diff --git a/website/concepts/knowledge-graph.md b/website/concepts/knowledge-graph.md
index d7969b3..97a318f 100644
--- a/website/concepts/knowledge-graph.md
+++ b/website/concepts/knowledge-graph.md
@@ -80,12 +80,11 @@ The knowledge graph uses SQLite with two tables:
 
 Database location: `~/.mempalace/knowledge_graph.sqlite3`
 
-## Comparison
+## Related Work
 
-| Feature | MemPalace | Zep (Graphiti) |
-|---------|-----------|----------------|
-| Storage | SQLite (local) | Neo4j (cloud) |
-| Cost | Free | $25/mo+ |
-| Temporal validity | Yes | Yes |
-| Self-hosted | Always | Enterprise only |
-| Privacy | Everything local | SOC 2, HIPAA |
+Temporal entity-relationship graphs are a familiar pattern — Zep's
+Graphiti, for example, also exposes a bi-temporal model. MemPalace's
+knowledge graph is local-first (SQLite, everything on disk) and free;
+Zep is a managed service backed by Neo4j with its own pricing, SLAs,
+and compliance surface. See Zep's own [documentation](https://www.getzep.com/)
+for authoritative details on their deployment model.
diff --git a/website/concepts/the-palace.md b/website/concepts/the-palace.md
index 2ef01a4..c2fd114 100644
--- a/website/concepts/the-palace.md
+++ b/website/concepts/the-palace.md
@@ -92,16 +92,9 @@ The original stored text chunks. This is the primary retrieval layer used by the
 
 ## Why Structure Matters
 
-Tested on 22,000+ real conversation memories:
+Wing and room identifiers become metadata filters at query time. Narrowing a search to a specific wing (or wing + room) means the vector store only scores candidates inside that scope, which is useful when you have many unrelated projects or people filed in the same palace.
 
-| Search scope | R@10 | Improvement |
-|-------------|------|-------------|
-| All closets | 60.9% | baseline |
-| Within wing | 73.1% | +12% |
-| Wing + hall | 84.8% | +24% |
-| Wing + room | 94.8% | +34% |
-
-The practical point is that structure improves retrieval. In the project benchmarks, narrowing the search scope by wing and room outperformed searching the entire corpus at once.
+This is standard metadata filtering in the underlying vector store, not a novel retrieval mechanism. The useful property here is operational — clear scoping rules that a human or an agent can apply predictably — not a magic retrieval boost.
 
 ## Navigation
 
diff --git a/website/guide/searching.md b/website/guide/searching.md
index e9c0bbb..ce2ce3b 100644
--- a/website/guide/searching.md
+++ b/website/guide/searching.md
@@ -23,23 +23,16 @@ mempalace search "deploy process" --results 10
 
 ## How Search Works
 
-1. Your query is embedded using ChromaDB's default model (`all-MiniLM-L6-v2`)
-2. The embedding is compared against all drawers using cosine similarity
-3. Optional wing/room filters narrow the search scope
-4. Results are returned with similarity scores and source metadata
+1. Your query is embedded using the vector store's default model (`all-MiniLM-L6-v2` with the default ChromaDB backend).
+2. The embedding is compared against all drawers using cosine similarity.
+3. Optional wing/room filters narrow the search scope — standard metadata filtering in the underlying vector store.
+4. Results are returned with similarity scores and source metadata.
 
-### Why Structure Matters
+### Why Scoping Matters
 
-Tested on 22,000+ real conversation memories:
+Wing/room filtering is useful when a single palace contains many unrelated projects or people. Narrowing the search to a specific wing (or wing + room) means the vector store only scores candidates inside that scope, which keeps retrieval predictable as the palace grows.
 
-```
-Search all closets:          60.9%  R@10
-Search within wing:          73.1%  (+12%)
-Search wing + hall:          84.8%  (+24%)
-Search wing + room:          94.8%  (+34%)
-```
-
-Wings and rooms aren't cosmetic — they're a **34% retrieval improvement**.
+This is a metadata-filter feature of the vector store, not a novel retrieval mechanism. Treat it as an operational convenience: clear scoping rules that a human or an agent can apply predictably.
 
 ## Programmatic Search
 
diff --git a/website/index.md b/website/index.md
index a8487fb..b32cc35 100644
--- a/website/index.md
+++ b/website/index.md
@@ -4,7 +4,7 @@ layout: home
 hero:
   name: MemPalace
   text: Give your AI a memory.
-  tagline: "96.6% recall on LongMemEval in raw mode. Local-first, open source, and usable without an API key."
+  tagline: "Local-first AI memory. Verbatim storage, pluggable backend, 96.6% R@5 raw on LongMemEval — zero API calls."
   image:
     src: /mempalace_logo.png
     alt: MemPalace
@@ -34,7 +34,7 @@ features:
       src: /icons/search.svg
       alt: Semantic Search
     title: Semantic Search
-    details: ChromaDB-powered vector search lets the model retrieve past discussions by topic, project, or room.
+    details: Vector search over verbatim content lets the model retrieve past discussions by topic, project, or room. Backend is pluggable.
   - icon:
       src: /icons/git-merge.svg
       alt: Knowledge Graph
@@ -49,7 +49,7 @@ features:
       src: /icons/shield-check.svg
       alt: Zero Cloud
     title: Zero Cloud
-    details: Core storage and retrieval run locally on ChromaDB and SQLite. Optional reranking features can add an API dependency.
+    details: Core storage and retrieval run locally. Optional reranking features can add an API dependency but are not required for the benchmark path.
 ---
 
 <style>
@@ -68,20 +68,21 @@ features:
 
 ## Verbatim Retrieval First
 
-MemPalace starts from a simple premise: **store the source text and retrieve it well**. The benchmarked raw mode does not require an LLM extraction step.
+MemPalace stores source text and retrieves it with semantic search. The benchmarked raw mode does not require an LLM at any stage — no extraction, no rerank, no summarisation.
 
-| System | LongMemEval R@5 | API Required | Cost |
-|--------|----------------|--------------|------|
-| **MemPalace (hybrid)** | **100%** | Optional | Free |
-| Supermemory ASMR | ~99% | Yes | — |
-| **MemPalace (raw)** | **96.6%** | **None** | **Free** |
-| Mastra | 94.87% | Yes | API costs |
-| Mem0 | ~85% | Yes | $19–249/mo |
+**LongMemEval retrieval recall (500 questions):**
 
-The raw 96.6% LongMemEval result is the baseline story: strong recall without requiring an API key or an LLM in the retrieval pipeline.
+| Mode | R@5 | LLM required |
+|---|---|---|
+| Raw (semantic search over verbatim text) | **96.6%** | None |
+| Hybrid v4, held-out 450q | **98.4%** | None |
+
+The raw 96.6% reproduces on any machine with the committed dataset: result JSONLs, the `seed=42` train/held-out split, and the `--mode raw` / `--held-out` runners are all in the `benchmarks/` directory of the repo.
+
+We deliberately do not publish a side-by-side comparison against other memory systems on this page. Retrieval recall (R@5) and end-to-end QA accuracy are different metrics and are not comparable; where MemPalace can be fairly compared on the same metric, we link to the other project's published source.
 
 <div style="text-align: center; padding-top: 16px;">
-  <a href="./reference/benchmarks" style="color: var(--vp-c-brand-1); font-weight: 500;">Full benchmark results →</a>
+  <a href="./reference/benchmarks" style="color: var(--vp-c-brand-1); font-weight: 500;">Full benchmark methodology →</a>
 </div>
 
 </div>
diff --git a/website/reference/benchmarks.md b/website/reference/benchmarks.md
index dae2df1..60bc8cb 100644
--- a/website/reference/benchmarks.md
+++ b/website/reference/benchmarks.md
@@ -1,28 +1,51 @@
 # Benchmarks
 
-Curated summary of MemPalace benchmark results. For the full 725-line progression with every experiment, see [`benchmarks/BENCHMARKS.md`](https://github.com/MemPalace/mempalace/blob/main/benchmarks/BENCHMARKS.md) in the repository.
+Curated summary of MemPalace's reproducible benchmark results. For the
+complete progression with every experiment, see
+[`benchmarks/BENCHMARKS.md`](https://github.com/MemPalace/mempalace/blob/main/benchmarks/BENCHMARKS.md).
+All headline numbers on this page are reproducible from the committed
+repository — datasets, scripts, and per-question result JSONLs are all
+checked in.
 
 ## The Core Finding
 
-MemPalace's benchmarked raw baseline stores the source text and searches it with ChromaDB's default embeddings. No extraction layer or summarization step is required for that baseline.
+MemPalace's benchmarked raw baseline stores the source text and searches
+it with the vector store's default embeddings. No extraction or
+summarisation step is required for that baseline, and it reproduces at
+**96.6% R@5** on LongMemEval with no LLM at any stage.
 
-**And it scores 96.6% on LongMemEval.**
+## LongMemEval — Retrieval Recall
 
-## LongMemEval Results
+Retrieval recall asks: is the labelled session for this question inside
+the top-K retrieved sessions? It is not the same metric as end-to-end QA
+accuracy; a system can have perfect retrieval recall and poor QA answer
+quality, and vice versa.
 
-| Mode | R@5 | LLM Required | Cost/query |
-|------|-----|-------------|------------|
-| Raw ChromaDB | **96.6%** | None | $0 |
-| Hybrid v3 + rerank | 99.4% | Haiku | ~$0.001 |
-| Palace + rerank | 99.4% | Haiku | ~$0.001 |
-| **Hybrid v4 + rerank** | **100%** | Haiku | ~$0.001 |
+**Full 500 questions:**
 
-The 96.6% raw score requires no API key, no cloud, and no LLM at any stage. The 100% result uses optional Haiku reranking.
+| Mode | R@5 | LLM required | Cost/query |
+|---|---|---|---|
+| Raw — vector search over verbatim sessions | **96.6%** | None | $0 |
+| Hybrid v4 — keyword/temporal/preference boosts, no LLM | 98.6% | None | $0 |
+| Hybrid v4 + LLM rerank (minimax-m2.7 via Ollama) | 99.2% | Any capable model | $0 local / varies cloud |
 
-### Per-Category Breakdown (Raw, 96.6%)
+**Held-out set (450 questions, never used during `hybrid_v4` development):**
 
-| Question Type | R@5 | Count |
-|---------------|-----|-------|
+| Mode | R@5 | R@10 | NDCG@10 |
+|---|---|---|---|
+| Hybrid v4 | **98.4%** | 99.8% | 0.938 |
+
+The held-out figure is the honest generalisable number. The full-500
+scores are higher but include the 50 "dev" questions that hybrid_v4's
+three targeted fixes (quoted-phrase boost, person-name boost, nostalgia
+patterns) were developed against. `benchmarks/BENCHMARKS.md` calls this
+"teaching to the test" and the held-out 98.4% is the clean number to
+quote when a single R@5 figure is needed for the hybrid pipeline.
+
+### Per-category breakdown (raw, 96.6%)
+
+| Question type | R@5 | Count |
+|---|---|---|
 | Knowledge update | 99.0% | 78 |
 | Multi-session | 98.5% | 133 |
 | Temporal reasoning | 96.2% | 133 |
@@ -30,66 +53,95 @@ The 96.6% raw score requires no API key, no cloud, and no LLM at any stage. The
 | Single-session preference | 93.3% | 30 |
 | Single-session assistant | 92.9% | 56 |
 
-### Held-Out Validation
+## LoCoMo — Retrieval Recall
 
-**98.4% R@5** on 450 questions that hybrid_v4 was never tuned on — confirming the improvements generalize.
+LoCoMo contains 1,986 questions across 10 long conversations (19–32
+sessions each).
 
-## Comparison vs Published Systems
+| Mode | R@10 | LLM required |
+|---|---|---|
+| Session, no rerank, top-10 | 60.3% | None |
+| Hybrid v5 (keyword + predicate boosts), top-10 | 88.9% | None |
 
-| System | LongMemEval R@5 | API Required | Cost |
-|--------|----------------|--------------|------|
-| **MemPalace (hybrid)** | **100%** | Optional | Free |
-| Supermemory ASMR | ~99% | Yes | — |
-| **MemPalace (raw)** | **96.6%** | **None** | **Free** |
-| Mastra | 94.87% | Yes | API costs |
-| Hindsight | 91.4% | Yes | API costs |
-| Mem0 | ~85% | Yes | $19–249/mo |
+We do not publish a "100% R@10" headline for LoCoMo. A reported 100% in
+earlier drafts used `top_k=50`, which exceeds the per-conversation
+session count (19–32) — so the retrieval stage returns every session in
+every conversation by construction. That number measures an LLM's
+reading comprehension over the whole conversation, not retrieval. The
+honest retrieval-recall number for LoCoMo is the top-10 figure.
 
 ## Other Benchmarks
 
-### ConvoMem (Salesforce, 75K+ QA pairs)
+**ConvoMem** (Salesforce; 50 items per category × 5 categories = 250
+items): MemPalace raw retrieval reaches **92.9% avg recall**. Strongest
+categories: Assistant Facts 100%, User Facts 98%. Weakest: Preferences
+86%. The Salesforce dataset contains ~75K items in total; our headline
+number is from the 250-item sample the benchmark script was designed
+around.
 
-| System | Score |
-|--------|-------|
-| **MemPalace** | **92.9%** |
-| Gemini (long context) | 70–82% |
-| Block extraction | 57–71% |
-| Mem0 (RAG) | 30–45% |
+**MemBench** (ACL 2025; 8,500 items, all topics): MemPalace hybrid
+top-5 reaches **80.3% R@5 overall**. Strongest: aggregative 99.3%,
+comparative 98.4%, lowlevel_rec 99.8%. Weakest: noisy 43.4%
+(distractor-heavy by design), conditional 57.3%.
 
-On this benchmark, MemPalace materially outperforms the Mem0 result cited in the comparison table.
+## Why We Don't Publish a Cross-System Comparison Table
 
-### LoCoMo (1,986 multi-hop QA pairs)
+Previous versions of this page placed MemPalace's retrieval recall (R@5)
+next to other projects' end-to-end QA accuracy figures under a single
+"LongMemEval R@5" column. Those are different metrics and are not
+comparable. A system can have 100% retrieval recall and 40% QA
+accuracy, and vice versa.
 
-| Mode | R@10 | LLM |
-|------|------|-----|
-| Hybrid v5 + Sonnet rerank (top-50) | **100%** | Sonnet |
-| bge-large + Haiku rerank (top-15) | 96.3% | Haiku |
-| Hybrid v5 (top-10, no rerank) | **88.9%** | None |
-| Session, no rerank (top-10) | 60.3% | None |
+If you are evaluating memory systems against MemPalace and want a fair
+comparison, use the retrieval-recall numbers above and the benchmark
+scripts in the repo; or pick the metric the other project publishes and
+compare on that. Each project's published source is the correct
+reference:
 
-### MemBench (ACL 2025, 8,500 items)
-
-**80.3% R@5** overall. Strongest categories: aggregative (99.3%), comparative (98.4%), lowlevel_rec (99.8%).
+- [Mastra — Observational Memory](https://mastra.ai/research/observational-memory)
+  (their published metric is binary QA accuracy with GPT-5-mini)
+- [Mem0 — Research](https://mem0.ai/research)
+  (their published LoCoMo metric is end-to-end QA accuracy, not retrieval recall)
+- [Supermemory — ASMR post](https://supermemory.ai/blog/we-broke-the-frontier-in-agent-memory-introducing-99-sota-memory-system/)
+  (their published metric is QA accuracy; authors explicitly frame the
+  ensemble as an experimental proof-of-concept, not production)
 
 ## Reproducing Results
 
-All benchmarks are reproducible with public datasets:
+Every benchmark runs deterministically from this repository.
 
 ```bash
 git clone https://github.com/MemPalace/mempalace.git
 cd mempalace
-pip install chromadb pyyaml
+pip install -e ".[dev]"
 
-# Download LongMemEval data
+# LongMemEval — raw (96.6%)
 curl -fsSL -o /tmp/longmemeval_s_cleaned.json \
   https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json
-
-# Run raw baseline (96.6%, no API key needed)
 python benchmarks/longmemeval_bench.py /tmp/longmemeval_s_cleaned.json
+
+# LongMemEval — hybrid v4 on the held-out 450 (98.4%)
+python benchmarks/longmemeval_bench.py /tmp/longmemeval_s_cleaned.json \
+  --mode hybrid_v4 --held-out --split-file benchmarks/lme_split_50_450.json
+
+# LoCoMo — session, top-10 (60.3%)
+git clone https://github.com/snap-research/locomo.git /tmp/locomo
+python benchmarks/locomo_bench.py /tmp/locomo/data/locomo10.json \
+  --granularity session --top-k 10
+
+# LongMemEval — hybrid v4 + rerank, any OpenAI-compatible endpoint
+python benchmarks/longmemeval_bench.py /tmp/longmemeval_s_cleaned.json \
+  --mode hybrid_v4 --llm-rerank \
+  --llm-backend ollama --llm-model <your-model-tag>
 ```
 
 ::: tip
-Results are deterministic. Same data + same script = same result every time. Every result JSONL file contains every question, every retrieved document, every score.
+Results are deterministic: same data, same script, same split seed →
+same score. The committed `benchmarks/results_*.jsonl` files include
+every question, every retrieved corpus id, and every score, so every
+individual answer is auditable — not just the aggregate.
 :::
 
-For complete reproduction instructions, benchmark integrity notes, and the full score progression, see the [full benchmark documentation](https://github.com/MemPalace/mempalace/blob/main/benchmarks/BENCHMARKS.md).
+For the complete progression (hybrid v1 → v4, diary mode, palace mode,
+LoCoMo architecture iterations, methodology integrity notes), see
+[`benchmarks/BENCHMARKS.md`](https://github.com/MemPalace/mempalace/blob/main/benchmarks/BENCHMARKS.md).
diff --git a/website/reference/contributing.md b/website/reference/contributing.md
index 17dc6d9..2b909b4 100644
--- a/website/reference/contributing.md
+++ b/website/reference/contributing.md
@@ -68,7 +68,7 @@ If you're planning a significant change, open an issue first. Key principles:
 - **Verbatim first** — never summarize user content. Store exact words.
 - **Local first** — everything runs on the user's machine. No cloud dependencies.
 - **Zero API by default** — core features must work without any API key.
-- **Palace structure matters** — wings, halls, and rooms aren't cosmetic — they drive a 34% retrieval improvement.
+- **Palace structure is scoping, not magic** — wings, halls, and rooms act as metadata filters in the underlying vector store. They make scoping predictable when a palace holds many unrelated projects; they are not a novel retrieval mechanism.
 
 ## Community