diff --git a/CLAUDE.md b/CLAUDE.md
index 390c03e..e60e07d 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -22,6 +22,7 @@ CASE v1.1 公式仕様との照合が必要な場合は `docs/reference/` 配下
 | docs/spec/web-ui.md                              | Web UIパス設計・ツリービュー・詳細ページ・URI生成ルール            |
 | docs/spec/csv-format.md                          | CSVフォーマット仕様（独自形式・OpenSALT形式・簡易形式）           |
 | docs/spec/import-logic.md                        | インポート/エクスポートのビジネスロジック・upsertルール             |
+| docs/spec/round-trip-fidelity.md                 | 経路別（CASE JSON / Excel / CSV）の全フィールド往復忠実度マトリクス |
 | docs/spec/cli.md                                 | CLIコマンド仕様                                   |
 | docs/guide/initial-setup.md                      | 初期データセットアップガイド（テナント作成〜ルーブリック取り込みまで）         |
 | docs/guide/opencase-interop.md                   | OpenCASE 相互運用ガイド（公開/プライベートの配置パターンと運用フロー）         |
diff --git a/docs/spec/csv-format.md b/docs/spec/csv-format.md
index f339e4b..4117755 100644
--- a/docs/spec/csv-format.md
+++ b/docs/spec/csv-format.md
@@ -14,6 +14,11 @@ Three CSV formats are supported. The format is auto-detected on import; on expor
 > (`import xlsx` / `export xlsx`). OpenSALT's CSV importer drops CFItemType / educationLevel,
 > whereas its 3-sheet Excel workbook preserves them. See
 > [round_trip_status.md](../dev/round_trip_status.md) (OpenSALT section).
+>
+> For the full per-field comparison of what each path (CASE JSON / Excel / CSV)
+> preserves or drops, see [round-trip-fidelity.md](round-trip-fidelity.md). The
+> short version: CASE JSON is full fidelity; Excel is an OpenSALT-compatible subset;
+> CSV is a smaller subset (and only expresses `isChildOf` associations).
 
 ## Auto-detection logic
 
diff --git a/docs/spec/round-trip-fidelity.md b/docs/spec/round-trip-fidelity.md
new file mode 100644
index 0000000..8f15be9
--- /dev/null
+++ b/docs/spec/round-trip-fidelity.md
@@ -0,0 +1,207 @@
+# Round-trip Fidelity (CASE JSON vs Excel vs CSV)
+
+Which CASE v1.1 fields survive an import → export round-trip on each interchange
+path. The short version: **CASE JSON (CFPackage) is full fidelity; Excel is an
+OpenSALT-compatible subset; CSV is a smaller subset.** Whether a field is kept or
+lost is a property of the *format's expressiveness*, not of the tool's CASE
+version — if a format has no place to put a field, even a fully v1.1-compliant
+tool drops it on that path.
+
+Related docs:
+[csv-format.md](csv-format.md) (column-level CSV/Excel spec),
+[import-logic.md](import-logic.md) (upsert rules),
+[../guide/opencase-interop.md](../guide/opencase-interop.md) (JSON handoff),
+[../dev/round_trip_status.md](../dev/round_trip_status.md) (JSON losslessness test status).
+
+## How to read the matrices
+
+| Mark | Meaning |
+|------|---------|
+| ✓ | Preserved across the round-trip |
+| △ | Partially preserved (only a sub-field such as `title`, or the value is regenerated) |
+| ✗ | Dropped (no column, hard-coded `None`, or not wired in import/export) |
+| — | Not applicable |
+
+CSV has three profiles. Where they differ, **Custom** and **OpenSALT** get their
+own columns below; the **Simple** profile is positional (auto-assigns
+`identifier`/`sequenceNumber`, keeps only the first few columns) and is covered in
+[Cross-cutting behaviour](#cross-cutting-behaviour).
+
+## TL;DR — what is JSON-only
+
+These are preserved by CASE JSON but lost (or only partially kept) on **both**
+Excel and CSV:
+
+- **`targetType`** (CFAssociation origin/destination, v1.1) — the cross-framework marker.
+- **`notes`** on CFAssociation (v1.1).
+- **`extensions`** on every resource (v1.1).
+- **CFDocument `frameworkType`, `caseVersion`** (v1.1).
+- **CFItem `subject` / `subjectURI`** (v1.1, item-level).
+- **CFConcept** (whole resource), and the non-label fields of **CFItemType /
+  CFSubject / CFLicense / CFAssociationGrouping**.
+- **CFRubric / CFRubricCriterion / CFRubricCriterionLevel** are absent from Excel
+  entirely (CSV handles them only via the dedicated rubric CSV).
+
+Notice the pattern: **almost everything CASE *added* in v1.1 is JSON-only.** The
+flat-file formats predate those fields and have no column for them.
+
+## CFDocument
+
+| Field | JSON | Excel | CSV (Custom) | CSV (OpenSALT) | Notes |
+|-------|:----:|:-----:|:------------:|:--------------:|-------|
+| identifier | ✓ | ✓ | ✓ | ✓ | find-or-create key |
+| title | ✓ | ✓ | ✓ | ✓ | |
+| creator | ✓ | ✓ | ✓ | ✓ | |
+| publisher | ✓ | ✓ | ✓ | ✓ | |
+| description | ✓ | ✓ | ✓ | ✓ | |
+| version | ✓ | ✓ | ✓ | ✓ | |
+| language | ✓ | ✓ | ✓ | ✓ | |
+| adoptionStatus | ✓ | ✓ | ✓ | ✓ | |
+| officialSourceURL | ✓ | ✓ | ✓ | ✓ | |
+| statusStartDate | ✓ | ✓ | ✓ | ✓ | |
+| statusEndDate | ✓ | ✓ | ✓ | ✓ | |
+| notes | ✓ | ✓ | ✓ | ✓ | |
+| subject | ✓ | ✓ | ✓ | ✓ | string array (`\|`-joined in Excel) |
+| subjectURI | ✓ | △ | △ | △ | flat formats keep only the `subject` title (Excel carries it in the subject column → rebuilt via `#subject`); identifier/uri regenerated |
+| licenseURI | ✓ | △ | △ | △ | Excel: `title` + `licenseText` (text exported but **not re-imported**); CSV: `title` only |
+| frameworkType | ✓ | ✗ | ✗ | ✗ | v1.1; no column |
+| caseVersion | ✓ | ✗ | ✗ | ✗ | v1.1; no column |
+| CFPackageURI | ✓ | ✗ | ✗ | ✗ | `.uri` kept verbatim in JSON |
+| extensions | ✓ | ✗ | ✗ | ✗ | v1.1; no column |
+| uri | ✓ | △ | △ | △ | JSON keeps source URI; flat formats rebuild from identifier |
+| lastChangeDateTime | △ | ✗ | ✗ | ✗ | re-stamped on import (see caveats) |
+
+## CFItem
+
+| Field | JSON | Excel | CSV (Custom) | CSV (OpenSALT) | Notes |
+|-------|:----:|:-----:|:------------:|:--------------:|-------|
+| identifier | ✓ | ✓ | ✓ | ✓ | blank → new UUID |
+| fullStatement | ✓ | ✓ | ✓ | ✓ | |
+| humanCodingScheme | ✓ | ✓ | ✓ | ✓ | |
+| abbreviatedStatement | ✓ | ✓ | ✓ | ✓ | |
+| notes | ✓ | ✓ | ✓ | ✓ | |
+| language | ✓ | ✓ | ✓ | ✓ | |
+| conceptKeywords | ✓ | ✓ | ✓ | ✓ | string array |
+| educationLevel | ✓ | ✓ | ✓ | ✓ | string array |
+| CFItemType (label) | ✓ | ✓ | ✓ | ✓ | find-or-create by title |
+| listEnumeration | ✓ | ✓ | ✓ | ✗ | no column in OpenSALT item sheet |
+| alternativeLabel | ✓ | ✗ | ✓ | ✗ | Excel forces `""`; not in OpenSALT |
+| statusStartDate | ✓ | ✗ | ✓ | ✗ | not in OpenSALT item sheet |
+| statusEndDate | ✓ | ✗ | ✓ | ✗ | not in OpenSALT item sheet |
+| licenseURI | ✓ | ✗ | △ | ✗ | Custom CSV: `title` only; Excel column unused; OpenSALT ignores it |
+| subject | ✓ | ✗ | ✗ | ✗ | v1.1 (item-level) |
+| subjectURI | ✓ | ✗ | ✗ | ✗ | v1.1 (item-level) |
+| conceptKeywordsURI | ✓ | ✗ | ✗ | ✗ | only the plain `conceptKeywords` strings survive |
+| CFItemTypeURI | ✓ | △ | △ | △ | only the type label survives; identifier/uri rebuilt from FK |
+| extensions | ✓ | ✗ | ✗ | ✗ | v1.1; no column |
+| CFDocumentURI | ✓ | ✓ | ✓ | ✓ | implicit (item belongs to the document) |
+| uri | ✓ | △ | △ | △ | JSON keeps source URI; flat formats rebuild from identifier |
+| lastChangeDateTime | △ | ✗ | ✗ | ✗ | re-stamped on import |
+
+The parent/child hierarchy (`isChildOf` + `sequenceNumber`) is **not** a CFItem
+field but is carried alongside items: Excel uses the `smartLevel` column, CSV
+Custom uses `parentIdentifier`/`sequenceNumber`, CSV Simple uses indentation.
+
+## CFAssociation
+
+| Field | JSON | Excel | CSV | Notes |
+|-------|:----:|:-----:|:---:|-------|
+| identifier | ✓ | ✓ | ✗ | CSV regenerates a UUID each round-trip |
+| associationType | ✓ (all) | ✓ (all) | ✗ | **CSV can only express `isChildOf`** (parent/child); Excel carries every type in the CF Association sheet |
+| originNodeURI.uri | ✓ | ✓ | ✓ | CSV value is generated for the isChildOf link |
+| originNodeURI.identifier | ✓ | ✓ | ✓ | |
+| originNodeURI.title | ✓ | ✗ | △ | Excel stores uri + identifier only |
+| **originNodeURI.targetType** | ✓ | ✗ | ✗ | v1.1; CSV hard-codes `None` |
+| destinationNodeURI.uri | ✓ | ✓ | ✓ | |
+| destinationNodeURI.identifier | ✓ | ✓ | ✓ | |
+| destinationNodeURI.title | ✓ | ✗ | △ | |
+| **destinationNodeURI.targetType** | ✓ | ✗ | ✗ | v1.1; CSV hard-codes `None` |
+| sequenceNumber | ✓ | ✗ | ✓ | CSV: only on isChildOf rows |
+| CFAssociationGroupingURI | ✓ | ✓ | ✗ | Excel carries grouping identifier + title |
+| notes | ✓ | ✗ | ✗ | v1.1; no column |
+| extensions | ✓ | ✗ | ✗ | v1.1; no column |
+| CFDocumentURI | ✓ | ✓ | ✓ | implicit |
+| uri | ✓ | ✓ | ✗ | CSV regenerates |
+
+> **Cross-tenant / cross-framework associations** (e.g. `exactMatchOf` pointing at
+> another framework) round-trip **fully in JSON**. In **Excel** the association and
+> its type survive, but `targetType` and the node `title`s are lost. In **CSV** the
+> association cannot be expressed at all (CSV only emits `isChildOf`). Note that
+> Excel splits the work: `isChildOf` lives in the CF Item `smartLevel` column and is
+> *excluded* from the CF Association sheet, which carries every other type.
+
+## CFDefinitions (lookup resources)
+
+Excel has three sheets (CF Doc / CF Item / CF Association) and CSV has item +
+rubric layouts; neither has dedicated sheets/rows for definition resources. They
+survive only as labels embedded in other rows, if at all.
+
+| Resource | JSON | Excel | CSV | What survives on flat-file paths |
+|----------|:----:|:-----:|:---:|----------------------------------|
+| CFItemType | ✓ | △ | △ | `title` label only; `description` / `hierarchyCode` / `typeCode` dropped |
+| CFConcept | ✓ | ✗ | ✗ | nothing (the plain `conceptKeywords` strings are a different thing) |
+| CFSubject | ✓ | △ | △ | `title` only; `hierarchyCode` / `description` dropped |
+| CFLicense | ✓ | △ | △ | `title` (+ `licenseText` exported-only in Excel); `identifier` / `uri` / `description` dropped |
+| CFAssociationGrouping | ✓ | △ | ✗ | Excel: `identifier` + `title`; `description` / `extensions` dropped |
+
+> A definition resource that is **not referenced** by any document/item/association
+> is omitted from a CFPackage export even on the JSON path (the package only emits
+> definitions reachable from the document). Fetch it via the single-resource API if
+> you need the orphan.
+
+## CFRubric / CFRubricCriterion / CFRubricCriterionLevel
+
+Excel has **no rubric sheet** — rubrics are dropped entirely on the Excel path.
+CSV handles rubrics only through the dedicated rubric CSV
+(`import rubric` / `export rubric`), separate from the item CSV.
+
+| Field | JSON | Excel | Rubric CSV | Notes |
+|-------|:----:|:-----:|:----------:|-------|
+| CFRubric.identifier / title / description | ✓ | ✗ | ✓ | |
+| CFRubricCriterion.identifier / category / description / weight / position / rubricId | ✓ | ✗ | ✓ | |
+| CFRubricCriterion.CFItemURI | ✓ | ✗ | △ | CSV keeps the linked item `identifier` only; title/uri rebuilt |
+| CFRubricCriterion.rubricCriterionText | ✗ | ✗ | ✗ | **dropped on every path** — the `rubric_criterion_text_plain` DB column exists but is not wired in import/export |
+| CFRubricCriterionLevel.identifier / description / quality / score / feedback / position / rubricCriterionId | ✓ | ✗ | ✓ | |
+| extensions (all rubric resources) | ✓ | ✗ | ✗ | v1.1 |
+| uri (all rubric resources) | ✓ | ✗ | ✗ | CSV regenerates; levels get a synthetic `urn:csv-import:{uuid}` |
+
+`rubricCriterionText` is the **only** CASE v1.1 field that even the JSON path does
+not round-trip.
+
+## Cross-cutting behaviour
+
+- **`lastChangeDateTime` is re-stamped on import on every path, including JSON.**
+  The DB stores it and export emits it, but import overwrites it with the current
+  time, so the *original* timestamp is not preserved. The JSON losslessness test
+  ignores this field by design (see
+  [round_trip_status.md](../dev/round_trip_status.md)).
+- **`uri`:** JSON import keeps the source `uri` verbatim (resolved gaps A/F/G in
+  the losslessness work). CSV/Excel rebuild every URI from the `identifier`, so the
+  stored URI string does not survive those paths.
+- **`identifier`:** preserved when a column carries it (CFDocument/CFItem
+  identifier on all paths; CFAssociation identifier in Excel). CSV regenerates
+  CFAssociation, definition, and rubric-level UUIDs on each round-trip.
+- **LinkURI sub-fields:** `title`/`identifier`/`uri` of *referenced-resource*
+  LinkURIs (`CFItemTypeURI`, `conceptKeywordsURI`, `licenseURI`,
+  `CFAssociationGroupingURI`, document `licenseURI`) are rebuilt from the FK target
+  in JSON — so they survive if the target is in the same package, reflecting the
+  target's current values. On flat-file paths usually only a label survives.
+- **Association node identifiers** that are UUIDs are lower-cased on JSON import
+  (value-equivalent; cosmetic only).
+- **CSV Simple profile** is positional: it auto-assigns `identifier` and
+  `sequenceNumber` and keeps only the leading columns, dropping most optional
+  fields. Use the **Custom** profile when you need to edit and re-import safely.
+
+## Choosing a path
+
+- **Full fidelity / archival / cross-framework** (need `targetType`, `notes`,
+  `extensions`, definition resources, rubrics): use **CASE JSON (CFPackage)**.
+  This is also what the OpenCASE interop flow uses end-to-end, so OpenCASE ↔
+  COMPEITO keeps everything.
+- **OpenSALT interchange:** use **Excel** (3-sheet workbook) — the highest-fidelity
+  flat-file path, but it still drops `targetType`, `notes`, `extensions`, and
+  rubrics.
+- **Quick human edits / first ingest:** use **CSV (Custom)**. Remember CSV can only
+  express `isChildOf`; other association types need Excel or JSON.
+- **Rubrics over a flat-file path:** only the dedicated **rubric CSV** carries them
+  (there is no Excel rubric sheet).