 matrix
 ```
 
-Each path must contain a `.agentv/` directory. Projects are stored in `~/.agentv/projects.yaml`.
+The matrix is available per-project under the **Compare** tab.
+
+
+## Benchmarks Dashboard
+
+By default, Studio shows results for the current directory. Register multiple benchmark repos to view them from a single dashboard.
+
+### Registering Benchmarks
+
+Register benchmark repos one at a time:
+
+```bash
+agentv studio --add /path/to/my-evals
+agentv studio --add /path/to/other-evals
+```
+
+Each path must contain a `.agentv/` directory. Registered benchmarks are stored in `~/.agentv/projects.yaml`.
 
 ### Auto-Discovery
 
-Scan a parent directory to find and register all projects:
+Scan a parent directory to find and register all benchmark repos:
 
 ```bash
 agentv studio --discover /path/to/repos
@@ -90,25 +114,29 @@ This recursively searches (up to 2 levels deep) for directories containing `.age
 
 ### Launching the Dashboard
 
-Once projects are registered, launch the multi-project dashboard:
+Studio auto-detects the mode based on how many benchmarks are registered:
+
+- `0` or `1` registered: single-project view
+- `2+` registered: Benchmarks dashboard
 
 ```bash
-agentv studio --multi
+agentv studio          # auto-detects
+agentv studio --single # force single-project view
 ```
 
-If you have any registered projects, `--multi` is automatically enabled. The landing page shows a card for each project with run count, pass rate, and last run time. Click a project to view its runs.
+The landing page shows a card for each benchmark with run count, pass rate, and last run time.
 
-<Image src={studioProjects} alt="AgentV Studio multi-project dashboard showing project cards with pass rates" />
+<Image src={studioProjectsMulti} alt="AgentV Studio benchmarks dashboard showing benchmark cards with pass rates" />
 
-### Removing Projects
+### Removing a Benchmark
 
-Unregister a project by its ID:
+Unregister by its ID:
 
 ```bash
-agentv studio --remove my-project
+agentv studio --remove my-evals
 ```
 
-Project IDs are derived from the directory name (e.g., `/home/user/repos/my-project` becomes `my-project`).
+IDs are derived from the directory name (e.g., `/home/user/repos/my-evals` becomes `my-evals`).
 
 ## Remote Results
 
diff --git a/examples/features/docker-workspace/README.md b/examples/features/docker-workspace/README.md
index 2e15e675b..7e27007bd 100644
--- a/examples/features/docker-workspace/README.md
+++ b/examples/features/docker-workspace/README.md
@@ -37,6 +37,8 @@ workspace:
     cpus: 2          # optional Docker CPU limit
 ```
 
+For evals that need a repo pinned to a dataset snapshot, prefer `workspace.repos[].checkout.base_commit`. `workspace.docker.base_commit` still works as a compatibility bridge for existing Docker-backed SWE-bench configs, but new configs should keep checkout state in the repo model rather than in the Docker block.
+
 ## Running
 
 ```bash
diff --git a/packages/core/src/evaluation/evaluators/code-evaluator.ts b/packages/core/src/evaluation/evaluators/code-evaluator.ts
index fa47981ab..8762cde27 100644
--- a/packages/core/src/evaluation/evaluators/code-evaluator.ts
+++ b/packages/core/src/evaluation/evaluators/code-evaluator.ts
@@ -11,6 +11,7 @@ import {
 import { toSnakeCaseDeep } from '../case-conversion.js';
 import { type ContentImage, isContentArray } from '../content.js';
 import type { AssertionEntry, JsonObject, TargetAccessConfig } from '../types.js';
+import { getRepoCheckoutTargets } from '../workspace/repo-checkout.js';
 import { clampScore, isNonEmptyString, parseJsonSafe, scoreToVerdict } from './scoring.js';
 import type { EvaluationContext, EvaluationScore, Evaluator } from './types.js';
 
@@ -218,6 +219,7 @@ export class CodeEvaluator implements Evaluator {
         const result = await dockerProvider.runGraderInContainer({
           command: [...this.command],
           stdin: inputPayload,
+          repoCheckouts: getRepoCheckoutTargets(context.evalCase.workspace?.repos),
         });
         if (result.exitCode !== 0) {
           const trimmedErr = result.stderr.trim();
diff --git a/packages/core/src/evaluation/loaders/evaluator-parser.ts b/packages/core/src/evaluation/loaders/evaluator-parser.ts
index 013f29b69..2e69b4639 100644
--- a/packages/core/src/evaluation/loaders/evaluator-parser.ts
+++ b/packages/core/src/evaluation/loaders/evaluator-parser.ts
@@ -1,6 +1,9 @@
+import { readFile } from 'node:fs/promises';
 import path from 'node:path';
+import { parse } from 'yaml';
 
 import { normalizePreprocessorType } from '../content-preprocessor.js';
+import { interpolateEnv } from '../interpolation.js';
 import type { ToolTrajectoryEvaluatorConfig, ToolTrajectoryExpectedItem } from '../trace.js';
 import type {
   ContentPreprocessorConfig,
@@ -15,6 +18,7 @@ import { resolveFileReference } from './file-resolver.js';
 
 const ANSI_YELLOW = '\u001b[33m';
 const ANSI_RESET = '\u001b[0m';
+const MAX_ASSERTION_INCLUDE_DEPTH = 3;
 
 /**
  * Prefix for explicit file references in prompt strings.
@@ -98,15 +102,121 @@ export async function parseEvaluators(
   return evaluators.length > 0 ? evaluators : undefined;
 }
 
-/**
- * Parse a raw evaluator array into typed EvaluatorConfig objects.
- */
-async function parseEvaluatorList(
+interface IncludeContext {
+  readonly depth: number;
+  readonly chain: readonly string[];
+}
+
+function isIncludeEntry(value: unknown): value is { include: string } {
+  return (
+    isJsonObject(value) && typeof value.include === 'string' && Object.keys(value).length === 1
+  );
+}
+
+function isTemplateReference(value: string): boolean {
+  return !value.startsWith('.') && !value.includes('/') && !value.includes('\\');
+}
+
+async function resolveAssertionTemplateReference(
+  include: string,
+  searchRoots: readonly string[],
+): Promise<{
+  readonly displayPath: string;
+  readonly resolvedPath: string;
+  readonly attempted: readonly string[];
+}> {
+  const templateCandidates = isTemplateReference(include)
+    ? [
+        path.join('.agentv', 'templates', `${include}.yaml`),
+        path.join('.agentv', 'templates', `${include}.yml`),
+      ]
+    : [include];
+
+  const attempted: string[] = [];
+  for (const candidate of templateCandidates) {
+    const resolved = await resolveFileReference(candidate, searchRoots);
+    attempted.push(...resolved.attempted);
+    if (resolved.resolvedPath) {
+      return {
+        displayPath: resolved.displayPath,
+        resolvedPath: resolved.resolvedPath,
+        attempted,
+      };
+    }
+  }
+
+  return {
+    displayPath: templateCandidates[0] ?? include,
+    resolvedPath: '',
+    attempted,
+  };
+}
+
+async function loadAssertionTemplateEntries(
+  include: string,
+  searchRoots: readonly string[],
+  evalId: string,
+  includeContext: IncludeContext,
+): Promise<readonly unknown[]> {
+  const nextDepth = includeContext.depth + 1;
+  if (nextDepth > MAX_ASSERTION_INCLUDE_DEPTH) {
+    const chain = [...includeContext.chain, include].join(' -> ');
+    throw new Error(
+      `Assertion template include depth exceeded ${MAX_ASSERTION_INCLUDE_DEPTH} in '${evalId}'. Include chain: ${chain}`,
+    );
+  }
+
+  const resolved = await resolveAssertionTemplateReference(include, searchRoots);
+  if (!resolved.resolvedPath) {
+    const attempted =
+      resolved.attempted.length > 0
+        ? `\n${resolved.attempted.map((attempt) => `  Tried: ${attempt}`).join('\n')}`
+        : '';
+    throw new Error(
+      `Assertion template not found in '${evalId}': ${resolved.displayPath}${attempted}`,
+    );
+  }
+
+  if (includeContext.chain.includes(resolved.resolvedPath)) {
+    const cycle = [...includeContext.chain, resolved.resolvedPath].join(' -> ');
+    throw new Error(`Assertion template cycle detected in '${evalId}': ${cycle}`);
+  }
+
+  const content = await readFile(resolved.resolvedPath, 'utf8');
+  const parsed = interpolateEnv(parse(content), process.env) as unknown;
+  if (!isJsonObject(parsed)) {
+    throw new Error(
+      `Invalid assertion template file in '${evalId}': ${resolved.resolvedPath} (expected a YAML object with an assertions array)`,
+    );
+  }
+
+  const assertions = (parsed as Record<string, unknown>).assertions;
+  if (!Array.isArray(assertions)) {
+    throw new Error(
+      `Invalid assertion template file in '${evalId}': ${resolved.resolvedPath} is missing a top-level assertions array`,
+    );
+  }
+
+  const templateDir = path.dirname(resolved.resolvedPath);
+  const nestedSearchRoots = [
+    templateDir,
+    ...searchRoots.filter((root) => path.resolve(root) !== templateDir),
+  ];
+
+  return (
+    (await expandEvaluatorEntries(assertions, nestedSearchRoots, evalId, {
+      depth: nextDepth,
+      chain: [...includeContext.chain, resolved.resolvedPath],
+    })) ?? []
+  );
+}
+
+async function expandEvaluatorEntries(
   candidateEvaluators: JsonValue | undefined,
   searchRoots: readonly string[],
   evalId: string,
-  defaultPreprocessors?: readonly ContentPreprocessorConfig[],
-): Promise<readonly EvaluatorConfig[] | undefined> {
+  includeContext: IncludeContext = { depth: 0, chain: [] },
+): Promise<readonly unknown[] | undefined> {
   if (candidateEvaluators === undefined) {
     return undefined;
   }
@@ -116,19 +226,51 @@ async function parseEvaluatorList(
     return undefined;
   }
 
+  const expanded: unknown[] = [];
+  for (const rawEvaluator of candidateEvaluators) {
+    if (isIncludeEntry(rawEvaluator)) {
+      const included = await loadAssertionTemplateEntries(
+        rawEvaluator.include,
+        searchRoots,
+        evalId,
+        includeContext,
+      );
+      expanded.push(...included);
+      continue;
+    }
+    expanded.push(rawEvaluator);
+  }
+
+  return expanded;
+}
+
+/**
+ * Parse a raw evaluator array into typed EvaluatorConfig objects.
+ */
+async function parseEvaluatorList(
+  candidateEvaluators: JsonValue | undefined,
+  searchRoots: readonly string[],
+  evalId: string,
+  defaultPreprocessors?: readonly ContentPreprocessorConfig[],
+): Promise<readonly EvaluatorConfig[] | undefined> {
+  const expandedEvaluators = await expandEvaluatorEntries(candidateEvaluators, searchRoots, evalId);
+  if (!expandedEvaluators) {
+    return undefined;
+  }
+
   // Pre-process: collect all string entries across the array (regardless of position) and
   // group them into a single rubrics evaluator inserted at the first-string position.
   // Non-string entries are preserved in their original relative order.
-  const firstStringIndex = candidateEvaluators.findIndex((e) => typeof e === 'string');
+  const firstStringIndex = expandedEvaluators.findIndex((e) => typeof e === 'string');
   const processedEvaluators: unknown[] =
     firstStringIndex === -1
-      ? [...candidateEvaluators]
+      ? [...expandedEvaluators]
       : (() => {
           const PLACEHOLDER = Symbol('rubric-placeholder');
           const strings: string[] = [];
           const result: unknown[] = [];
           let rubricInserted = false;
-          for (const item of candidateEvaluators) {
+          for (const item of expandedEvaluators) {
             if (typeof item === 'string') {
               const trimmed = item.trim();
               if (trimmed.length === 0) {
@@ -395,9 +537,18 @@ async function parseEvaluatorList(
         continue;
       }
 
+      const expandedMembers = await expandEvaluatorEntries(
+        rawMembers,
+        searchRoots,
+        `${evalId}:${name}`,
+      );
+      if (!expandedMembers) {
+        continue;
+      }
+
       // Recursively parse member evaluators
       const memberEvaluators: EvaluatorConfig[] = [];
-      for (const rawMember of rawMembers) {
+      for (const rawMember of expandedMembers) {
         if (!isJsonObject(rawMember)) {
           logWarning(`Skipping invalid member evaluator in composite '${name}' (expected object)`);
           continue;
diff --git a/packages/core/src/evaluation/types.ts b/packages/core/src/evaluation/types.ts
index 8614d54ee..7e0214095 100644
--- a/packages/core/src/evaluation/types.ts
+++ b/packages/core/src/evaluation/types.ts
@@ -232,6 +232,8 @@ export type RepoSource =
 
 export type RepoCheckout = {
   readonly ref?: string;
+  /** SWE-bench-friendly alias for ref when pinning a dataset snapshot commit */
+  readonly base_commit?: string;
   readonly resolve?: 'remote' | 'local';
   readonly ancestor?: number;
 };
@@ -290,6 +292,8 @@ export type DockerWorkspaceConfig = {
   readonly memory?: string;
   /** CPU limit (e.g. 2, 0.5) */
   readonly cpus?: number;
+  /** @deprecated Prefer workspace.repos[].checkout.base_commit as the checkout source of truth */
+  readonly base_commit?: string;
 };
 
 export type WorkspaceConfig = {
diff --git a/packages/core/src/evaluation/validation/eval-file.schema.ts b/packages/core/src/evaluation/validation/eval-file.schema.ts
index c980e6baf..f45ec40a3 100644
--- a/packages/core/src/evaluation/validation/eval-file.schema.ts
+++ b/packages/core/src/evaluation/validation/eval-file.schema.ts
@@ -101,6 +101,12 @@ const LlmGraderSchema = EvaluatorCommonSchema.extend({
   preprocessors: z.array(PreprocessorSchema).optional(),
 });
 
+const IncludeSchema = z
+  .object({
+    include: z.string().min(1),
+  })
+  .strict();
+
 /** Aggregator configs for composite evaluator */
 const AggregatorSchema = z.discriminatedUnion('type', [
   z.object({
@@ -228,6 +234,7 @@ const RubricsSchema = EvaluatorCommonSchema.extend({
 const EvaluatorSchema = z.union([
   CodeGraderSchema,
   LlmGraderSchema,
+  IncludeSchema,
   CompositeSchema,
   ToolTrajectorySchema,
   FieldAccuracySchema,
@@ -264,6 +271,7 @@ const RepoSourceSchema = z.discriminatedUnion('type', [
 
 const RepoCheckoutSchema = z.object({
   ref: z.string().optional(),
+  base_commit: z.string().min(1).optional(),
   resolve: z.enum(['remote', 'local']).optional(),
   ancestor: z.number().int().min(0).optional(),
 });
@@ -303,6 +311,7 @@ const DockerWorkspaceSchema = z.object({
   timeout: z.number().int().min(1).optional(),
   memory: z.string().optional(),
   cpus: z.number().min(0.1).optional(),
+  base_commit: z.string().min(1).optional(),
 });
 
 const WorkspaceSchema = z
diff --git a/packages/core/src/evaluation/validation/eval-validator.ts b/packages/core/src/evaluation/validation/eval-validator.ts
index d1ce1aa5c..6f7f1bbda 100644
--- a/packages/core/src/evaluation/validation/eval-validator.ts
+++ b/packages/core/src/evaluation/validation/eval-validator.ts
@@ -420,7 +420,7 @@ function validateWorkspaceRepoConfig(
             location: `workspace.repos[path=${repo.path}]`,
             message:
               'checkout.resolve has no effect for a local source. ' +
-              'Use source.type to choose where the repo comes from; keep checkout.ref or checkout.ancestor only when pinning a local source.',
+              'Use source.type to choose where the repo comes from; keep checkout.ref, checkout.base_commit, or checkout.ancestor only when pinning a local source.',
           });
         }
       }
@@ -454,6 +454,17 @@ function validateWorkspaceRepoConfig(
     }
   }
 
+  const docker = workspace.docker;
+  if (isObject(docker) && typeof docker.base_commit === 'string') {
+    errors.push({
+      severity: 'warning',
+      filePath,
+      location: 'workspace.docker.base_commit',
+      message:
+        'workspace.docker.base_commit is deprecated. Prefer workspace.repos[].checkout.base_commit so checkout state remains backend-agnostic.',
+    });
+  }
+
   // after_each reset with per_test isolation warning
   if (isObject(afterEachHook) && afterEachHook.reset && isolation === 'per_test') {
     errors.push({
diff --git a/packages/core/src/evaluation/workspace/docker-workspace.ts b/packages/core/src/evaluation/workspace/docker-workspace.ts
index 700e90fbe..26592c824 100644
--- a/packages/core/src/evaluation/workspace/docker-workspace.ts
+++ b/packages/core/src/evaluation/workspace/docker-workspace.ts
@@ -13,6 +13,7 @@
  */
 
 import type { DockerWorkspaceConfig } from '../types.js';
+import type { RepoCheckoutTarget } from './repo-checkout.js';
 
 /** Result of a command execution */
 export interface ExecResult {
@@ -62,6 +63,16 @@ export interface ExecInContainerOptions {
 
 const DEFAULT_TIMEOUT_S = 1800;
 
+function buildGitCommand(
+  target: RepoCheckoutTarget | undefined,
+  args: readonly string[],
+): string[] {
+  if (!target?.path) {
+    return ['git', ...args];
+  }
+  return ['git', '-C', target.path, ...args];
+}
+
 /**
  * Manages Docker container lifecycle for workspace-based evaluations.
  *
@@ -155,6 +166,57 @@ export class DockerWorkspaceProvider {
     }
   }
 
+  /**
+   * Reset the container checkout to the configured base commit, if present.
+   * This is used for SWE-bench images where the repo state must match the
+   * dataset's base snapshot before grading begins.
+   */
+  async resetContainerCheckout(
+    containerId: string,
+    repoCheckouts?: readonly RepoCheckoutTarget[],
+  ): Promise<void> {
+    const checkoutTargets =
+      repoCheckouts && repoCheckouts.length > 0
+        ? repoCheckouts
+        : this.config.base_commit
+          ? [{ ref: this.config.base_commit }]
+          : [];
+
+    if (checkoutTargets.length === 0) {
+      return;
+    }
+
+    for (const target of checkoutTargets) {
+      const resetResult = await this.execInContainer({
+        containerId,
+        command: buildGitCommand(target, ['reset', '--hard', target.ref]),
+      });
+      if (resetResult.exitCode !== 0) {
+        throw new Error(
+          `docker git reset failed (exit ${resetResult.exitCode}): ${resetResult.stderr.trim()}`,
+        );
+      }
+
+      const verifyResult = await this.execInContainer({
+        containerId,
+        command: buildGitCommand(target, ['rev-parse', 'HEAD']),
+        timeoutMs: 30_000,
+      });
+      if (verifyResult.exitCode !== 0) {
+        throw new Error(
+          `docker checkout verification failed (exit ${verifyResult.exitCode}): ${verifyResult.stderr.trim()}`,
+        );
+      }
+
+      const head = verifyResult.stdout.trim();
+      if (head !== target.ref) {
+        throw new Error(
+          `docker checkout verification failed: expected ${target.ref} but found ${head || '<empty>'}`,
+        );
+      }
+    }
+  }
+
   /** Copy a local file or directory into a running container. */
   async copyToContainer(
     containerId: string,
@@ -206,10 +268,12 @@ export class DockerWorkspaceProvider {
     readonly command: readonly string[];
     readonly stdin?: string;
     readonly copyFiles?: ReadonlyArray<{ localPath: string; containerPath: string }>;
+    readonly repoCheckouts?: readonly RepoCheckoutTarget[];
   }): Promise<ExecResult> {
     const containerId = await this.createContainer();
     try {
       await this.startContainer(containerId);
+      await this.resetContainerCheckout(containerId, options.repoCheckouts);
 
       if (options.copyFiles) {
         for (const file of options.copyFiles) {
diff --git a/packages/core/src/evaluation/workspace/pool-manager.ts b/packages/core/src/evaluation/workspace/pool-manager.ts
index 86f2f3cef..f1e9913fd 100644
--- a/packages/core/src/evaluation/workspace/pool-manager.ts
+++ b/packages/core/src/evaluation/workspace/pool-manager.ts
@@ -7,6 +7,7 @@ import { promisify } from 'node:util';
 
 import { getWorkspacePoolRoot } from '../../paths.js';
 import type { RepoConfig } from '../types.js';
+import { getRepoCheckoutRef } from './repo-checkout.js';
 import type { RepoManager } from './repo-manager.js';
 
 const execFileAsync = promisify(execFile);
@@ -74,7 +75,7 @@ function normalizeRepoForFingerprint(repo: RepoConfig): Record<string, unknown>
   const result: Record<string, unknown> = {
     path: repo.path,
     source,
-    ref: repo.checkout?.ref ?? 'HEAD',
+    ref: getRepoCheckoutRef(repo.checkout),
   };
 
   if (repo.clone?.depth !== undefined) {
@@ -372,7 +373,7 @@ export class WorkspacePoolManager {
       if (poolReset === 'none') {
         continue;
       }
-      const ref = repo.checkout?.ref ?? 'HEAD';
+      const ref = getRepoCheckoutRef(repo.checkout);
       const resolve = repo.checkout?.resolve ?? 'remote';
 
       // When resolve is 'remote', fetch latest from origin before resetting
diff --git a/packages/core/src/evaluation/workspace/repo-checkout.ts b/packages/core/src/evaluation/workspace/repo-checkout.ts
new file mode 100644
index 000000000..edf3f2ee5
--- /dev/null
+++ b/packages/core/src/evaluation/workspace/repo-checkout.ts
@@ -0,0 +1,22 @@
+import type { RepoCheckout, RepoConfig } from '../types.js';
+
+export interface RepoCheckoutTarget {
+  readonly path?: string;
+  readonly ref: string;
+}
+
+export function getRepoCheckoutRef(checkout: RepoCheckout | undefined): string {
+  return checkout?.base_commit ?? checkout?.ref ?? 'HEAD';
+}
+
+export function getRepoCheckoutTargets(
+  repos: readonly RepoConfig[] | undefined,
+): RepoCheckoutTarget[] {
+  if (!repos) return [];
+  return repos
+    .filter((repo) => repo.checkout?.base_commit || repo.checkout?.ref)
+    .map((repo) => ({
+      path: repo.path,
+      ref: getRepoCheckoutRef(repo.checkout),
+    }));
+}
diff --git a/packages/core/src/evaluation/workspace/repo-config-parser.ts b/packages/core/src/evaluation/workspace/repo-config-parser.ts
index ef2dca082..cd7e62820 100644
--- a/packages/core/src/evaluation/workspace/repo-config-parser.ts
+++ b/packages/core/src/evaluation/workspace/repo-config-parser.ts
@@ -23,11 +23,13 @@ export function parseRepoCheckout(raw: unknown): RepoCheckout | undefined {
   if (!isJsonObject(raw)) return undefined;
   const obj = raw as Record<string, unknown>;
   const ref = typeof obj.ref === 'string' ? obj.ref : undefined;
+  const baseCommit = typeof obj.base_commit === 'string' ? obj.base_commit : undefined;
   const resolve = obj.resolve === 'remote' || obj.resolve === 'local' ? obj.resolve : undefined;
   const ancestor = typeof obj.ancestor === 'number' ? obj.ancestor : undefined;
-  if (!ref && !resolve && ancestor === undefined) return undefined;
+  if (!ref && !baseCommit && !resolve && ancestor === undefined) return undefined;
   return {
     ...(ref !== undefined && { ref }),
+    ...(baseCommit !== undefined && { base_commit: baseCommit }),
     ...(resolve !== undefined && { resolve }),
     ...(ancestor !== undefined && { ancestor }),
   };
diff --git a/packages/core/src/evaluation/workspace/repo-manager.ts b/packages/core/src/evaluation/workspace/repo-manager.ts
index cfa2cd031..0b67fea28 100644
--- a/packages/core/src/evaluation/workspace/repo-manager.ts
+++ b/packages/core/src/evaluation/workspace/repo-manager.ts
@@ -4,6 +4,7 @@ import path from 'node:path';
 import { promisify } from 'node:util';
 
 import type { RepoConfig, RepoSource } from '../types.js';
+import { getRepoCheckoutRef } from './repo-checkout.js';
 
 /**
  * Validation error for a local repo source path that doesn't exist or is unresolved.
@@ -160,7 +161,7 @@ export class RepoManager {
     }
 
     // Resolve ref
-    const ref = repo.checkout?.ref ?? 'HEAD';
+    const ref = getRepoCheckoutRef(repo.checkout);
     const resolve = repo.checkout?.resolve ?? 'remote';
 
     let resolvedSha: string;
diff --git a/packages/core/src/evaluation/yaml-parser.ts b/packages/core/src/evaluation/yaml-parser.ts
index 9e6f7de1e..0f4f1dbae 100644
--- a/packages/core/src/evaluation/yaml-parser.ts
+++ b/packages/core/src/evaluation/yaml-parser.ts
@@ -737,6 +737,7 @@ function parseDockerWorkspaceConfig(raw: unknown): DockerWorkspaceConfig | undef
     ...(typeof obj.timeout === 'number' && { timeout: obj.timeout }),
     ...(typeof obj.memory === 'string' && { memory: obj.memory }),
     ...(typeof obj.cpus === 'number' && { cpus: obj.cpus }),
+    ...(typeof obj.base_commit === 'string' && { base_commit: obj.base_commit }),
   };
 }
 
diff --git a/packages/core/test/evaluation/loaders/evaluator-parser.test.ts b/packages/core/test/evaluation/loaders/evaluator-parser.test.ts
index 00a3002b1..5d7b393f0 100644
--- a/packages/core/test/evaluation/loaders/evaluator-parser.test.ts
+++ b/packages/core/test/evaluation/loaders/evaluator-parser.test.ts
@@ -1380,6 +1380,201 @@ describe('parseEvaluators - assert field', () => {
   });
 });
 
+describe('parseEvaluators - assertion templates', () => {
+  let tempDir: string;
+
+  beforeAll(async () => {
+    tempDir = path.join(os.tmpdir(), `agentv-test-templates-${Date.now()}`);
+    await mkdir(tempDir, { recursive: true });
+  });
+
+  afterAll(async () => {
+    await rm(tempDir, { recursive: true, force: true });
+  });
+
+  async function writeTemplate(relativePath: string, content: string): Promise<string> {
+    const filePath = path.join(tempDir, relativePath);
+    await mkdir(path.dirname(filePath), { recursive: true });
+    await writeFile(filePath, content);
+    return filePath;
+  }
+
+  it('resolves convention-based includes from .agentv/templates', async () => {
+    await writeTemplate(
+      '.agentv/templates/shared.yaml',
+      `
+assertions:
+  - type: contains
+    value: shared
+`,
+    );
+
+    const evaluators = await parseEvaluators(
+      {
+        evaluators: [{ include: 'shared' }],
+      },
+      undefined,
+      [tempDir],
+      'test-1',
+    );
+
+    expect(evaluators).toHaveLength(1);
+    expect(evaluators?.[0]).toMatchObject({ type: 'contains', value: 'shared' });
+  });
+
+  it('resolves relative template paths and respects skip_defaults', async () => {
+    await writeTemplate(
+      'shared/local.yaml',
+      `
+assertions:
+  - type: contains
+    value: local
+`,
+    );
+
+    const evaluators = await parseEvaluators(
+      {
+        assertions: [{ type: 'contains', value: 'case-only' }],
+        execution: { skip_defaults: true },
+      },
+      {
+        assertions: [{ include: './shared/local.yaml' }],
+      },
+      [tempDir],
+      'test-2',
+    );
+
+    expect(evaluators).toHaveLength(1);
+    expect(evaluators?.[0]).toMatchObject({ type: 'contains', value: 'case-only' });
+  });
+
+  it('resolves nested relative includes from the template file directory', async () => {
+    await writeTemplate(
+      'templates/nested/outer.yaml',
+      `
+assertions:
+  - include: ./inner.yaml
+  - type: regex
+    value: nested
+`,
+    );
+    await writeTemplate(
+      'templates/nested/inner.yaml',
+      `
+assertions:
+  - type: contains
+    value: inner
+`,
+    );
+
+    const evaluators = await parseEvaluators(
+      {
+        assertions: [{ include: './templates/nested/outer.yaml' }],
+      },
+      undefined,
+      [tempDir],
+      'test-relative-nested',
+    );
+
+    expect(evaluators).toHaveLength(2);
+    expect(evaluators?.[0]).toMatchObject({ type: 'contains', value: 'inner' });
+    expect(evaluators?.[1]).toMatchObject({ type: 'regex', value: 'nested' });
+  });
+
+  it('expands nested template includes up to depth 3', async () => {
+    await writeTemplate(
+      '.agentv/templates/level-a.yaml',
+      `
+assertions:
+  - include: level-b
+`,
+    );
+    await writeTemplate(
+      '.agentv/templates/level-b.yaml',
+      `
+assertions:
+  - include: level-c
+`,
+    );
+    await writeTemplate(
+      '.agentv/templates/level-c.yaml',
+      `
+assertions:
+  - type: contains
+    value: nested-ok
+`,
+    );
+
+    const evaluators = await parseEvaluators(
+      {
+        evaluators: [{ include: 'level-a' }],
+      },
+      undefined,
+      [tempDir],
+      'test-3',
+    );
+
+    expect(evaluators).toHaveLength(1);
+    expect(evaluators?.[0]).toMatchObject({ type: 'contains', value: 'nested-ok' });
+  });
+
+  it('throws when nested includes exceed depth 3', async () => {
+    await writeTemplate(
+      '.agentv/templates/depth-a.yaml',
+      `
+assertions:
+  - include: depth-b
+`,
+    );
+    await writeTemplate(
+      '.agentv/templates/depth-b.yaml',
+      `
+assertions:
+  - include: depth-c
+`,
+    );
+    await writeTemplate(
+      '.agentv/templates/depth-c.yaml',
+      `
+assertions:
+  - include: depth-d
+`,
+    );
+    await writeTemplate(
+      '.agentv/templates/depth-d.yaml',
+      `
+assertions:
+  - type: contains
+    value: too-deep
+`,
+    );
+
+    await expect(
+      parseEvaluators(
+        {
+          evaluators: [{ include: 'depth-a' }],
+        },
+        undefined,
+        [tempDir],
+        'test-4',
+      ),
+    ).rejects.toThrow(/depth exceeded 3/i);
+  });
+
+  it('throws a clear error when a template is missing', async () => {
+    await expect(
+      parseEvaluators(
+        {
+          evaluators: [{ include: 'missing-template' }],
+        },
+        undefined,
+        [tempDir],
+        'test-5',
+      ),
+    ).rejects.toThrow(/\.agentv\/templates\/missing-template\.yaml/);
+  });
+});
+
 describe('parseEvaluators - type: rubrics with criteria', () => {
   let tempDir: string;
 
diff --git a/packages/core/test/evaluation/workspace-config-parsing.test.ts b/packages/core/test/evaluation/workspace-config-parsing.test.ts
index 5779b606d..22685b1b7 100644
--- a/packages/core/test/evaluation/workspace-config-parsing.test.ts
+++ b/packages/core/test/evaluation/workspace-config-parsing.test.ts
@@ -188,6 +188,57 @@ tests:
     expect(cases[0].workspace?.template).toBe(path.join(testDir, 'workspace-template'));
   });
 
+  it('should parse docker workspace base_commit', async () => {
+    const evalFile = path.join(testDir, 'workspace-docker-base-commit.yaml');
+    await writeFile(
+      evalFile,
+      `
+tests:
+  - id: docker-base-commit
+    input: "Do something"
+    criteria: "Should work"
+    workspace:
+      docker:
+        image: swebench/sweb.eval.django__django:latest
+        base_commit: abc123def
+`,
+    );
+
+    const cases = await loadTests(evalFile, testDir);
+    expect(cases).toHaveLength(1);
+    expect(cases[0].workspace?.docker).toEqual({
+      image: 'swebench/sweb.eval.django__django:latest',
+      base_commit: 'abc123def',
+    });
+  });
+
+  it('should parse repo checkout base_commit', async () => {
+    const evalFile = path.join(testDir, 'workspace-repo-base-commit.yaml');
+    await writeFile(
+      evalFile,
+      `
+tests:
+  - id: repo-base-commit
+    input: "Do something"
+    criteria: "Should work"
+    workspace:
+      repos:
+        - path: /testbed
+          source:
+            type: git
+            url: https://github.com/org/repo.git
+          checkout:
+            base_commit: abc123def
+`,
+    );
+
+    const cases = await loadTests(evalFile, testDir);
+    expect(cases).toHaveLength(1);
+    expect(cases[0].workspace?.repos?.[0].checkout).toEqual({
+      base_commit: 'abc123def',
+    });
+  });
+
   it('parses workspace repos from YAML', async () => {
     const evalFile = path.join(testDir, 'workspace-repos.yaml');
     await writeFile(
diff --git a/packages/core/test/evaluation/workspace/docker-workspace.test.ts b/packages/core/test/evaluation/workspace/docker-workspace.test.ts
index 08bff49d2..2b48aca60 100644
--- a/packages/core/test/evaluation/workspace/docker-workspace.test.ts
+++ b/packages/core/test/evaluation/workspace/docker-workspace.test.ts
@@ -181,6 +181,121 @@ describe('DockerWorkspaceProvider', () => {
     });
   });
 
+  describe('resetContainerCheckout', () => {
+    it('resets the container to repo checkout targets and verifies HEAD', async () => {
+      executor.pushResponse({ exitCode: 0 }); // docker start
+      executor.pushResponse({ exitCode: 0 }); // git reset --hard
+      executor.pushResponse({ stdout: 'abc123\n', exitCode: 0 }); // git rev-parse HEAD
+
+      const provider = new DockerWorkspaceProvider({ image: 'img:1' }, executor);
+
+      await provider.startContainer('container-1');
+      await provider.resetContainerCheckout('container-1', [{ path: '/testbed', ref: 'abc123' }]);
+
+      expect(executor.callArgv(1)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        '-C',
+        '/testbed',
+        'reset',
+        '--hard',
+        'abc123',
+      ]);
+      expect(executor.callArgv(2)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        '-C',
+        '/testbed',
+        'rev-parse',
+        'HEAD',
+      ]);
+    });
+
+    it('skips reset when base_commit is not set', async () => {
+      const provider = new DockerWorkspaceProvider({ image: 'img:1' }, executor);
+      await provider.resetContainerCheckout('container-1');
+      expect(executor.calls).toHaveLength(0);
+    });
+
+    it('falls back to deprecated docker base_commit when repo checkout targets are absent', async () => {
+      executor.pushResponse({ exitCode: 0 }); // git reset --hard
+      executor.pushResponse({ stdout: 'abc123\n', exitCode: 0 }); // git rev-parse HEAD
+
+      const provider = new DockerWorkspaceProvider(
+        { image: 'img:1', base_commit: 'abc123' },
+        executor,
+      );
+
+      await provider.resetContainerCheckout('container-1');
+
+      expect(executor.callArgv(0)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        'reset',
+        '--hard',
+        'abc123',
+      ]);
+      expect(executor.callArgv(1)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        'rev-parse',
+        'HEAD',
+      ]);
+    });
+  });
+
+  describe('runGraderInContainer', () => {
+    it('resets the container to repo checkout targets before running the grader', async () => {
+      executor.pushResponse({ stdout: 'container-1\n', exitCode: 0 }); // docker create
+      executor.pushResponse({ exitCode: 0 }); // docker start
+      executor.pushResponse({ exitCode: 0 }); // git reset --hard
+      executor.pushResponse({ stdout: 'abc123\n', exitCode: 0 }); // git rev-parse HEAD
+      executor.pushResponse({ stdout: '{"score": 1, "assertions": []}', exitCode: 0 }); // grader
+      executor.pushResponse({ exitCode: 0 }); // docker rm -f
+
+      const provider = new DockerWorkspaceProvider({ image: 'img:1' }, executor);
+
+      const result = await provider.runGraderInContainer({
+        command: ['python', 'grade.py'],
+        repoCheckouts: [{ path: '/testbed', ref: 'abc123' }],
+      });
+
+      expect(result.exitCode).toBe(0);
+      expect(executor.callArgv(0)).toEqual(['docker', 'create', 'img:1', 'sleep', 'infinity']);
+      expect(executor.callArgv(1)).toEqual(['docker', 'start', 'container-1']);
+      expect(executor.callArgv(2)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        '-C',
+        '/testbed',
+        'reset',
+        '--hard',
+        'abc123',
+      ]);
+      expect(executor.callArgv(3)).toEqual([
+        'docker',
+        'exec',
+        'container-1',
+        'git',
+        '-C',
+        '/testbed',
+        'rev-parse',
+        'HEAD',
+      ]);
+      expect(executor.callArgv(4)).toEqual(['docker', 'exec', 'container-1', 'python', 'grade.py']);
+    });
+  });
+
   describe('copyToContainer', () => {
     it('copies local path to container path', async () => {
       executor.pushResponse({ exitCode: 0 });
diff --git a/plugins/agentv-dev/skills/agentv-eval-writer/SKILL.md b/plugins/agentv-dev/skills/agentv-eval-writer/SKILL.md
index 754346dba..05d2e3403 100644
--- a/plugins/agentv-dev/skills/agentv-eval-writer/SKILL.md
+++ b/plugins/agentv-dev/skills/agentv-eval-writer/SKILL.md
@@ -295,16 +295,23 @@ tests:
     criteria: Bug is fixed
     metadata:
       repo: sympy/sympy
-      base_commit: "abc123"
     workspace:
-      setup:
-        command: ["python", "custom-setup.py"]  # overrides suite-level
+      repos:
+        - path: /testbed
+          source:
+            type: git
+            url: https://github.com/sympy/sympy.git
+          checkout:
+            base_commit: "abc123"
+      docker:
+        image: swebench/sweb.eval.django__django:latest
 ```
 
 **Lifecycle:** template copy → repo clone → setup → git baseline → agent → file changes → teardown → repo reset → cleanup
 **Merge:** Case-level fields replace suite-level fields.
 **Commands receive stdin JSON:** `{workspace_path, test_id, eval_run_id, case_input, case_metadata}`
 **Setup failure:** aborts case. **Teardown failure:** non-fatal (warning).
+For SWE-bench-style evals, keep operational checkout state under `workspace.repos[].checkout.base_commit`; treat `metadata.base_commit` as informational only. `workspace.docker.base_commit` remains a deprecated compatibility bridge for legacy Docker-backed evals.
 
 ### Repository Lifecycle
 
diff --git a/plugins/agentv-dev/skills/agentv-eval-writer/references/eval-schema.json b/plugins/agentv-dev/skills/agentv-eval-writer/references/eval-schema.json
index 1f96f0f2b..d261b0bdb 100644
--- a/plugins/agentv-dev/skills/agentv-eval-writer/references/eval-schema.json
+++ b/plugins/agentv-dev/skills/agentv-eval-writer/references/eval-schema.json
@@ -517,6 +517,17 @@
                           "required": ["type"],
                           "additionalProperties": false
                         },
+                        {
+                          "type": "object",
+                          "properties": {
+                            "include": {
+                              "type": "string",
+                              "minLength": 1
+                            }
+                          },
+                          "required": ["include"],
+                          "additionalProperties": false
+                        },
                         {
                           "type": "object",
                           "properties": {
@@ -1645,6 +1656,17 @@
                           "required": ["type"],
                           "additionalProperties": false
                         },
+                        {
+                          "type": "object",
+                          "properties": {
+                            "include": {
+                              "type": "string",
+                              "minLength": 1
+                            }
+                          },
+                          "required": ["include"],
+                          "additionalProperties": false
+                        },
                         {
                           "type": "object",
                           "properties": {
@@ -2790,6 +2812,17 @@
                               "required": ["type"],
                               "additionalProperties": false
                             },
+                            {
+                              "type": "object",
+                              "properties": {
+                                "include": {
+                                  "type": "string",
+                                  "minLength": 1
+                                }
+                              },
+                              "required": ["include"],
+                              "additionalProperties": false
+                            },
                             {
                               "type": "object",
                               "properties": {
@@ -3918,6 +3951,17 @@
                               "required": ["type"],
                               "additionalProperties": false
                             },
+                            {
+                              "type": "object",
+                              "properties": {
+                                "include": {
+                                  "type": "string",
+                                  "minLength": 1
+                                }
+                              },
+                              "required": ["include"],
+                              "additionalProperties": false
+                            },
                             {
                               "type": "object",
                               "properties": {
@@ -4846,6 +4890,10 @@
                                 "ref": {
                                   "type": "string"
                                 },
+                                "base_commit": {
+                                  "type": "string",
+                                  "minLength": 1
+                                },
                                 "resolve": {
                                   "type": "string",
                                   "enum": ["remote", "local"]
@@ -5037,6 +5085,10 @@
                           "cpus": {
                             "type": "number",
                             "minimum": 0.1
+                          },
+                          "base_commit": {
+                            "type": "string",
+                            "minLength": 1
                           }
                         },
                         "required": ["image"],
@@ -5490,6 +5542,17 @@
                           "required": ["type"],
                           "additionalProperties": false
                         },
+                        {
+                          "type": "object",
+                          "properties": {
+                            "include": {
+                              "type": "string",
+                              "minLength": 1
+                            }
+                          },
+                          "required": ["include"],
+                          "additionalProperties": false
+                        },
                         {
                           "type": "object",
                           "properties": {
@@ -6618,6 +6681,17 @@
                           "required": ["type"],
                           "additionalProperties": false
                         },
+                        {
+                          "type": "object",
+                          "properties": {
+                            "include": {
+                              "type": "string",
+                              "minLength": 1
+                            }
+                          },
+                          "required": ["include"],
+                          "additionalProperties": false
+                        },
                         {
                           "type": "object",
                           "properties": {
@@ -7763,6 +7837,17 @@
                               "required": ["type"],
                               "additionalProperties": false
                             },
+                            {
+                              "type": "object",
+                              "properties": {
+                                "include": {
+                                  "type": "string",
+                                  "minLength": 1
+                                }
+                              },
+                              "required": ["include"],
+                              "additionalProperties": false
+                            },
                             {
                               "type": "object",
                               "properties": {
@@ -8891,6 +8976,17 @@
                               "required": ["type"],
                               "additionalProperties": false
                             },
+                            {
+                              "type": "object",
+                              "properties": {
+                                "include": {
+                                  "type": "string",
+                                  "minLength": 1
+                                }
+                              },
+                              "required": ["include"],
+                              "additionalProperties": false
+                            },
                             {
                               "type": "object",
                               "properties": {
@@ -9819,6 +9915,10 @@
                                 "ref": {
                                   "type": "string"
                                 },
+                                "base_commit": {
+                                  "type": "string",
+                                  "minLength": 1
+                                },
                                 "resolve": {
                                   "type": "string",
                                   "enum": ["remote", "local"]
@@ -10010,6 +10110,10 @@
                           "cpus": {
                             "type": "number",
                             "minimum": 0.1
+                          },
+                          "base_commit": {
+                            "type": "string",
+                            "minLength": 1
                           }
                         },
                         "required": ["image"],
@@ -10369,6 +10473,17 @@
                     "required": ["type"],
                     "additionalProperties": false
                   },
+                  {
+                    "type": "object",
+                    "properties": {
+                      "include": {
+                        "type": "string",
+                        "minLength": 1
+                      }
+                    },
+                    "required": ["include"],
+                    "additionalProperties": false
+                  },
                   {
                     "type": "object",
                     "properties": {
@@ -11497,6 +11612,17 @@
                     "required": ["type"],
                     "additionalProperties": false
                   },
+                  {
+                    "type": "object",
+                    "properties": {
+                      "include": {
+                        "type": "string",
+                        "minLength": 1
+                      }
+                    },
+                    "required": ["include"],
+                    "additionalProperties": false
+                  },
                   {
                     "type": "object",
                     "properties": {
@@ -12676,6 +12802,17 @@
                 "required": ["type"],
                 "additionalProperties": false
               },
+              {
+                "type": "object",
+                "properties": {
+                  "include": {
+                    "type": "string",
+                    "minLength": 1
+                  }
+                },
+                "required": ["include"],
+                "additionalProperties": false
+              },
               {
                 "type": "object",
                 "properties": {
@@ -13582,6 +13719,10 @@
                           "ref": {
                             "type": "string"
                           },
+                          "base_commit": {
+                            "type": "string",
+                            "minLength": 1
+                          },
                           "resolve": {
                             "type": "string",
                             "enum": ["remote", "local"]
@@ -13773,6 +13914,10 @@
                     "cpus": {
                       "type": "number",
                       "minimum": 0.1
+                    },
+                    "base_commit": {
+                      "type": "string",
+                      "minLength": 1
                     }
                   },
                   "required": ["image"],
diff --git a/scripts/import-huggingface.py b/scripts/import-huggingface.py
index fd5208230..66244425b 100644
--- a/scripts/import-huggingface.py
+++ b/scripts/import-huggingface.py
@@ -21,7 +21,9 @@
 SWE-bench field mapping:
     instance_id        -> test id
     problem_statement  -> input (user message)
-    repo + base_commit -> workspace.docker metadata
+    repo               -> metadata.repo
+    instance_id        -> workspace.docker.image (ghcr.io/epoch-research/swe-bench.eval.x86_64.<id>:latest)
+    base_commit        -> legacy workspace.docker compatibility bridge
     FAIL_TO_PASS       -> assertions (code-grader commands)
     difficulty         -> metadata.difficulty
 
@@ -77,15 +79,13 @@ def _to_eval_name(instance_id: str) -> str:
     return name.strip("-")
 
 
-def _docker_image_for_repo(repo: str) -> str:
-    """Derive a Docker image name for a SWE-bench repo.
+def _docker_image_for_instance(instance_id: str) -> str:
+    """Derive a Docker image name for a SWE-bench instance.
 
-    Uses the swebench Docker image naming convention:
-    swebench/sweb.eval.<owner>__<repo>:latest
+    Uses the Epoch Research GHCR naming convention:
+    ghcr.io/epoch-research/swe-bench.eval.x86_64.<instance_id>:latest
     """
-    # repo format: "owner/name" e.g. "django/django"
-    safe = repo.replace("/", "__")
-    return f"swebench/sweb.eval.{safe}:latest"
+    return f"ghcr.io/epoch-research/swe-bench.eval.x86_64.{instance_id}:latest"
 
 
 def _parse_test_list(value: Any) -> list[str]:
@@ -120,13 +120,15 @@ def _convert_swebench_instance(row: dict[str, Any]) -> dict[str, Any]:
     difficulty = row.get("difficulty")
 
     # Build assertions from FAIL_TO_PASS and PASS_TO_PASS test names
+    # SWE-bench Docker images use a conda env named "testbed" for all repo dependencies.
+    # Commands must run through `conda run -n testbed` to pick up pytest and the installed repo.
     assertions: list[dict[str, Any]] = []
     if fail_to_pass:
         # Code-grader that runs the previously-failing tests (should pass after patch)
         assertions.append({
             "type": "code-grader",
             "command": [
-                "python", "-c",
+                "conda", "run", "-n", "testbed", "python", "-c",
                 (
                     "import subprocess, sys, json; "
                     f"result = subprocess.run({json.dumps(['python', '-m', 'pytest'] + fail_to_pass)}, "
@@ -143,7 +145,7 @@ def _convert_swebench_instance(row: dict[str, Any]) -> dict[str, Any]:
         assertions.append({
             "type": "code-grader",
             "command": [
-                "python", "-c",
+                "conda", "run", "-n", "testbed", "python", "-c",
                 (
                     "import subprocess, sys, json; "
                     f"result = subprocess.run({json.dumps(['python', '-m', 'pytest'] + pass_to_pass)}, "
@@ -165,7 +167,7 @@ def _convert_swebench_instance(row: dict[str, Any]) -> dict[str, Any]:
     if assertions:
         test_case["assertions"] = assertions
 
-    # Add metadata (informational only — repo/base_commit are in workspace.docker)
+    # Add metadata (informational only)
     metadata: dict[str, Any] = {}
     if repo:
         metadata["repo"] = repo
@@ -181,11 +183,11 @@ def _convert_swebench_instance(row: dict[str, Any]) -> dict[str, Any]:
     }
 
     # Docker workspace config
-    # base_commit is part of the workspace, not metadata — the container must be
-    # checked out at this commit for the patch to apply and tests to match.
+    # base_commit remains here as a compatibility bridge until Docker-backed
+    # prebuilt images consume workspace.repos[].checkout directly.
     if repo:
         docker_config: dict[str, Any] = {
-            "image": _docker_image_for_repo(repo),
+            "image": _docker_image_for_instance(instance_id),
             "timeout": 600,
             "memory": "4g",
         }