diff --git a/docs/WIKI_BENCH.md b/docs/WIKI_BENCH.md
new file mode 100644
index 000000000..d4fbdd24d
--- /dev/null
+++ b/docs/WIKI_BENCH.md
@@ -0,0 +1,117 @@
+# wiki-bench v0
+
+Retrieval quality benchmark over Tobi Lütke’s public-appearances wiki
+(`tobi/wiki`). Ground-truth labels come from `meta/registry.json` and page
+YAML frontmatter only — **no LLM-as-judge**.
+
+## Privacy: fixture only in this repo
+
+**This public repository ships the fixture (queries + expected paths/ids)
+only.** It does **not** vendor wiki page bodies.
+
+| Artifact | Path | In public repo? |
+|----------|------|-----------------|
+| Fixture (queries + expected paths) | `src/bench/fixtures/wiki-v0.json` | Yes |
+| Fixture schema test | `test/wiki-bench-fixture.test.ts` | Yes |
+| Optional BM25 floors (env-gated) | `test/wiki-bench-bm25.test.ts` | Yes (skips without corpus) |
+| Local BM25 runner (temp index) | `scripts/run-wiki-bench-local.mjs` | Yes (requires local corpus) |
+| Wiki markdown corpus | private `tobi/wiki` (or a local checkout) | **No — never commit** |
+
+Do **not** add `test/wiki-bench-docs/**` or any wiki `.md` bodies to this repo.
+CI must stay green without access to the private corpus.
+
+Expected file paths in the fixture are **wiki-relative**:
+`concepts|sources|entities|syntheses/...md` (collection rooted at the wiki
+markdown directory, typically `wiki/` inside `tobi/wiki`).
+
+## Validate the fixture (no corpus required)
+
+```bash
+npx vitest run test/wiki-bench-fixture.test.ts
+# or
+bun test test/wiki-bench-fixture.test.ts
+```
+
+This checks version/collection, query shape, allowed `type` values
+(`exact|semantic|topical|cross-domain|alias`), unique ids, and that
+`expected_files` are safe wiki-relative `.md` paths.
+
+## Index a local wiki checkout
+
+Point QMD at your local clone of the private wiki. Markdown usually lives
+under `wiki/`:
+
+```bash
+# If markdown is under <wiki-repo>/wiki/{concepts,sources,entities,...}:
+qmd collection add /path/to/wiki/wiki --name wiki-bench
+
+# If you already keep a bare wiki-root directory of those folders:
+qmd collection add /path/to/wiki-root --name wiki-bench
+```
+
+Then update/embed as usual for your install (`qmd update`, etc.).
+
+## Run the full multi-backend bench
+
+`qmd bench` is wired in `src/cli/qmd.ts` and implemented by
+`src/bench/bench.ts`. It loads a fixture JSON and scores bm25 / vector /
+hybrid / full against an already-indexed collection:
+
+```bash
+qmd bench src/bench/fixtures/wiki-v0.json -c wiki-bench
+qmd bench src/bench/fixtures/wiki-v0.json -c wiki-bench --json
+```
+
+Usage (from CLI): `qmd bench <fixture.json> [--json] [-c collection]`.
+
+## Local BM25 runner (private corpus)
+
+`scripts/run-wiki-bench-local.mjs` indexes a **temporary** collection named
+`wiki-bench` (under `$TMPDIR`; never writes into the git tree), runs BM25 via
+`runBenchmark`, prints summary metrics, and exits non-zero if quality floors
+fail. Same env vars / floors as the skip-gated vitest suite:
+
+```bash
+QMD_WIKI_PATH=~/src/wiki node scripts/run-wiki-bench-local.mjs
+# or:
+QMD_WIKI_BENCH_DOCS=/path/to/wiki/wiki node scripts/run-wiki-bench-local.mjs
+```
+
+- `QMD_WIKI_PATH` — checkout of private `tobi/wiki`; docs root is
+  `$QMD_WIKI_PATH/wiki`
+- `QMD_WIKI_BENCH_DOCS` — directory already laid out as
+  `concepts|sources|entities/...`
+
+Exits with a clear error if neither env var is set or the path is missing.
+Does **not** copy or commit wiki markdown into this repository.
+
+## Optional BM25 quality floors (local corpus)
+
+`test/wiki-bench-bm25.test.ts` indexes a **local** corpus into a temp DB and
+asserts floors on `exact` + `alias` queries. It **skips** unless one of these
+env vars points at an existing directory:
+
+- `QMD_WIKI_BENCH_DOCS` — preferred; directory that contains
+  `concepts/`, `sources/`, etc.
+- `QMD_WIKI_PATH` — wiki repo root or wiki markdown root; if a nested `wiki/`
+  subdirectory contains those folders, that nested path is used.
+
+```bash
+export QMD_WIKI_BENCH_DOCS=/path/to/wiki/wiki
+npx vitest run test/wiki-bench-bm25.test.ts
+```
+
+Without the env var (default CI), the suite skips and stays green.
+
+Asserted floors (measured baseline when corpus is present):
+
+- exact: mean recall@3 ≥ 0.85
+- alias: mean recall@3 ≥ 0.80
+- exact+alias: mean MRR ≥ 0.70
+
+## Notes
+
+- Eval/benchmark scope only — this does not change search ranking code.
+- Labels were derived from registry titles + frontmatter (`title`, `sources`,
+  entity names), never from an LLM judge.
+- Never commit private wiki `.md` content into `tobi/qmd`.
diff --git a/scripts/run-wiki-bench-local.mjs b/scripts/run-wiki-bench-local.mjs
new file mode 100755
index 000000000..db4a72e2f
--- /dev/null
+++ b/scripts/run-wiki-bench-local.mjs
@@ -0,0 +1,57 @@
+#!/usr/bin/env node
+/**
+ * wiki-bench v0 — private/local BM25 runner.
+ *
+ * Requires a local corpus via env (never vendors wiki bodies into the repo):
+ *   QMD_WIKI_PATH       — checkout of tobi/wiki; uses $QMD_WIKI_PATH/wiki
+ *   QMD_WIKI_BENCH_DOCS — markdown root with concepts|sources|entities/...
+ *
+ * Builds a temporary QMD index/collection named `wiki-bench`, runs BM25 against
+ * src/bench/fixtures/wiki-v0.json via runBenchmark, prints metrics, and exits
+ * non-zero if exact R@3 < 0.85, alias R@3 < 0.80, or exact+alias MRR < 0.70.
+ *
+ * Usage:
+ *   QMD_WIKI_PATH=~/src/wiki node scripts/run-wiki-bench-local.mjs
+ */
+import { spawnSync } from "node:child_process";
+import { existsSync } from "node:fs";
+import { join } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const root = fileURLToPath(new URL("..", import.meta.url));
+const helper = join(root, "scripts", "wiki-bench-local.ts");
+
+function resolveRunner() {
+  const tsx = join(root, "node_modules", ".bin", "tsx");
+  if (existsSync(tsx)) {
+    return { command: tsx, args: [helper], label: "tsx" };
+  }
+  const bun = spawnSync("bun", ["--version"], { encoding: "utf8" });
+  if (bun.status === 0) {
+    return { command: "bun", args: [helper], label: "bun" };
+  }
+  console.error(
+    "wiki-bench-local: need local TypeScript runner (node_modules/.bin/tsx or bun).",
+  );
+  process.exit(1);
+}
+
+if (!existsSync(helper)) {
+  console.error(`wiki-bench-local: missing helper ${helper}`);
+  process.exit(1);
+}
+
+const runner = resolveRunner();
+const result = spawnSync(runner.command, runner.args, {
+  cwd: root,
+  stdio: "inherit",
+  env: process.env,
+  shell: process.platform === "win32",
+});
+
+if (result.error) {
+  console.error(`wiki-bench-local: failed to launch ${runner.label}:`, result.error.message);
+  process.exit(1);
+}
+
+process.exit(result.status ?? 1);
diff --git a/scripts/wiki-bench-local.ts b/scripts/wiki-bench-local.ts
new file mode 100755
index 000000000..0830f725f
--- /dev/null
+++ b/scripts/wiki-bench-local.ts
@@ -0,0 +1,211 @@
+/**
+ * wiki-bench v0 — private/local BM25 runner body.
+ *
+ * Invoked by scripts/run-wiki-bench-local.mjs. Builds a temporary QMD index
+ * (collection `wiki-bench`) outside the git tree, runs BM25 via runBenchmark,
+ * prints metrics, and exits non-zero if quality floors fail.
+ *
+ * Never vendors wiki page bodies into the repo.
+ */
+
+import {
+  existsSync,
+  mkdtempSync,
+  rmSync,
+  statSync,
+} from "node:fs";
+import { join, resolve } from "node:path";
+import { tmpdir } from "node:os";
+import { fileURLToPath } from "node:url";
+import { createStore } from "../src/index.js";
+import { runBenchmark } from "../src/bench/bench.js";
+import type { BenchmarkResult, QueryResult } from "../src/bench/types.js";
+
+const COLLECTION = "wiki-bench";
+const FLOOR_EXACT_RECALL_AT_3 = 0.85;
+const FLOOR_ALIAS_RECALL_AT_3 = 0.80;
+const FLOOR_COMBINED_MRR = 0.70;
+
+const root = fileURLToPath(new URL("..", import.meta.url));
+const FIXTURE_PATH = join(root, "src", "bench", "fixtures", "wiki-v0.json");
+
+function die(message: string, code = 1): never {
+  console.error(`wiki-bench-local: ${message}`);
+  process.exit(code);
+}
+
+function looksLikeWikiDocsRoot(dir: string): boolean {
+  return existsSync(join(dir, "concepts")) || existsSync(join(dir, "sources"));
+}
+
+/** Resolve corpus root from env. Never looks under test/wiki-bench-docs. */
+function resolveWikiDocsDir(): string {
+  const benchDocs = process.env.QMD_WIKI_BENCH_DOCS?.trim();
+  if (benchDocs) {
+    const abs = resolve(benchDocs);
+    if (!existsSync(abs) || !statSync(abs).isDirectory()) {
+      die(`QMD_WIKI_BENCH_DOCS is set but not a directory: ${abs}`);
+    }
+    if (!looksLikeWikiDocsRoot(abs)) {
+      die(
+        `QMD_WIKI_BENCH_DOCS must contain wiki-relative folders (concepts|sources|...): ${abs}`,
+      );
+    }
+    return abs;
+  }
+
+  const wikiPath = process.env.QMD_WIKI_PATH?.trim();
+  if (wikiPath) {
+    const abs = resolve(wikiPath);
+    if (!existsSync(abs) || !statSync(abs).isDirectory()) {
+      die(`QMD_WIKI_PATH is set but not a directory: ${abs}`);
+    }
+    // Spec: use $QMD_WIKI_PATH/wiki (Obsidian wiki/ folder) as the docs root.
+    const nested = join(abs, "wiki");
+    if (!existsSync(nested) || !statSync(nested).isDirectory()) {
+      die(
+        `QMD_WIKI_PATH set but missing Obsidian wiki/ folder at: ${nested}\n` +
+          `  (expected layout: $QMD_WIKI_PATH/wiki/{concepts,sources,entities,...})`,
+      );
+    }
+    if (!looksLikeWikiDocsRoot(nested)) {
+      die(
+        `QMD_WIKI_PATH/wiki must contain wiki-relative folders (concepts|sources|...): ${nested}`,
+      );
+    }
+    return nested;
+  }
+
+  die(
+    "Set QMD_WIKI_PATH (checkout of tobi/wiki) or QMD_WIKI_BENCH_DOCS (markdown root with concepts|sources|...).\n" +
+      "  Example: QMD_WIKI_PATH=~/src/wiki node scripts/run-wiki-bench-local.mjs\n" +
+      "Private wiki page bodies must never be committed into this repo.",
+  );
+}
+
+function mean(values: number[]): number {
+  if (values.length === 0) return 0;
+  return values.reduce((a, b) => a + b, 0) / values.length;
+}
+
+function bm25Metrics(results: QueryResult[], type: string): { r3: number; mrr: number; n: number } {
+  const subset = results.filter((r) => r.type === type);
+  const scores = subset
+    .map((r) => r.backends.bm25)
+    .filter((b): b is NonNullable<typeof b> => !!b);
+  return {
+    r3: mean(scores.map((s) => s.recall_at_3)),
+    mrr: mean(scores.map((s) => s.mrr)),
+    n: scores.length,
+  };
+}
+
+function printFloorSummary(result: BenchmarkResult): {
+  exactR3: number;
+  aliasR3: number;
+  combinedMrr: number;
+} {
+  const exact = bm25Metrics(result.results, "exact");
+  const alias = bm25Metrics(result.results, "alias");
+  const combinedScores = result.results
+    .filter((r) => r.type === "exact" || r.type === "alias")
+    .map((r) => r.backends.bm25)
+    .filter((b): b is NonNullable<typeof b> => !!b);
+  const combinedMrr = mean(combinedScores.map((s) => s.mrr));
+
+  console.log("\n[wiki-bench-local] BM25 quality floors (exact + alias):");
+  console.log(
+    `  exact mean recall@3 = ${exact.r3.toFixed(4)} (n=${exact.n}; floor ${FLOOR_EXACT_RECALL_AT_3})`,
+  );
+  console.log(
+    `  alias mean recall@3 = ${alias.r3.toFixed(4)} (n=${alias.n}; floor ${FLOOR_ALIAS_RECALL_AT_3})`,
+  );
+  console.log(
+    `  exact+alias mean MRR = ${combinedMrr.toFixed(4)} (n=${combinedScores.length}; floor ${FLOOR_COMBINED_MRR})`,
+  );
+
+  return { exactR3: exact.r3, aliasR3: alias.r3, combinedMrr };
+}
+
+async function main(): Promise<void> {
+  if (!existsSync(FIXTURE_PATH)) {
+    die(`fixture not found: ${FIXTURE_PATH}`);
+  }
+
+  const docsDir = resolveWikiDocsDir();
+  const tempDir = mkdtempSync(join(tmpdir(), "qmd-wiki-bench-local-"));
+  const dbPath = join(tempDir, "wiki-bench.sqlite");
+
+  console.log(`[wiki-bench-local] docs: ${docsDir}`);
+  console.log(`[wiki-bench-local] temp index: ${dbPath}`);
+  console.log(`[wiki-bench-local] fixture: ${FIXTURE_PATH}`);
+
+  let exitCode = 0;
+  try {
+    const store = await createStore({
+      dbPath,
+      config: {
+        collections: {
+          [COLLECTION]: { path: docsDir, pattern: "**/*.md" },
+        },
+      },
+    });
+
+    const updateResult = await store.update({ collections: [COLLECTION] });
+    await store.close();
+
+    const indexed =
+      updateResult.indexed + updateResult.updated + updateResult.unchanged;
+    if (indexed <= 0) {
+      die(`no markdown documents indexed from ${docsDir}`);
+    }
+    console.log(
+      `[wiki-bench-local] indexed collection '${COLLECTION}': ` +
+        `${updateResult.indexed} new, ${updateResult.updated} updated, ` +
+        `${updateResult.unchanged} unchanged (${indexed} total)`,
+    );
+
+    const benchResult = await runBenchmark(FIXTURE_PATH, {
+      backends: ["bm25"],
+      collection: COLLECTION,
+      dbPath,
+    });
+
+    const { exactR3, aliasR3, combinedMrr } = printFloorSummary(benchResult);
+
+    const failures: string[] = [];
+    if (exactR3 < FLOOR_EXACT_RECALL_AT_3) {
+      failures.push(
+        `exact R@3 ${exactR3.toFixed(4)} < ${FLOOR_EXACT_RECALL_AT_3}`,
+      );
+    }
+    if (aliasR3 < FLOOR_ALIAS_RECALL_AT_3) {
+      failures.push(
+        `alias R@3 ${aliasR3.toFixed(4)} < ${FLOOR_ALIAS_RECALL_AT_3}`,
+      );
+    }
+    if (combinedMrr < FLOOR_COMBINED_MRR) {
+      failures.push(
+        `exact+alias MRR ${combinedMrr.toFixed(4)} < ${FLOOR_COMBINED_MRR}`,
+      );
+    }
+
+    if (failures.length > 0) {
+      console.error("\nwiki-bench-local: FAILED quality floors:");
+      for (const f of failures) console.error(`  - ${f}`);
+      exitCode = 1;
+    } else {
+      console.log("\nwiki-bench-local: PASSED quality floors.");
+    }
+  } finally {
+    rmSync(tempDir, { recursive: true, force: true });
+  }
+
+  process.exit(exitCode);
+}
+
+main().catch((err) => {
+  console.error("wiki-bench-local: unexpected error");
+  console.error(err);
+  process.exit(1);
+});
diff --git a/src/bench/fixtures/wiki-v0.json b/src/bench/fixtures/wiki-v0.json
new file mode 100644
index 000000000..d6398f0cc
--- /dev/null
+++ b/src/bench/fixtures/wiki-v0.json
@@ -0,0 +1,365 @@
+{
+  "description": "wiki-bench v0: retrieval over Tobi Lütke public-appearances wiki. Ground truth from meta/registry.json + page YAML frontmatter only (no LLM judge). Collection rooted at wiki/ so paths are concepts|sources|entities/*.md.",
+  "version": 1,
+  "collection": "wiki-bench",
+  "queries": [
+    {
+      "id": "src-kp41",
+      "query": "Knowledge Project Episode 41 Trust Battery",
+      "type": "exact",
+      "description": "Exact show + episode title → source page (registry title)",
+      "expected_files": ["sources/2018-knowledge-project.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-hibt",
+      "query": "How I Built This Guy Raz Shopify",
+      "type": "exact",
+      "description": "Show + host → 2019 NPR biography source",
+      "expected_files": ["sources/2019-how-i-built-this.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-kara",
+      "query": "Kara Swisher Startupfest 2016 Rise of Shopify",
+      "type": "exact",
+      "description": "Host + event + year → source",
+      "expected_files": ["sources/2016-kara-swisher.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-mos",
+      "query": "Masters of Scale Reid Hoffman be a platform",
+      "type": "exact",
+      "description": "Show + host → 2020 Masters of Scale source",
+      "expected_files": ["sources/2020-masters-of-scale-reid-hoffman.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-motley-2017",
+      "query": "Motley Fool Tom Gardner 2017 trust battery",
+      "type": "exact",
+      "description": "Show + host + year → source that first names trust battery",
+      "expected_files": ["sources/2017-motley-fool-tom-gardner.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-leo",
+      "query": "Leo Laporte Call for Help Liquid 2006",
+      "type": "exact",
+      "description": "Early TV interview → source",
+      "expected_files": ["sources/2006-leo-laporte.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-mixergy",
+      "query": "Mixergy Andrew Warner Shopify profitable",
+      "type": "exact",
+      "description": "Show + host → 2010 Mixergy source",
+      "expected_files": ["sources/2010-mixergy.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-twist",
+      "query": "This Week in Startups Jason Calacanis Ottawa",
+      "type": "exact",
+      "description": "Show + host → 2013 TWiST source",
+      "expected_files": ["sources/2013-this-week-in-startups.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-balaji",
+      "query": "Network State Balaji country-sized economy",
+      "type": "exact",
+      "description": "Show + host + phrase → 2023 source",
+      "expected_files": ["sources/2023-network-state-balaji.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "src-iltb173",
+      "query": "Invest Like the Best 173 Patrick O'Shaughnessy",
+      "type": "exact",
+      "description": "Show + episode number → source",
+      "expected_files": ["sources/2020-invest-like-the-best-173.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-trust-battery",
+      "query": "trust battery",
+      "type": "exact",
+      "description": "Concept title exact match",
+      "expected_files": ["concepts/trust-battery.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-arming-rebels",
+      "query": "arming the rebels",
+      "type": "exact",
+      "description": "Concept title exact match",
+      "expected_files": ["concepts/arming-the-rebels.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-antifragility",
+      "query": "antifragility Taleb",
+      "type": "exact",
+      "description": "Concept + author cue from tldr",
+      "expected_files": ["concepts/antifragility.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-liquid",
+      "query": "Liquid template language Shopify",
+      "type": "exact",
+      "description": "Concept title + tldr keywords",
+      "expected_files": ["concepts/liquid.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-tobi-test",
+      "query": "Tobi test",
+      "type": "exact",
+      "description": "Concept title exact match",
+      "expected_files": ["concepts/tobi-test.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-pendulum",
+      "query": "pendulum model management",
+      "type": "exact",
+      "description": "Concept title + domain word",
+      "expected_files": ["concepts/pendulum-model.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-hundred-year",
+      "query": "hundred year vision",
+      "type": "exact",
+      "description": "Concept title exact match",
+      "expected_files": ["concepts/hundred-year-vision.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-apprenticeship",
+      "query": "apprenticeship German craft",
+      "type": "exact",
+      "description": "Concept + tldr keywords",
+      "expected_files": ["concepts/apprenticeship.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-autonomy",
+      "query": "autonomy NUMMI rip cord",
+      "type": "exact",
+      "description": "Concept + distinctive exhibit from page body/frontmatter framing",
+      "expected_files": ["concepts/autonomy.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "concept-platform-shift",
+      "query": "AI as platform shift",
+      "type": "exact",
+      "description": "Concept title phrasing",
+      "expected_files": ["concepts/ai-as-platform-shift.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "sem-trust-paraphrase",
+      "query": "interpersonal trust that charges with follow-through and buys autonomy",
+      "type": "semantic",
+      "description": "Paraphrase of trust-battery tldr (no title words)",
+      "expected_files": ["concepts/trust-battery.md"],
+      "expected_in_top_k": 3
+    },
+    {
+      "id": "sem-rebels-paraphrase",
+      "query": "give independent merchants platform tools to compete with Amazon",
+      "type": "semantic",
+      "description": "Paraphrase of arming-the-rebels tldr",
+      "expected_files": ["concepts/arming-the-rebels.md"],
+      "expected_in_top_k": 3
+    },
+    {
+      "id": "sem-tobi-test-paraphrase",
+      "query": "randomly turning off production services to train for failure",
+      "type": "semantic",
+      "description": "Paraphrase of tobi-test definition (chaos drill)",
+      "expected_files": ["concepts/tobi-test.md"],
+      "expected_in_top_k": 3
+    },
+    {
+      "id": "sem-calm-progress",
+      "query": "counter-philosophy to hype and calm progress at scale",
+      "type": "semantic",
+      "description": "Paraphrase tied to calm-progress / KP 152 framing",
+      "expected_files": ["concepts/calm-progress.md"],
+      "expected_in_top_k": 3
+    },
+    {
+      "id": "cit-trust-battery-intro",
+      "query": "which interview first named the trust battery metaphor",
+      "type": "topical",
+      "description": "Citation grounding: concept frontmatter sources include 2017 Motley Fool (first naming) and 2018 KP",
+      "expected_files": [
+        "sources/2017-motley-fool-tom-gardner.md",
+        "sources/2018-knowledge-project.md"
+      ],
+      "expected_in_top_k": 2
+    },
+    {
+      "id": "cit-liquid-debut",
+      "query": "interview where Liquid template engine was first demoed",
+      "type": "topical",
+      "description": "Citation: liquid frontmatter sources → 2006 Leo Laporte",
+      "expected_files": ["sources/2006-leo-laporte.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-tobi-test-named",
+      "query": "podcast episode where the Tobi test chaos drill was named",
+      "type": "topical",
+      "description": "Citation: tobi-test frontmatter → 2018 Knowledge Project",
+      "expected_files": ["sources/2018-knowledge-project.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-apprenticeship",
+      "query": "appearance discussing Siemens apprenticeship and German craft education",
+      "type": "topical",
+      "description": "Citation: apprenticeship frontmatter → How I Built This",
+      "expected_files": ["sources/2019-how-i-built-this.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-arming-rebels",
+      "query": "podcast where Tobi said Shopify arms the rebels against Amazon",
+      "type": "topical",
+      "description": "Citation: arming-the-rebels frontmatter → Escape Velocity",
+      "expected_files": ["sources/2019-escape-velocity-dan-martell.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-autonomy-nummi",
+      "query": "talk where Tobi told the NUMMI Toyota autonomy rip cord story",
+      "type": "topical",
+      "description": "Citation: autonomy frontmatter → Business of Software 2011",
+      "expected_files": ["sources/2011-business-of-software.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-calm-progress",
+      "query": "2022 Farnam Street return interview about calm progress",
+      "type": "topical",
+      "description": "Citation: calm-progress frontmatter → KP 152",
+      "expected_files": ["sources/2022-knowledge-project-152.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-country-economy",
+      "query": "conversation with Balaji about Shopify as a country-sized economy",
+      "type": "topical",
+      "description": "Citation: country-sized-economy frontmatter → Network State",
+      "expected_files": ["sources/2023-network-state-balaji.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-pendulum",
+      "query": "Motley Fool interview introducing the pendulum management metaphor",
+      "type": "topical",
+      "description": "Citation: pendulum-model frontmatter → 2017 Motley Fool",
+      "expected_files": ["sources/2017-motley-fool-tom-gardner.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "cit-antifragile",
+      "query": "appearances discussing Taleb antifragile and thriving on change",
+      "type": "topical",
+      "description": "Citation: antifragility frontmatter sources",
+      "expected_files": [
+        "sources/2017-motley-fool-tom-gardner.md",
+        "sources/2016-kara-swisher.md"
+      ],
+      "expected_in_top_k": 2
+    },
+    {
+      "id": "alias-guy-raz",
+      "query": "Guy Raz",
+      "type": "alias",
+      "description": "Person entity by common name",
+      "expected_files": ["entities/p-guy-raz.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-shane-parrish",
+      "query": "Shane Parrish",
+      "type": "alias",
+      "description": "Person entity by common name",
+      "expected_files": ["entities/p-shane-parrish.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-hibt-show",
+      "query": "How I Built This",
+      "type": "alias",
+      "description": "Show entity title",
+      "expected_files": ["entities/s-how-i-built-this.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-knowledge-project",
+      "query": "The Knowledge Project Farnam Street",
+      "type": "alias",
+      "description": "Show entity + alternate name",
+      "expected_files": ["entities/s-the-knowledge-project.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-kara",
+      "query": "Kara Swisher",
+      "type": "alias",
+      "description": "Person entity",
+      "expected_files": ["entities/p-kara-swisher.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-reid",
+      "query": "Reid Hoffman",
+      "type": "alias",
+      "description": "Person entity",
+      "expected_files": ["entities/p-reid-hoffman.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-amazon",
+      "query": "Amazon aggregator",
+      "type": "alias",
+      "description": "Org entity + tldr cue",
+      "expected_files": ["entities/o-amazon.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-calacanis",
+      "query": "Jason Calacanis",
+      "type": "alias",
+      "description": "Person entity",
+      "expected_files": ["entities/p-jason-calacanis.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-shopify",
+      "query": "Shopify",
+      "type": "alias",
+      "description": "Primary org entity",
+      "expected_files": ["entities/shopify.md"],
+      "expected_in_top_k": 1
+    },
+    {
+      "id": "alias-ferriss",
+      "query": "Tim Ferriss",
+      "type": "alias",
+      "description": "Person entity",
+      "expected_files": ["entities/p-tim-ferriss.md"],
+      "expected_in_top_k": 1
+    }
+  ]
+}
diff --git a/test/wiki-bench-bm25.test.ts b/test/wiki-bench-bm25.test.ts
new file mode 100644
index 000000000..a8ebc2200
--- /dev/null
+++ b/test/wiki-bench-bm25.test.ts
@@ -0,0 +1,222 @@
+/**
+ * wiki-bench v0 — optional BM25/FTS quality floors.
+ *
+ * Skips unless QMD_WIKI_BENCH_DOCS or QMD_WIKI_PATH points at an existing
+ * corpus directory (the wiki markdown root containing concepts|sources|entities).
+ * Default CI has neither set, so this suite is a no-op and stays green.
+ *
+ * Does NOT vendor wiki page bodies into this repo.
+ *
+ * Measured BM25 baseline (lex/FTS only, when corpus is present):
+ *   exact mean recall@3:  1.0000 (n=20)
+ *   alias mean recall@3:  0.9000 (n=10)
+ *   exact+alias mean MRR: 0.8811 (n=30)
+ * Floors: 0.85 / 0.80 / 0.70
+ */
+
+import { describe, test, expect, beforeAll, afterAll } from "vitest";
+import {
+  mkdtempSync,
+  rmSync,
+  readFileSync,
+  readdirSync,
+  statSync,
+  existsSync,
+} from "fs";
+import { join, dirname, relative } from "path";
+import { tmpdir } from "os";
+import { createHash } from "crypto";
+import { fileURLToPath } from "url";
+import type { Database } from "../src/db.js";
+import {
+  createStore,
+  searchFTS,
+  insertDocument,
+  insertContent,
+} from "../src/store";
+import { scoreResults } from "../src/bench/score.js";
+import type { BenchmarkFixture, BenchmarkQuery } from "../src/bench/types.js";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const FIXTURE_PATH = join(__dirname, "..", "src", "bench", "fixtures", "wiki-v0.json");
+const COLLECTION = "wiki-bench";
+
+const FLOOR_EXACT_RECALL_AT_3 = 0.85;
+const FLOOR_ALIAS_RECALL_AT_3 = 0.80;
+const FLOOR_COMBINED_MRR = 0.70;
+
+/** Resolve local corpus dir; never looks under test/wiki-bench-docs in-repo. */
+function resolveWikiDocsDir(): string | null {
+  const candidates = [
+    process.env.QMD_WIKI_BENCH_DOCS,
+    process.env.QMD_WIKI_PATH,
+  ].filter((v): v is string => typeof v === "string" && v.length > 0);
+
+  for (const c of candidates) {
+    if (!existsSync(c) || !statSync(c).isDirectory()) continue;
+    if (existsSync(join(c, "concepts")) || existsSync(join(c, "sources"))) {
+      return c;
+    }
+    const nested = join(c, "wiki");
+    if (
+      existsSync(nested) &&
+      (existsSync(join(nested, "concepts")) || existsSync(join(nested, "sources")))
+    ) {
+      return nested;
+    }
+  }
+  return null;
+}
+
+const DOCS_DIR = resolveWikiDocsDir();
+
+function walkMarkdownFiles(dir: string, base: string = dir): string[] {
+  const out: string[] = [];
+  for (const ent of readdirSync(dir, { withFileTypes: true })) {
+    const full = join(dir, ent.name);
+    if (ent.isDirectory()) {
+      out.push(...walkMarkdownFiles(full, base));
+    } else if (ent.isFile() && ent.name.endsWith(".md")) {
+      out.push(relative(base, full));
+    }
+  }
+  return out.sort();
+}
+
+function extractTitle(content: string, fallback: string): string {
+  const fm = content.match(/^---\r?\n([\s\S]*?)\r?\n---/);
+  if (fm) {
+    const m = fm[1]!.match(/^title:\s*(.+)$/m);
+    if (m) {
+      const raw = m[1]!.trim();
+      const quoted = raw.match(/^["'](.*)["']$/);
+      if (quoted) return quoted[1]!.trim() || fallback;
+      return raw || fallback;
+    }
+  }
+  const heading = content.match(/^#\s+(.+)$/m);
+  if (heading?.[1]) return heading[1]!.trim();
+  return fallback;
+}
+
+function mean(values: number[]): number {
+  if (values.length === 0) return 0;
+  return values.reduce((a, b) => a + b, 0) / values.length;
+}
+
+type Scored = {
+  id: string;
+  type: string;
+  recall_at_3: number;
+  mrr: number;
+  top_files: string[];
+};
+
+function scoreQuery(db: Database, q: BenchmarkQuery): Scored {
+  const results = searchFTS(db, q.query, 10, COLLECTION);
+  const files = results.map((r) => r.filepath);
+  const scores = scoreResults(files, q.expected_files, Math.max(q.expected_in_top_k, 3));
+  return {
+    id: q.id,
+    type: q.type,
+    recall_at_3: scores.recall_at_3,
+    mrr: scores.mrr,
+    top_files: files.slice(0, 3),
+  };
+}
+
+describe.skipIf(!DOCS_DIR)("wiki-bench BM25 (FTS, local corpus)", () => {
+  let store: ReturnType<typeof createStore>;
+  let db: Database;
+  let tempDir: string;
+  let exactScores: Scored[];
+  let aliasScores: Scored[];
+  let indexedCount: number;
+
+  beforeAll(() => {
+    if (!DOCS_DIR) return;
+
+    tempDir = mkdtempSync(join(tmpdir(), "qmd-wiki-bench-"));
+    process.env.INDEX_PATH = join(tempDir, "wiki-bench.sqlite");
+
+    store = createStore();
+    db = store.db;
+
+    const files = walkMarkdownFiles(DOCS_DIR);
+    expect(files.length).toBeGreaterThan(0);
+
+    for (const rel of files) {
+      const full = join(DOCS_DIR, rel);
+      expect(statSync(full).size).toBeGreaterThan(0);
+      const content = readFileSync(full, "utf-8");
+      const title = extractTitle(content, rel);
+      const hash = createHash("sha256").update(content).digest("hex").slice(0, 12);
+      const now = new Date().toISOString();
+      insertContent(db, hash, content, now);
+      insertDocument(db, COLLECTION, rel, title, hash, now, now);
+    }
+    indexedCount = files.length;
+
+    const fixture = JSON.parse(readFileSync(FIXTURE_PATH, "utf-8")) as BenchmarkFixture;
+    const exactQueries = fixture.queries.filter((q) => q.type === "exact");
+    const aliasQueries = fixture.queries.filter((q) => q.type === "alias");
+    expect(exactQueries.length).toBeGreaterThan(0);
+    expect(aliasQueries.length).toBeGreaterThan(0);
+
+    exactScores = exactQueries.map((q) => scoreQuery(db, q));
+    aliasScores = aliasQueries.map((q) => scoreQuery(db, q));
+
+    const exactR3 = mean(exactScores.map((s) => s.recall_at_3));
+    const aliasR3 = mean(aliasScores.map((s) => s.recall_at_3));
+    const combinedMrr = mean([...exactScores, ...aliasScores].map((s) => s.mrr));
+
+    console.log("\n[wiki-bench] measured BM25 metrics (local corpus):");
+    console.log(`  docs = ${indexedCount} from ${DOCS_DIR}`);
+    console.log(`  exact mean recall@3 = ${exactR3.toFixed(4)} (n=${exactScores.length})`);
+    console.log(`  alias mean recall@3 = ${aliasR3.toFixed(4)} (n=${aliasScores.length})`);
+    console.log(`  exact+alias mean MRR = ${combinedMrr.toFixed(4)}`);
+  });
+
+  afterAll(() => {
+    store?.close();
+    if (tempDir) rmSync(tempDir, { recursive: true, force: true });
+  });
+
+  test("indexes non-empty wiki pages from local corpus", () => {
+    const count = (
+      db
+        .prepare(`SELECT COUNT(*) as n FROM documents WHERE collection = ? AND active = 1`)
+        .get(COLLECTION) as { n: number }
+    ).n;
+    expect(count).toBe(indexedCount);
+    expect(count).toBeGreaterThan(0);
+  });
+
+  test("exact: mean recall_at_3 meets floor", () => {
+    const exactR3 = mean(exactScores.map((s) => s.recall_at_3));
+    expect(exactR3).toBeGreaterThanOrEqual(FLOOR_EXACT_RECALL_AT_3);
+  });
+
+  test("alias: mean recall_at_3 meets floor", () => {
+    const aliasR3 = mean(aliasScores.map((s) => s.recall_at_3));
+    expect(aliasR3).toBeGreaterThanOrEqual(FLOOR_ALIAS_RECALL_AT_3);
+  });
+
+  test("exact+alias: mean MRR meets floor", () => {
+    const combinedMrr = mean([...exactScores, ...aliasScores].map((s) => s.mrr));
+    expect(combinedMrr).toBeGreaterThanOrEqual(FLOOR_COMBINED_MRR);
+  });
+});
+
+// Explicit always-on guard so CI reports a passing skip reason rather than an empty file.
+describe("wiki-bench BM25 gate", () => {
+  test("skips quality floors unless local corpus env is set", () => {
+    if (!DOCS_DIR) {
+      expect(
+        process.env.QMD_WIKI_BENCH_DOCS || process.env.QMD_WIKI_PATH || "",
+      ).toBe("");
+    } else {
+      expect(existsSync(DOCS_DIR)).toBe(true);
+    }
+  });
+});
diff --git a/test/wiki-bench-fixture.test.ts b/test/wiki-bench-fixture.test.ts
new file mode 100644
index 000000000..e116b03a2
--- /dev/null
+++ b/test/wiki-bench-fixture.test.ts
@@ -0,0 +1,84 @@
+/**
+ * wiki-bench v0 — fixture schema validation (no wiki corpus required).
+ *
+ * Asserts that src/bench/fixtures/wiki-v0.json is well-formed for the
+ * benchmark harness. Does not read or require any wiki page bodies.
+ */
+
+import { describe, test, expect } from "vitest";
+import { readFileSync } from "fs";
+import { join, dirname } from "path";
+import { fileURLToPath } from "url";
+import type { BenchmarkFixture, BenchmarkQuery } from "../src/bench/types.js";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const FIXTURE_PATH = join(__dirname, "..", "src", "bench", "fixtures", "wiki-v0.json");
+
+const ALLOWED_TYPES = new Set([
+  "exact",
+  "semantic",
+  "topical",
+  "cross-domain",
+  "alias",
+] as const);
+
+/** Wiki-relative paths only: concepts|sources|entities|syntheses/...md */
+const WIKI_REL_PATH =
+  /^(concepts|sources|entities|syntheses)\/(?:[\w.-]+\/)*[\w.-]+\.md$/;
+
+describe("wiki-bench fixture (schema only)", () => {
+  const raw = readFileSync(FIXTURE_PATH, "utf-8");
+  const fixture = JSON.parse(raw) as BenchmarkFixture;
+
+  test("has version, collection, and non-empty queries", () => {
+    expect(typeof fixture.version).toBe("number");
+    expect(fixture.version).toBeGreaterThanOrEqual(1);
+    expect(typeof fixture.collection).toBe("string");
+    expect(fixture.collection!.length).toBeGreaterThan(0);
+    expect(Array.isArray(fixture.queries)).toBe(true);
+    expect(fixture.queries.length).toBeGreaterThan(0);
+  });
+
+  test("has at least 40 queries", () => {
+    expect(fixture.queries.length).toBeGreaterThanOrEqual(40);
+  });
+
+  test("each query has required fields and allowed type", () => {
+    for (const q of fixture.queries) {
+      expect(typeof q.id).toBe("string");
+      expect(q.id.length).toBeGreaterThan(0);
+      expect(typeof q.query).toBe("string");
+      expect(q.query.length).toBeGreaterThan(0);
+      expect(ALLOWED_TYPES.has(q.type as BenchmarkQuery["type"])).toBe(true);
+      expect(typeof q.description).toBe("string");
+      expect(q.description.length).toBeGreaterThan(0);
+      expect(Array.isArray(q.expected_files)).toBe(true);
+      expect(q.expected_files.length).toBeGreaterThan(0);
+      expect(typeof q.expected_in_top_k).toBe("number");
+      expect(q.expected_in_top_k).toBeGreaterThanOrEqual(1);
+    }
+  });
+
+  test("query ids are unique", () => {
+    const ids = fixture.queries.map((q) => q.id);
+    expect(new Set(ids).size).toBe(ids.length);
+  });
+
+  test("expected_files are wiki-relative .md paths (no abs, no ..)", () => {
+    for (const q of fixture.queries) {
+      for (const file of q.expected_files) {
+        expect(typeof file).toBe("string");
+        expect(file.endsWith(".md")).toBe(true);
+        expect(file.startsWith("/")).toBe(false);
+        expect(file.includes("..")).toBe(false);
+        expect(file).toMatch(WIKI_REL_PATH);
+      }
+    }
+  });
+
+  test("does not vendor or require wiki corpus under test/", () => {
+    // Schema-only suite: fixture JSON is the only input.
+    expect(FIXTURE_PATH.endsWith("wiki-v0.json")).toBe(true);
+    expect(fixture.queries.length).toBeGreaterThan(0);
+  });
+});