diff --git a/brain/eval/baseline-pre-fix.txt b/brain/eval/baseline-pre-fix.txt
new file mode 100644
index 0000000..a9a96bb
--- /dev/null
+++ b/brain/eval/baseline-pre-fix.txt
@@ -0,0 +1,167 @@
+# baseline-pre-fix — 20 questions, k=5
+
+top-1 hit rate: 4/20 = 20%
+top-3 hit rate: 13/20 = 65%
+
+## per-question detail
+
+· rank=3  expected=dex-in-memory-storage-wipes-oauth-tokens-on-every-pod-restart
+     q: how do I stop dex from logging users out on every pod restart?
+     1. homelab-network-perimeter-model
+     2. 2026-05-12-koala-machine-state
+     3. dex-in-memory-storage-wipes-oauth-tokens-on-every-pod-restart  <-- expected
+     4. infra-litellm-absorption-2026-05-16
+     5. Financial Sentiment Analysis on Stock Market Headlines With FinBERT & HuggingFace
+
+★ rank=1  expected=postgres-least-privilege-migration-tenant-grant-bypass-2026-05
+     q: my postgres-exporter broke after revoking PUBLIC CONNECT — why?
+     1. postgres-least-privilege-migration-tenant-grant-bypass-2026-05  <-- expected
+     2. infra-litellm-absorption-2026-05-16
+     3. brain-mcp-activation-runbook
+     4. extension-version-lags-platform-major-upgrade
+     5. ntfy-deny-all-rollout-ordering-keep-alert-pipeline-live-during-auth-flip
+
+★ rank=1  expected=homelab-network-perimeter-model
+     q: when is a NodePort acceptable vs needing a public ingress with bearer gate?
+     1. homelab-network-perimeter-model  <-- expected
+     2. qwen3-thinking-model-empty-content-trap
+     3. mcpclient-empty-token-silent-401-envfrom-missing-key
+     4. 2026-05-12-koala-machine-state
+     5. koala-llama-swap-native-tool-calls-survey-2026-05
+
+· rank=3  expected=exit-255-unknown-reason-not-oom
+     q: what does container exit code 255 with reason Unknown mean?
+     1. qwen3-thinking-model-empty-content-trap
+     2. infra-litellm-absorption-2026-05-16
+     3. exit-255-unknown-reason-not-oom  <-- expected
+     4. mcpclient-empty-token-silent-401-envfrom-missing-key
+     5. koala-llama-swap-native-tool-calls-survey-2026-05
+
+· rank=3  expected=gitea-push-mirror-cannot-create-remote-repo-needs-pre-existing-github-repo
+     q: can gitea push-mirror create the github repo automatically?
+     1. infra-litellm-absorption-2026-05-16
+     2. Autoresearch
+     3. gitea-push-mirror-cannot-create-remote-repo-needs-pre-existing-github-repo  <-- expected
+     4. adr-new-project-gitea-first-github-mirror
+     5. adr-github-as-primary-remote
+
+✗ rank=0  expected=flux-healthcheck-stale-on-resource-removal
+     q: a flux kustomization is stuck after I removed a resource — why?
+     1. qwen3-thinking-model-empty-content-trap
+     2. 2026-05-12-koala-machine-state
+     3. homelab-architecture-principles-2026-05
+     4. gitea-mcp: full stack shipped end-to-end (2026-05-05)
+     5. k8s-configmap-mount-no-reload-needs-pod-restart
+
+· rank=2  expected=go-bytes-buffer-bytes-reset-aliasing-trap
+     q: the bytes buffer aliasing trap with Reset in a loop — what's the bug?
+     1. Financial Sentiment Analysis on Stock Market Headlines With FinBERT & HuggingFace
+     2. go-bytes-buffer-bytes-reset-aliasing-trap  <-- expected
+     3. homelab-security-chains-not-bugs
+     4. training-on-rtx-5070-pretraining-vs-finetuning
+     5. Hash Encoding
+
+★ rank=1  expected=homelab-architecture-principles-2026-05
+     q: what are the homelab architecture principles from may 2026?
+     1. homelab-architecture-principles-2026-05  <-- expected
+     2. homelab-network-perimeter-model
+     3. Claude Managed Agents — architecture notes relevant to homelab agent platform
+     4. homelab-core-glossary
+     5. 2026-05-12-koala-machine-state
+
+✗ rank=0  expected=2026-05-04-sops-age-key-from-flux-cluster
+     q: where does the sops age private key live in the cluster?
+     1. 2026-05-12-koala-machine-state
+     2. homelab-network-perimeter-model
+     3. postgres-least-privilege-migration-tenant-grant-bypass-2026-05
+     4. brain-mcp-activation-runbook
+     5. dex-in-memory-storage-wipes-oauth-tokens-on-every-pod-restart
+
+✗ rank=0  expected=grafana-dashboards-as-code-not-ui-state
+     q: why do my grafana dashboards disappear after a pod restart?
+     1. infra-litellm-absorption-2026-05-16
+     2. 2026-05-12-koala-machine-state
+     3. Financial Sentiment Analysis on Stock Market Headlines With FinBERT & HuggingFace
+     4. brain-mcp-activation-runbook
+     5. dex-in-memory-storage-wipes-oauth-tokens-on-every-pod-restart
+
+· rank=2  expected=double-diamond-methodology
+     q: what is the double diamond methodology?
+     1. Harnessing the Power of Hash Encoding for Categorical Data in Data Science
+     2. double-diamond-methodology  <-- expected
+     3. unified-methodology-diamond-futures-autoresearch
+     4. futures-thinking-extended-double-diamond
+     5. insight-exploration-as-diamond-1
+
+· rank=3  expected=2026-05-04-mcp-transport-version-claude-ai-strict
+     q: my MCP server works from claude code but fails on claude.ai — what's different?
+     1. qwen3-thinking-model-empty-content-trap
+     2. mcp-resource-url-empty-breaks-claude-ai-discovery-silently
+     3. 2026-05-04-mcp-transport-version-claude-ai-strict  <-- expected
+     4. 2026-05-04-claude-ai-custom-mcp-connectors
+     5. finding-github-mcp-claudeai-vs-claudecode
+
+· rank=2  expected=homelab-security-chains-not-bugs
+     q: how should I rate security findings — isolated bugs or exploit chains?
+     1. homelab-network-perimeter-model
+     2. homelab-security-chains-not-bugs  <-- expected
+     3. Financial Sentiment Analysis on Stock Market Headlines With FinBERT & HuggingFace
+     4. policy-audit-mode-blocks-nothing
+     5. homelab-document-accepted-risk-to-break-audit-cycle
+
+· rank=2  expected=2026-05-03-canonical-vs-derived-context-flow
+     q: how should canonical context files relate to derived adapter files?
+     1. qwen3-thinking-model-empty-content-trap
+     2. 2026-05-03-canonical-vs-derived-context-flow  <-- expected
+     3. 2026-05-12-koala-machine-state
+     4. 2026-05-04-claude-ai-custom-mcp-connectors
+     5. koala-llama-swap-native-tool-calls-survey-2026-05
+
+· rank=2  expected=homelab-core-glossary
+     q: what is the homelab core vocabulary glossary?
+     1. homelab-architecture-principles-2026-05
+     2. homelab-core-glossary  <-- expected
+     3. Claude Managed Agents — architecture notes relevant to homelab agent platform
+     4. 2026-05-12-koala-machine-state
+     5. Autoresearch
+
+★ rank=1  expected=koala-llama-swap-native-tool-calls-survey-2026-05
+     q: which models on koala llama-swap actually emit native tool_calls correctly?
+     1. koala-llama-swap-native-tool-calls-survey-2026-05  <-- expected
+     2. 2026-05-12-koala-machine-state
+     3. infra-litellm-absorption-2026-05-16
+     4. training-on-rtx-5070-pretraining-vs-finetuning
+     5. qwen3-thinking-model-empty-content-trap
+
+✗ rank=0  expected=qwen35-9b-fast
+     q: what is qwen35-9b-fast and what's it used for?
+     1. koala-llama-swap-native-tool-calls-survey-2026-05
+     2. qwen3-thinking-model-empty-content-trap
+     3. Qwen35-9b-fast
+     4. infra-litellm-absorption-2026-05-16
+     5. 2026-05-12-koala-machine-state
+
+✗ rank=0  expected=go-defer-errcheck-body-close
+     q: in go, how do I prevent defer body close from silently dropping errors?
+     1. infra-litellm-absorption-2026-05-16
+     2. homelab-network-perimeter-model
+     3. go-bytes-buffer-bytes-reset-aliasing-trap
+     4. mcpclient-empty-token-silent-401-envfrom-missing-key
+     5. brain-mcp-activation-runbook
+
+✗ rank=0  expected=hyperguild-level3-pipeline-rewrite
+     q: what was the level 3 rewrite of hyperguild's ingestion pipeline?
+     1. 2026-05-12-koala-machine-state
+     2. homelab-core-glossary
+     3. brain-mcp-activation-runbook
+     4. koala-llama-swap-native-tool-calls-survey-2026-05
+     5. infra-litellm-absorption-2026-05-16
+
+? rank=4  expected=adr-new-project-gitea-first-github-mirror
+     q: what's the new-project ADR — is it gitea-first or github-first?
+     1. gitea-push-mirror-cannot-create-remote-repo-needs-pre-existing-github-repo
+     2. gitea-mcp: full stack shipped end-to-end (2026-05-05)
+     3. mcp-tool-design-get-needs-list-partner
+     4. adr-new-project-gitea-first-github-mirror  <-- expected
+     5. 2026-05-04-gitea-mcp-build-session
+
diff --git a/brain/eval/qa-2026-05.md b/brain/eval/qa-2026-05.md
new file mode 100644
index 0000000..f0c99d5
--- /dev/null
+++ b/brain/eval/qa-2026-05.md
@@ -0,0 +1,76 @@
+# Brain retrieval eval set — 2026-05-24
+
+20 hand-authored Q→expected-top-1-slug pairs. Used by `score.sh` to
+measure brain_query top-1 + top-3 hit rate against the live brain.
+
+Authoring rules:
+- Each question maps to **one** clear-best entry. Avoid ambiguous
+  questions where multiple slugs could be the right answer.
+- Questions are phrased the way a future-me would actually ask, not
+  the way the entry's title reads. Some lexical distance is the point.
+- `expected` is the slug as stored in `brain_entities.slug`. Update
+  if the slug renames.
+
+## Pairs
+
+```
+q: how do I stop dex from logging users out on every pod restart?
+expected: dex-in-memory-storage-wipes-oauth-tokens-on-every-pod-restart
+
+q: my postgres-exporter broke after revoking PUBLIC CONNECT — why?
+expected: postgres-least-privilege-migration-tenant-grant-bypass-2026-05
+
+q: when is a NodePort acceptable vs needing a public ingress with bearer gate?
+expected: homelab-network-perimeter-model
+
+q: what does container exit code 255 with reason Unknown mean?
+expected: exit-255-unknown-reason-not-oom
+
+q: can gitea push-mirror create the github repo automatically?
+expected: gitea-push-mirror-cannot-create-remote-repo-needs-pre-existing-github-repo
+
+q: a flux kustomization is stuck after I removed a resource — why?
+expected: flux-healthcheck-stale-on-resource-removal
+
+q: the bytes buffer aliasing trap with Reset in a loop — what's the bug?
+expected: go-bytes-buffer-bytes-reset-aliasing-trap
+
+q: what are the homelab architecture principles from may 2026?
+expected: homelab-architecture-principles-2026-05
+
+q: where does the sops age private key live in the cluster?
+expected: 2026-05-04-sops-age-key-from-flux-cluster
+
+q: why do my grafana dashboards disappear after a pod restart?
+expected: grafana-dashboards-as-code-not-ui-state
+
+q: what is the double diamond methodology?
+expected: double-diamond-methodology
+
+q: my MCP server works from claude code but fails on claude.ai — what's different?
+expected: 2026-05-04-mcp-transport-version-claude-ai-strict
+
+q: how should I rate security findings — isolated bugs or exploit chains?
+expected: homelab-security-chains-not-bugs
+
+q: how should canonical context files relate to derived adapter files?
+expected: 2026-05-03-canonical-vs-derived-context-flow
+
+q: what is the homelab core vocabulary glossary?
+expected: homelab-core-glossary
+
+q: which models on koala llama-swap actually emit native tool_calls correctly?
+expected: koala-llama-swap-native-tool-calls-survey-2026-05
+
+q: what is qwen35-9b-fast and what's it used for?
+expected: qwen35-9b-fast
+
+q: in go, how do I prevent defer body close from silently dropping errors?
+expected: go-defer-errcheck-body-close
+
+q: what was the level 3 rewrite of hyperguild's ingestion pipeline?
+expected: hyperguild-level3-pipeline-rewrite
+
+q: what's the new-project ADR — is it gitea-first or github-first?
+expected: adr-new-project-gitea-first-github-mirror
+```
diff --git a/brain/eval/score.py b/brain/eval/score.py
new file mode 100644
index 0000000..25f73c2
--- /dev/null
+++ b/brain/eval/score.py
@@ -0,0 +1,131 @@
+#!/usr/bin/env python3
+"""Score brain_query against the qa-2026-05.md eval set.
+
+Reads `q:` / `expected:` pairs, calls brain_query MCP for each, records
+top-1 + top-3 hit rate. Run:
+
+    BRAIN_MCP_TOKEN=$(grep '^export BRAIN_MCP_TOKEN=' ~/.llmkeys | cut -d= -f2-) \\
+      python3 score.py qa-2026-05.md
+
+Optionally pass --baseline <name> to save the result as a labeled run.
+"""
+import argparse
+import json
+import os
+import re
+import sys
+import time
+import urllib.request
+
+ENDPOINT = "https://brain-mcp.d-ma.be/mcp"
+
+
+def load_pairs(path):
+    pairs = []
+    q = None
+    with open(path) as f:
+        for line in f:
+            line = line.rstrip()
+            if line.startswith("q:"):
+                q = line[2:].strip()
+            elif line.startswith("expected:") and q is not None:
+                expected = line[len("expected:"):].strip()
+                pairs.append((q, expected))
+                q = None
+    return pairs
+
+
+def brain_query(token, query, k=5):
+    body = json.dumps({
+        "jsonrpc": "2.0",
+        "id": 1,
+        "method": "tools/call",
+        "params": {"name": "brain_query", "arguments": {"query": query, "k": k}},
+    }).encode()
+    req = urllib.request.Request(
+        ENDPOINT,
+        data=body,
+        headers={
+            "Authorization": f"Bearer {token}",
+            "Content-Type": "application/json",
+            "Accept": "application/json, text/event-stream",
+        },
+        method="POST",
+    )
+    with urllib.request.urlopen(req, timeout=30) as r:
+        raw = r.read().decode()
+    for line in raw.splitlines():
+        if line.startswith("data:"):
+            raw = line[5:].strip()
+            break
+    d = json.loads(raw)
+    if "error" in d:
+        raise RuntimeError(d["error"])
+    text = d["result"]["content"][0]["text"]
+    return json.loads(text).get("results", [])
+
+
+def slug_of(result):
+    # `title` mirrors the slug in brain_entities for normal entries.
+    # Fall back to basename(path) if title is missing.
+    t = result.get("title", "")
+    if t:
+        return t
+    p = result.get("path", "")
+    return re.sub(r"\.md$", "", os.path.basename(p))
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("evalset")
+    ap.add_argument("--baseline", default="run")
+    ap.add_argument("--k", type=int, default=5)
+    args = ap.parse_args()
+
+    token = os.environ.get("BRAIN_MCP_TOKEN")
+    if not token:
+        sys.exit("BRAIN_MCP_TOKEN not set")
+
+    pairs = load_pairs(args.evalset)
+    if not pairs:
+        sys.exit(f"no pairs in {args.evalset}")
+
+    print(f"# {args.baseline} — {len(pairs)} questions, k={args.k}")
+    print()
+    hits1 = 0
+    hits3 = 0
+    detail = []
+    for q, expected in pairs:
+        try:
+            results = brain_query(token, q, k=args.k)
+        except Exception as e:
+            detail.append((q, expected, [], f"ERR {e}"))
+            continue
+        slugs = [slug_of(r) for r in results]
+        rank = slugs.index(expected) + 1 if expected in slugs else 0
+        h1 = 1 if rank == 1 else 0
+        h3 = 1 if 0 < rank <= 3 else 0
+        hits1 += h1
+        hits3 += h3
+        detail.append((q, expected, slugs, rank))
+
+    total = len(pairs)
+    print(f"top-1 hit rate: {hits1}/{total} = {100*hits1/total:.0f}%")
+    print(f"top-3 hit rate: {hits3}/{total} = {100*hits3/total:.0f}%")
+    print()
+    print("## per-question detail")
+    print()
+    for q, expected, slugs, rank in detail:
+        marker = {0: "✗", 1: "★", 2: "·", 3: "·"}.get(rank, "?")
+        if isinstance(rank, str):
+            marker = "!"
+        print(f"{marker} rank={rank}  expected={expected}")
+        print(f"     q: {q}")
+        for i, s in enumerate(slugs[:args.k], 1):
+            mark = "  <-- expected" if s == expected else ""
+            print(f"     {i}. {s}{mark}")
+        print()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/ingestion/internal/graph/extract.go b/ingestion/internal/graph/extract.go
index 57550cd..0281665 100644
--- a/ingestion/internal/graph/extract.go
+++ b/ingestion/internal/graph/extract.go
@@ -83,12 +83,23 @@ func slugFromPath(docPath string) string {
 // classifyByPath fills Type / Wing / Hall from the path layout when the
 // doc lives under brain/wiki/. Layout: wiki/<wing>/<hall>/<slug>.md
 // or wiki/<bucket>/<slug>.md for the legacy concept/entity/source dirs.
+//
+// Files directly under wiki/ (no subdirectory — e.g. wiki/index.md) used
+// to incorrectly land Type="hall" Wing="index.md" because the path's
+// second segment was the file itself. Now they fall through to Type
+// "knowledge" and leave wing/hall to frontmatter.
 func classifyByPath(e *Entity, docPath string) {
 	parts := strings.Split(docPath, "/")
 	if len(parts) < 2 || parts[0] != "wiki" {
 		e.Type = "knowledge"
 		return
 	}
+	if len(parts) < 3 {
+		// wiki/<slug>.md — no subdirectory. Treat as plain knowledge
+		// and let frontmatter set wing/hall if they're present.
+		e.Type = "knowledge"
+		return
+	}
 	switch parts[1] {
 	case "concepts":
 		e.Type = "concept"
diff --git a/ingestion/internal/graph/extract_test.go b/ingestion/internal/graph/extract_test.go
index c785837..53faa0a 100644
--- a/ingestion/internal/graph/extract_test.go
+++ b/ingestion/internal/graph/extract_test.go
@@ -104,3 +104,31 @@ func TestExtract_LineNumbersAre1Indexed(t *testing.T) {
 	require.Len(t, edges, 1)
 	assert.Equal(t, 2, edges[0].SrcLine)
 }
+
+// Files directly under wiki/ (no subdirectory) used to land
+// Type="hall" Wing="<filename>.md" because the path's second segment
+// was the file itself. The fix routes them to Type="knowledge" with
+// empty Wing/Hall and lets frontmatter set them if present.
+func TestExtract_WikiRootFileIsKnowledgeNotHall(t *testing.T) {
+	content := []byte("# Index\n\n- [[foo]]\n")
+	ent, _, ok := Extract("wiki/index.md", content)
+	require.True(t, ok)
+	assert.Equal(t, "index", ent.Slug)
+	assert.Equal(t, "knowledge", ent.Type)
+	assert.Empty(t, ent.Wing)
+	assert.Empty(t, ent.Hall)
+}
+
+func TestExtract_WikiRootFileWithFrontmatterWingHall(t *testing.T) {
+	content := []byte(`---
+wing: homelab
+hall: facts
+---
+# Some root note
+`)
+	ent, _, ok := Extract("wiki/some-note.md", content)
+	require.True(t, ok)
+	assert.Equal(t, "knowledge", ent.Type)
+	assert.Equal(t, "homelab", ent.Wing)
+	assert.Equal(t, "facts", ent.Hall)
+}