diff --git a/README.ko.md b/README.ko.md
index f6e5aff..34edd90 100644
--- a/README.ko.md
+++ b/README.ko.md
@@ -1,6 +1,6 @@
 # Synaptic Memory
 
-**인덱싱에 API 호출 0회. 인프라 0. 락인 0.**
+**기본 경로는 인덱싱 API 호출 0회. 인프라 0. 락인 0.**
 LLM 에이전트용 지식 그래프 + MCP 도구 서버. 하이브리드 검색, CDC 기반 실시간 DB 동기화, 한국어 FTS 내장.
 
 [![PyPI](https://img.shields.io/pypi/v/synaptic-memory)](https://pypi.org/project/synaptic-memory/)
@@ -82,7 +82,7 @@ MCP 도구 → LLM 에이전트가 그래프 기반 멀티턴으로 탐색
 ```
 
 **라이브러리가 하는 건 딱 두 가지:**
-1. **그래프를 잘 구축한다** — 인덱싱에 LLM 비용 0원
+1. **그래프를 잘 구축한다** — 기본은 비용 없는 deterministic extraction
 2. **LLM에게 좋은 도구를 쥐어준다** — 판단은 LLM이, 코드는 데이터만
 
 ---
@@ -90,9 +90,12 @@ MCP 도구 → LLM 에이전트가 그래프 기반 멀티턴으로 탐색
 ## 설치
 
 ```bash
-# 추천 — README의 모든 예제가 작동하는 조합
+# 일반 로컬 그래프 + MCP 조합
 pip install "synaptic-memory[sqlite,korean,vector,mcp]"
 
+# LangChain retriever 예제를 실행할 때 추가
+pip install "synaptic-memory[langchain]"
+
 # 또는 Postgres / Kuzu / Qdrant / MinIO 까지 전부
 pip install "synaptic-memory[all]"
 ```
@@ -210,13 +213,14 @@ asyncio.run(main())
 |------|---------------|---------------|
 | GraphRAG 계열 (MS GraphRAG, Cognee, Graphiti) | LLM으로 엔티티 + 관계 + 커뮤니티 요약 추출 | 서사형 코퍼스에서 recall 최상. 대신 문서 추가마다 LLM 토큰 비용 |
 | LightRAG 계열 | LLM 호출을 쿼리 시점으로 지연 | 인덱스 비용 낮음. 대신 쿼리마다 비용 |
-| **Synaptic** | **없음.** 구조·통계 시그널만 (FK, NEXT_CHUNK, phrase DF 허브, MENTIONS) | 비용 0 + 결정론적. 단, 새로운 관계를 스스로 합성하지 않음 |
+| **Synaptic 기본 경로** | **없음.** 구조·통계 시그널만 (FK, NEXT_CHUNK, phrase DF 허브, MENTIONS) | 비용 0 + 결정론적. 명시적 cross-reference를 LLM 없이 엣지화 |
 
-인덱싱에 LLM을 쓰지 않습니다. 그래프는 지식 저장소가 아니라 검색 인덱스입니다.
-LLM이 합성한 요약이 필요하면 그래프 위에 별도 에이전트 레이어로 쌓으세요 —
-Synaptic은 primitive를 제공하고, 합성 여부는 사용자가 선택합니다.
+기본 인덱싱은 LLM-free입니다. 그래프는 지식 저장소가 아니라 검색 인덱스입니다.
+문서에 명시된 cross-reference는 LLM 없이 `REFERENCES` 엣지로 만들 수 있습니다.
+OpenIE를 opt-in하면 bounded/revertible 방식으로 LLM 기반 semantic relation을
+추가할 수 있지만, 기본 deterministic 경로에는 포함되지 않습니다.
 
-> **v0.28+**: `graph.search()`는 하나의 경로만 사용합니다. BM25 + HNSW +
+> **현재 API**: `graph.search()`는 하나의 경로만 사용합니다. BM25 + HNSW +
 > PPR + cross-encoder + MMR 기반 EvidenceSearch 파이프라인입니다.
 > 예전 `engine=` 스위치는 제거되었으므로 예제는
 > `graph.search("질문")`처럼 바로 호출하면 됩니다.
@@ -230,7 +234,7 @@ Synaptic은 primitive를 제공하고, 합성 여부는 사용자가 선택합
 |------|------|
 | `deep_search` | **추천.** 검색 → 확장 → 문서 읽기를 한 번에 |
 | `compare_search` | 복합 질문 자동 분해 + 병렬 검색 |
-| `knowledge_search` | 핵심 의미 검색 (v0.14.2+에서 EvidenceSearch 경유) |
+| `knowledge_search` | EvidenceSearch 기반 핵심 의미 검색 |
 | `agent_search` | FTS + 벡터 하이브리드 + intent routing |
 | `expand` | 1-hop 그래프 이웃 탐색 |
 | `get_document` | 쿼리 관련 청크만 선별한 문서 전문 |
@@ -243,8 +247,9 @@ Synaptic은 primitive를 제공하고, 합성 여부는 사용자가 선택합
 | `filter_nodes` | 속성 필터 (>=, <=, contains) — `{total, showing}` 반환으로 카운팅 정확 |
 | `aggregate_nodes` | GROUP BY + COUNT/SUM/AVG/MAX/MIN + WHERE 사전 필터 |
 | `join_related` | FK 기반 관련 레코드 조회 — RELATED 엣지 순회 (O(degree)) |
+| `top_nodes` | “가장 X한”, “top N”, “최대/최소”, “최근” 질의를 단일 호출로 처리 |
 
-### 인제스트 / CDC 도구 (v0.14.0+)
+### 인제스트 / CDC 도구
 대화 중에 Claude가 새 자료를 배울 수 있도록 하는 도구.
 
 | 도구 | 용도 |
@@ -255,7 +260,7 @@ Synaptic은 primitive를 제공하고, 합성 여부는 사용자가 선택합
 | `knowledge_ingest_path` | 로컬 CSV/JSONL/TXT 파일 단건 인제스트 |
 | `knowledge_remove` | 단건 노드 삭제 (엣지 cascade) |
 | `knowledge_sync_from_database` | CDC 증분 동기화 |
-| `knowledge_backfill` | 누락된 임베딩·phrase 허브 복구 (v0.14.4+) |
+| `knowledge_backfill` | 누락된 임베딩·phrase 허브 복구 |
 
 ### 탐색 도구
 | 도구 | 용도 |
@@ -301,88 +306,64 @@ graph.reranker_weights = RerankerWeights(
 await graph.reinforce([node_id], success=True)  # 이 결과가 도움 됨 → 다음엔 상위로
 ```
 
----
+### Memory operating layer
 
-## 벤치마크
+검색을 항상 stateful하게 만들지 않고도 관찰할 수 있습니다.
 
-아래 숫자는 버전 태그가 붙은 역사적 스냅샷입니다. 정확한 재현은 링크된
-스크립트/보고서를 기준으로 하고, 지속적으로 정리되는 결과 로그는
-[`docs/comparison/synaptic_results.md`](docs/comparison/synaptic_results.md)를 보세요.
+```python
+from synaptic import FeedbackSignal, MemoryScope
 
-### 재현 가능한 임베더-프리 베이스라인 (노트북 2초)
+scope = MemoryScope(workspace_id="docs", user_id="alice")
+result = await graph.search("환불 예외", record=True, scope=scope)
 
-```bash
-pip install "synaptic-memory[korean]"
-python examples/benchmark_allganize.py
+await graph.record_feedback(
+    event_id=result.event_id,
+    signal=FeedbackSignal.EXPLICIT_POSITIVE,
+    success=True,
+    scope=scope,
+)
+
+health = await graph.memory_health(scope=scope)
+signals = await graph.scan_memory_signals(scope=scope)
 ```
 
-결정론적 출력 (v0.16.0):
+event, feedback, provenance, health signal은 그래프 metadata로 저장됩니다.
+`Node.content`에 섞지 않고, LLM prompt에 원본 metadata 전체를 자동으로 밀어 넣지도 않습니다.
 
-```
-Dataset                  Corpus  Queries      MRR     R@10        Hit     Time
---------------------------------------------------------------------------------
-Allganize RAG-ko            200      200    0.947    1.000   200/200     9.3s
-Allganize RAG-Eval          300      300    0.911    0.950   285/300     5.9s
-```
+---
 
-임베더·reranker 없이 **EvidenceSearch 파이프라인(BM25 + PPR + MMR)** 만으로
-낸 결과입니다. 전체 소스:
-[`examples/benchmark_allganize.py`](examples/benchmark_allganize.py).
-원본: [allganize/RAG-Evaluation-Dataset-KO](https://huggingface.co/datasets/allganize/RAG-Evaluation-Dataset-KO).
+## 벤치마크와 보고서
 
-> **v0.16.0 개선 누적**. v0.15.1의 query-mode Kiwi + v0.16.0의 engine default
-> 전환으로 Korean 공개 벤치마크 전반에서 MRR이 **+0.22~+0.33 상승**. 영어
-> HotPotQA-24도 **+0.148**. 자세한 ablation:
-> [`examples/ablation/run_ablation.py`](examples/ablation/run_ablation.py).
+루트 README는 현재 설치 경로와 공개 API를 기준으로 유지합니다. 상세 수치는
+오래된 측정값이 현재 API 계약처럼 보이지 않도록 버전이 붙은 보고서에 둡니다.
 
-### 영어 multi-hop 표준 벤치마크 (500q subset, v0.16.0)
+빠른 로컬 smoke:
 
 ```bash
-pip install "synaptic-memory[eval]"
-python examples/ablation/download_benchmarks.py
-python examples/ablation/run_tier1_benchmarks.py --subset 500
+synaptic-quickstart --json
 ```
 
-| 데이터셋 | 소스 | Docs | MRR@10 | R@5 | Hit@10 |
-|---------|------|------|--------|-----|--------|
-| HotPotQA dev (distractor) | HuggingFace | 66,635 | **0.784** | 0.585 | 459/500 (91.8%) |
-| 2WikiMultihopQA dev | HuggingFace | 56,687 | **0.795** | 0.501 | 456/500 (91.2%) |
-| MuSiQue-Ans dev | HuggingFace | 21,100 | 0.590 | 0.379 | 381/500 (76.2%) |
+가벼운 한국어 FTS 벤치마크:
 
-HippoRAG2 등 선행 연구와의 해석은
-[docs/comparison/synaptic_results.md](docs/comparison/synaptic_results.md#tier-15--english-multi-hop-standard-benchmarks-v0160)
-참고. 직접 비교 지표가 아니므로 엄격한 head-to-head는 조심하지만,
-**MuSiQue R@5 0.379 vs HippoRAG2 0.747 격차**는 embedder 도입 시 좁힐 여지.
-
-### 전체 파이프라인 (임베더 + reranker, pre-v0.16.0 측정)
-
-아래 수치는 **v0.16.0 engine flip 이전** 측정값으로, **EvidenceSearch + 임베더
-(Ollama `qwen3-embedding:4b`) + cross-encoder reranker(TEI `bge-reranker-v2-m3`)**
-조합이었습니다. 재현에는 GPU 환경이 필요합니다. 전체 harness:
-[`eval/run_all.py`](eval/run_all.py). v0.16.1에서 재측정 예정.
-
-### 단일 검색
-
-| 데이터셋 | 유형 | 노드 | MRR | Hit |
-|---------|------|------|-----|-----|
-| KRRA Easy | 한국어 문서 (비공개) | 19,720 | **0.967** | 20/20 |
-| KRRA Hard | 한국어 문서 (비공개) | 19,720 | **1.000** | 15/15 |
-| X2BEE Easy | PostgreSQL 이커머스 (비공개) | 19,843 | **1.000** | 20/20 |
-| assort Easy | 패션 CSV (비공개) | 13,909 | **0.867** | 13/15 |
-| HotPotQA-24 | 영어 multi-hop (공개 서브셋) | 226 | **0.964** | 24/24 |
+```bash
+pip install "synaptic-memory[korean]"
+python examples/benchmark_allganize.py
+```
 
-> HotPotQA-24는 24문항 서브셋. 전체 HotPotQA-dev(7,405q) 전체 실행은
-> v0.16.1에서 예정 (PPR O(corpus) 최적화 후).
+선택 패키지/API key가 준비된 경우 competitor harness:
 
-### 멀티턴 에이전트 (GPT-4o-mini, 최대 5턴)
+```bash
+python examples/benchmark_vs_competitors/run_comparison.py --only synaptic
+```
 
-| 데이터셋 | 결과 |
-|---------|------|
-| KRRA Hard agent | 10-13/15 (67-87%) |
-| **X2BEE Hard agent** | **17/19 (89%)** |
-| **assort Hard agent** | **12/15 (80%)** |
+참고 보고서:
 
-필터 / 집계 / FK 조인 / 카운팅 같은 정형 데이터 질의가 그래프 기반 도구로 end-to-end 동작합니다.
+| 보고서 | 내용 |
+|--------|------|
+| [docs/comparison/synaptic_results.md](docs/comparison/synaptic_results.md) | 재현 가능한 Synaptic 벤치마크 결과와 provenance |
+| [docs/REPORT-rag-vs-synaptic.md](docs/REPORT-rag-vs-synaptic.md) | 금융 법령 multi-hop 검색에서 RAG와 synaptic-memory 비교 |
+| [docs/REPORT-memory-operating-layer-eval.md](docs/REPORT-memory-operating-layer-eval.md) | memory operating layer 평가와 health/reporting gate |
+| [examples/benchmark_vs_competitors/README.md](examples/benchmark_vs_competitors/README.md) | competitor adapter 공정성 주의사항 |
 
 ---
 
diff --git a/README.md b/README.md
index fe3b145..a3f7de1 100644
--- a/README.md
+++ b/README.md
@@ -1,6 +1,6 @@
 # Synaptic Memory
 
-**Zero API calls at index time. Zero infra. Zero lock-in.**
+**Default path: zero API calls at index time. Zero infra. Zero lock-in.**
 A knowledge graph + MCP tool server for LLM agents, with hybrid retrieval,
 CDC-based live database sync, and Korean FTS built in.
 
@@ -86,7 +86,7 @@ MCP tools → LLM agent explores via graph-aware multi-turn tool use
 ```
 
 **Two jobs, nothing else:**
-1. **Build the graph well** — cheap extraction, no LLM at index time
+1. **Build the graph well** — cheap deterministic extraction by default
 2. **Give the LLM good tools** — the agent decides what to search
 
 ---
@@ -94,9 +94,12 @@ MCP tools → LLM agent explores via graph-aware multi-turn tool use
 ## Install
 
 ```bash
-# Recommended — covers every example in this README
+# Recommended local graph + MCP setup
 pip install "synaptic-memory[sqlite,korean,vector,mcp]"
 
+# Add this for the LangChain retriever example
+pip install "synaptic-memory[langchain]"
+
 # Or everything, including Postgres / Kuzu / Qdrant / MinIO
 pip install "synaptic-memory[all]"
 ```
@@ -112,6 +115,7 @@ pip install synaptic-memory[vector]        # + usearch HNSW index
 pip install synaptic-memory[mcp]           # + MCP server for Claude
 pip install synaptic-memory[embedding]     # + aiohttp for embedding APIs
 pip install synaptic-memory[reranker]      # + flashrank cross-encoder
+pip install synaptic-memory[langchain]     # + LangChain retriever adapter
 pip install synaptic-memory[postgresql]    # + asyncpg + pgvector
 pip install synaptic-memory[docs]          # + xgen-doc2chunk (PDF/DOCX/PPTX/XLSX/HWP)
 ```
@@ -250,17 +254,17 @@ goes from **32% (standard dense RAG)** / **31% (HippoRAG2)** to
 |----------|-----------------|-----------|
 | GraphRAG-style (MS GraphRAG, Cognee, Graphiti) | LLM extracts entities + relations + community summaries | Highest recall on narrative corpora, but every new document costs LLM tokens |
 | LightRAG-style | LLM deferred to query time | Less index cost, but each query pays |
-| **Synaptic** | **None.** Structural + statistical signals (FK, NEXT_CHUNK, phrase DF hubs, MENTIONS) + rule-based REFERENCES edges | Cheapest, deterministic; extracts explicit cross-references but won't *LLM-synthesize* fuzzy semantic relations |
+| **Synaptic default** | **None.** Structural + statistical signals (FK, NEXT_CHUNK, phrase DF hubs, MENTIONS) + rule-based REFERENCES edges | Cheapest, deterministic; extracts explicit cross-references without LLM calls |
 
-No LLM at indexing. The graph is a search index, not a knowledge base.
+By default, indexing is LLM-free. The graph is a search index, not a knowledge base.
 Cross-references that documents state explicitly (statute article
 citations, clause codes) are turned into `REFERENCES` edges with
-zero LLM (see [Option D](#option-d-relation-ontology-multi-hop)). What
-Synaptic does *not* do is LLM-synthesize fuzzy semantic relations or
-community summaries — if you need those, layer them with your own
-agent; Synaptic gives you the primitives.
+zero LLM (see [Option D](#option-d-relation-ontology-multi-hop)).
+If you opt into OpenIE, Synaptic can add bounded, revertible
+LLM-extracted semantic relations; that is not part of the default
+deterministic path.
 
-> **v0.28+**: `graph.search()` has one retrieval path: the hybrid
+> **Current API**: `graph.search()` has one retrieval path: the hybrid
 > EvidenceSearch pipeline (BM25 + HNSW + PPR + cross-encoder + MMR).
 > The old `engine=` switch was removed, so examples should call
 > `graph.search("question")` directly.
@@ -274,7 +278,7 @@ agent; Synaptic gives you the primitives.
 |------|---------|
 | `deep_search` | **Recommended.** Search → expand → read documents in ONE call |
 | `compare_search` | Auto-decompose multi-topic queries, search in parallel |
-| `knowledge_search` | Core semantic search (routes through EvidenceSearch in v0.14.2+) |
+| `knowledge_search` | Core semantic search through EvidenceSearch |
 | `agent_search` | FTS + vector hybrid search with intent routing |
 | `expand` | 1-hop graph neighbours |
 | `get_document` | Full document with query-relevant chunks |
@@ -287,13 +291,13 @@ agent; Synaptic gives you the primitives.
 | `filter_nodes` | Property filter (>=, <=, contains) — returns `{total, showing}` for accurate counting |
 | `aggregate_nodes` | GROUP BY + COUNT/SUM/AVG/MAX/MIN with optional WHERE pre-filter |
 | `join_related` | FK-based related record lookup — walks RELATED edges (O(degree)) |
-| `top_nodes` | **Single-call top-N ranking** — "가장 X한" / "top N" / "최대/최소" / "최근" questions without composing aggregate_nodes. Each row carries `sort_value` for chaining into join_related / filter_nodes(from_ids=...). v0.18-β2+. |
+| `top_nodes` | **Single-call top-N ranking** — "가장 X한" / "top N" / "최대/최소" / "최근" questions without composing aggregate_nodes. Each row carries `sort_value` for chaining into join_related / filter_nodes(from_ids=...). |
 
 All four structured tools emit `hints` on 0-result returns (alternate
 operator, dropped WHERE, fuzzy column match) so the agent's next turn
 gets a concrete corrective action instead of a retry loop.
 
-### Ingest / CDC tools (v0.14.0+)
+### Ingest / CDC tools
 Mid-conversation ingestion so Claude can teach itself new material without leaving the chat.
 
 | Tool | Purpose |
@@ -304,7 +308,7 @@ Mid-conversation ingestion so Claude can teach itself new material without leavi
 | `knowledge_ingest_path` | Ingest a CSV / JSONL / text file from the local filesystem |
 | `knowledge_remove` | Delete a single node with edge cascade |
 | `knowledge_sync_from_database` | Incremental sync from a live database (CDC) |
-| `knowledge_backfill` | Repair graphs missing embeddings or phrase hubs (v0.14.4+) |
+| `knowledge_backfill` | Repair graphs missing embeddings or phrase hubs |
 
 ### Navigation tools
 | Tool | Purpose |
@@ -351,209 +355,66 @@ graph.reranker_weights = RerankerWeights(
 await graph.reinforce([node_id], success=True)  # this result helped → lift it next time
 ```
 
----
-
-## Benchmarks
+### Memory operating layer
 
-The numbers below are version-tagged historical snapshots. Use the linked
-scripts/reports for exact reproduction, and prefer
-[`docs/comparison/synaptic_results.md`](docs/comparison/synaptic_results.md)
-for a continuously curated result log.
+Retrieval can be observed without making every search stateful:
 
-### Reproducible FTS-only baseline (< 2 seconds on a laptop)
+```python
+from synaptic import FeedbackSignal, MemoryScope
 
-```bash
-pip install "synaptic-memory[korean]"
-python examples/benchmark_allganize.py
-```
+scope = MemoryScope(workspace_id="docs", user_id="alice")
+result = await graph.search("refund exception", record=True, scope=scope)
 
-Output (deterministic, v0.16.0):
+await graph.record_feedback(
+    event_id=result.event_id,
+    signal=FeedbackSignal.EXPLICIT_POSITIVE,
+    success=True,
+    scope=scope,
+)
 
+health = await graph.memory_health(scope=scope)
+signals = await graph.scan_memory_signals(scope=scope)
 ```
-Dataset                  Corpus  Queries      MRR     R@10        Hit     Time
---------------------------------------------------------------------------------
-Allganize RAG-ko            200      200    0.947    1.000   200/200     9.3s
-Allganize RAG-Eval          300      300    0.911    0.950   285/300     5.9s
-```
-
-This is the **embedder-free baseline** (EvidenceSearch pipeline: BM25 +
-PPR + MMR, no vector index, no cross-encoder). Full source:
-[`examples/benchmark_allganize.py`](examples/benchmark_allganize.py).
-Data source: [allganize/RAG-Evaluation-Dataset-KO](https://huggingface.co/datasets/allganize/RAG-Evaluation-Dataset-KO).
-
-> **v0.16.0 — engine default flipped to `"evidence"`.** Combined with
-> the v0.15.1 query-mode Kiwi improvement, FTS-only Korean retrieval
-> moved from **Allganize RAG-ko MRR 0.621 (v0.15.0) → 0.947
-> (v0.16.0)** without any embedder or reranker. English (HotPotQA-24)
-> 0.727 → 0.875. Full ablation:
-> [`examples/ablation/run_ablation.py`](examples/ablation/run_ablation.py).
-> Reproducibility under streaming ingest sharpens in lockstep: top-1
-> rank invariance rose from 54.5 % → **100 %**, bit-wise top-10
-> identical from 51.5 % → **96 %**, with MRR drift exactly zero.
 
-### RAG vs HippoRAG2 vs synaptic-memory — multi-hop retrieval (v0.25)
+Events, feedback, provenance, and health signals are stored as graph metadata.
+They are not appended to `Node.content`, and they are not automatically dumped
+into LLM prompts.
 
-RAG (chunk → top-k → one LLM answer), HippoRAG2 (NeurIPS'24 LLM-entity
-graph + PPR), and the synaptic-memory agent measured head-to-head on
-the same corpus, ground truth, and strict scoring (`finreg` — 4,417
-Korean financial-statute articles, 120 multi-hop queries):
-
-```
-System                        multi-hop solved
----------------------------------------------------
-RAG  (FTS / BM25)                   0/120   (0%)
-RAG  (dense, bge-m3)               39/120  (32%)
-HippoRAG2 (LLM graph + PPR)        37/120  (31%)
-synaptic-memory (agent)            88/120  (73%)
-```
+---
 
-A statute article that cites another article ("제30조에 따라 …") is a
-**multi-hop** query: the cited provision shares no query vocabulary, so
-single-shot retrieval structurally cannot reach it. Notably, **HippoRAG2
-ties plain dense RAG (31% ≈ 32%)** — an LLM-extracted *entity* graph
-cannot capture an exact *structural* citation. synaptic-memory turns
-cross-references into `REFERENCES` graph edges
-([`StructuralReferenceLinker`](src/synaptic/extensions/structural_reference_linker.py),
-LLM-free, auto-derived from the corpus) and the agent follows them —
-2.3× the nearest competitor. Full report, model notes, and limits:
-[`docs/REPORT-rag-vs-synaptic.md`](docs/REPORT-rag-vs-synaptic.md).
+## Benchmarks And Reports
 
-### Head-to-head vs Mem0 / Cognee / HippoRAG2
+The root README stays current with the install path and public API. Detailed
+numbers are versioned in reports so old measurements do not look like the
+current API contract.
 
-A runnable harness (BEIR-style corpus, same MRR / R@10 scoring code
-for every system):
+Run a quick local smoke:
 
 ```bash
-# Synaptic only (no API keys, ~2 s)
-python examples/benchmark_vs_competitors/run_comparison.py --only synaptic
-
-# Adapters for Mem0, Cognee, HippoRAG2 ship in-tree — add them to
-# --only when the respective packages are installed and API keys are set
-python examples/benchmark_vs_competitors/run_comparison.py --only synaptic,mem0 --subset 10
+synaptic-quickstart --json
 ```
 
-See [examples/benchmark_vs_competitors/README.md](examples/benchmark_vs_competitors/README.md)
-for fairness caveats. Competitor self-reported numbers (Mem0 LoCoMo
-91.6, HippoRAG2 MuSiQue F1 51.9, the Zep 84→58 correction incident,
-etc.) are catalogued with sources in
-[docs/comparison/published_numbers.md](docs/comparison/published_numbers.md).
-
-### Embedder-free full-dataset summary (v0.16.0)
-
-Run via `python examples/ablation/run_ablation.py`:
-
-| Dataset | Lang | Queries | MRR | Hit @ 10 |
-|---------|------|---------|-----|----------|
-| Allganize RAG-ko | ko | 200 | **0.947** | 200/200 |
-| Allganize RAG-Eval | ko | 300 | **0.911** | 285/300 |
-| AutoRAG KO | ko | 114 | **0.906** | 114/114 |
-| PublicHealthQA KO | ko | 77 | **0.546** | 64/77 |
-| HotPotQA-24 EN | en | 24 | **0.875** | 24/24 |
-
-### English multi-hop standard benchmarks (v0.16.0, subset)
-
-Run via:
+Run the lightweight Korean FTS benchmark:
 
 ```bash
-pip install "synaptic-memory[eval]"   # adds `datasets` for HuggingFace download
-python examples/ablation/download_benchmarks.py
-python examples/ablation/run_tier1_benchmarks.py --subset 500
+pip install "synaptic-memory[korean]"
+python examples/benchmark_allganize.py
 ```
 
-Adds HotPotQA-dev (66 k corpus), MuSiQue-Ans-dev (21 k), and
-2WikiMultihopQA-dev (57 k) — the three retrieval corpora the
-HippoRAG / GraphRAG line of research uses for head-to-head.
-Numbers go in
-[docs/comparison/synaptic_results.md](docs/comparison/synaptic_results.md#tier-15--english-multi-hop-standard-benchmarks-v0160).
-
-### Full pipeline (BGE-M3 + cross-encoder) — v0.17.1
-
-Measured 2026-04-19 on an H100, `BAAI/bge-m3` + `BAAI/bge-reranker-v2-m3`
-loaded in-process via `transformers` (no external TEI / Ollama
-endpoint needed):
+Run the competitor harness when optional packages/API keys are available:
 
 ```bash
-python eval/run_all.py --quick --local-bge
+python examples/benchmark_vs_competitors/run_comparison.py --only synaptic
 ```
 
-#### 14-bench single-shot (5 public + 9 private)
-
-| Dataset | Lang | Queries | FTS-only | Full pipeline | Δ |
-|---------|------|--------:|---------:|--------------:|---:|
-| HotPotQA-24 | en | 24 | 0.875 | **0.979** | +0.104 |
-| Allganize RAG-ko | ko | 200 | 0.947 | **0.983** | +0.036 |
-| Allganize RAG-Eval | ko | 300 | 0.911 | **0.955** | +0.044 |
-| PublicHealthQA | ko | 77 | 0.547 | **0.748** | +0.201 |
-| AutoRAG KO | ko | 114 | **0.906** | 0.806 | −0.100 ⚠ |
-| KRRA Easy | ko | 20 | 0.967 | **0.975** | +0.008 |
-| KRRA Hard | ko | 40 | 0.583 | **0.589** | +0.006 |
-| KRRA Conv | ko | 30 | 0.146 | **0.166** | +0.020 |
-| assort Easy | ko | 15 | 0.760 | **0.856** | +0.096 |
-| assort Hard | ko | 40 | 0.000 | 0.000 | 0 |
-| assort Conv | ko | 30 | 0.425 | **0.472** | +0.047 |
-| X2BEE Easy | en | 20 | 1.000 | 1.000 | 0 |
-| X2BEE Hard | en/ko | 20 | **0.379** | 0.368 | −0.011 |
-| X2BEE Conv | en/ko | 30 | 0.167 | 0.164 | −0.003 |
-| **Mean** | | | 0.615 | **0.647** | **+0.032 (+5.2pp)** |
-
-v0.17.1 is the **first release where mean Full-pipeline MRR exceeds
-mean FTS-only MRR** across all 14 benches (v0.17.0 was net −1.1 %).
-12/14 benches improve or hold; the 3 mild residual regressions
-(X2BEE Hard / Conv, AutoRAG) are at or below single-query noise
-except for AutoRAG where the structural reranker mismatch persists.
-
-**When the cross-encoder helps — and when it hurts.** The reranker
-shines on paraphrase-heavy corpora (PublicHealthQA +20 pp,
-HotPotQA +10 pp). On retrieval-style corpora where FTS ranking is
-already near-optimal (AutoRAG, X2BEE Hard) it injects noise that
-displaces the gold rank. v0.17.1's adaptive blend (
-`std/3` discriminator) and structured-row reranker skip recover
-most of that — AutoRAG went −0.264 → −0.100 — but pure
-**`reranker=None` is the only way to hit the FTS-only ceiling** on
-those corpora. Diagnostic: [`examples/ablation/diagnose_autorag.py`](examples/ablation/diagnose_autorag.py).
-
-#### Multi-turn agent (Qwen3.5-27B via vLLM, 5 turns, LLM-judge)
-
-| Dataset | Single-shot | Agent solved | vs v0.13 (gpt-4o-mini) |
-|---------|------------:|------------:|----------------------:|
-| KRRA Hard | 0.589 | 30/39 (77 %) | 11/15 (73 %) +4 pp |
-| assort Hard | 0.000 | 30/33 (91 %) | 13/15 (87 %) +4 pp |
-| X2BEE Hard | 0.368 | **19/19 (100 %)** | 17/19 (89 %) +11 pp |
-| KRRA Conv | 0.166 | 14/30 (47 %) | 21/30 (70 %) −23 pp |
-| assort Conv | 0.472 | 22/24 (92 %) | 20/24 (83 %) +9 pp |
-| X2BEE Conv | 0.164 | 25/27 (93 %) | 22/27 (81 %) +12 pp |
-| **Mean** | | **140/172 = 81 %** | |
-
-**This is Synaptic's real number.** Single-shot retrieval is the
-floor; the multi-turn agent (`deep_search` + `compare_search` +
-graph-context injection) brings the same questions from 0–47 %
-to 47–100 %. assort Hard 0/40 → 91 % under agent shows what the
-graph + structured tools can do that no single-shot pipeline can.
-5 of 6 benches beat the v0.13 GPT-4o-mini baseline (Qwen3.5-27B
-upgrade). KRRA Conv regression (Qwen Korean conversational gap)
-is the one open issue — v0.18 track.
-
-#### Known structural gap — MuSiQue (English multi-hop)
-
-MuSiQue-Ans-dev 500q full pipeline R@5 **0.453** vs HippoRAG2's
-published **0.747** (−0.294). Three rounds of targeted fixes (LLM
-query decomposition, inline phrase hub, DF-filtered entity linker)
-all regressed the score — the gap is structural. Closing it
-requires OpenIE triple extraction + query→triple dense linking.
-That work now lives as an opt-in, bounded, revertible layer rather
-than a default deterministic path. Synaptic's strength is Korean / structured-data RAG;
-English Wikipedia multi-hop is honestly documented as a trade-off.
-See the OpenIE and memory operating layer plans under `docs/PLAN-v0.30+`.
-
-### Multi-turn agent (GPT-4o-mini, 5 turns max)
-
-| Dataset | Result |
-|---------|--------|
-| KRRA Hard agent | 10-13/15 (67-87%) |
-| **X2BEE Hard agent** | **17/19 (89%)** |
-| **assort Hard agent** | **12/15 (80%)** |
-
-Structured data queries (filter / aggregate / FK join / count) work end-to-end through graph-aware tools.
+Reference reports:
+
+| Report | What it covers |
+|--------|----------------|
+| [docs/comparison/synaptic_results.md](docs/comparison/synaptic_results.md) | Reproducible Synaptic benchmark results with provenance |
+| [docs/REPORT-rag-vs-synaptic.md](docs/REPORT-rag-vs-synaptic.md) | RAG vs synaptic-memory on multi-hop financial-statute retrieval |
+| [docs/REPORT-memory-operating-layer-eval.md](docs/REPORT-memory-operating-layer-eval.md) | Memory operating layer evaluation and health/reporting gates |
+| [examples/benchmark_vs_competitors/README.md](examples/benchmark_vs_competitors/README.md) | Fairness caveats for competitor adapters |
 
 ---
 
@@ -599,6 +460,8 @@ Agent tools → MCP server → LLM agent
 | `embedding` | aiohttp for embedding API calls |
 | `mcp` | MCP server for Claude Desktop/Code |
 | `sqlite` | aiosqlite backend |
+| `langchain` | LangChain retriever adapter |
+| `postgresql` | asyncpg + pgvector |
 | `docs` | xgen-doc2chunk for PDF/DOCX/PPTX/XLSX/HWP loading |
 
 ---