diff --git a/README.md b/README.md
index f822c16..2ee0474 100644
--- a/README.md
+++ b/README.md
@@ -47,6 +47,12 @@ print(record)
 
 It also offers a interactive shell `moneywiz-cli`.
 
+## Read completeness
+
+The read-completeness API documents whether records and relationships were
+fully observed, partially interpreted, or not loaded. See
+[Read completeness](docs/read-completeness.md) for its contract and usage.
+
 ## Tests
 
 Run unit tests without a database:
@@ -67,6 +73,9 @@ MONEYWIZ_TEST_DB_PATH=/absolute/path/to/test.sqlite uv run pytest tests
 The default remains an eager load of every manager. Use an explicit manager list
 to isolate a read from unrelated malformed records:
 
+See [Read completeness](docs/read-completeness.md) for the complete public
+contract, status meanings, safe diagnostics, and snapshot guarantees.
+
 ```python
 from pathlib import Path
 
diff --git a/docs/read-completeness.md b/docs/read-completeness.md
new file mode 100644
index 0000000..073b487
--- /dev/null
+++ b/docs/read-completeness.md
@@ -0,0 +1,113 @@
+# Read completeness
+
+## Purpose
+
+Reading a database is not sufficient to establish that the returned records are
+complete. A row or relationship may be malformed, use an unsupported schema
+shape, or fail validation. Treating the remaining parsed records as a complete
+result can make an absent record indistinguishable from an unreadable one.
+
+The read-completeness API makes that distinction explicit. It reports what was
+observed, what was parsed, and identity-only diagnostics for source rows that
+could not be interpreted. It does not expose the rejected row's payload.
+
+## Public contract
+
+`MoneywizApi` provides three related operations:
+
+| Operation | Result |
+| --- | --- |
+| `load(managers=...)` | Reloads the requested managers and returns their aggregate completeness evidence. |
+| `completeness(managers=...)` | Returns completeness evidence for managers that have already loaded. |
+| `snapshot(managers=...)` | Returns JSON-safe records, completeness evidence, and the selected schema profile. |
+
+The supported manager names are `accounts`, `payees`, `categories`,
+`transactions`, `investment_holdings`, and `tags`.
+
+```python
+from pathlib import Path
+
+from moneywiz_api import MoneywizApi
+
+with MoneywizApi(
+    Path("/absolute/path/to/moneywiz.sqlite"),
+    managers=("accounts", "transactions"),
+) as api:
+    completeness = api.completeness().as_dict()
+    snapshot = api.snapshot().as_dict()
+```
+
+The returned records remain regular library data. The completeness report is
+separate evidence about how those records were obtained; it does not transform
+model `as_dict()` compatibility output.
+
+## Reading the result
+
+Each manager has a `ManagerLoadReport` with the following states:
+
+| Status | Meaning |
+| --- | --- |
+| `complete` | Every observed source record and required relationship was parsed. |
+| `partial` | At least one observed record or relationship could not be fully interpreted. |
+| `error` | Observation occurred, but no usable result could be published. |
+| `unloaded` | The manager has not completed a load; this is different from a valid empty result. |
+
+For each manager, the report includes source and parsed IDs, counts, skipped
+record diagnostics, and relationship reports. A skipped-record diagnostic
+contains only the source identity, entity name, bounded error kind, and exception
+type. It intentionally omits the source values that failed parsing.
+
+Transaction relationship reports cover category assignments, refunds, and tags.
+Their storage state is `present`, `absent`, or `unknown`. An unknown or malformed
+relationship layout prevents the containing manager from being complete.
+
+## Using partial results safely
+
+Partial data can still be useful for inspection, reporting, or diagnosing an
+unsupported database layout. It must not be silently presented as a complete
+answer.
+
+```python
+report = api.completeness()
+if not report.complete:
+    raise RuntimeError(f"read is {report.status}; inspect diagnostics first")
+
+snapshot = api.snapshot().as_dict()
+```
+
+Callers choose their own policy for partial results. The library reports the
+evidence and preserves parsed records; it does not infer the caller's intent or
+discard usable data solely because another row failed.
+
+## Consistent observation
+
+Database access is read-only. A multi-manager load runs inside one SQLite read
+transaction, so records and completeness evidence describe the same observed
+database generation. A later scoped load refreshes the union of the newly
+requested and previously loaded managers before publishing it.
+
+If a reload fails, the API retains the previous published records and
+completeness evidence. A direct manager load follows the same boundary: it
+publishes records, relationships, and its report together only after successful
+observation.
+
+Schema metadata and physical table definitions are checked before cache-dependent
+reads. If they change, callers must close and reopen the API instead of combining
+current rows with stale schema metadata.
+
+## Snapshot and ownership guarantees
+
+`snapshot()` converts supported values into JSON-safe data and returns detached,
+mutable dictionaries and lists. It rejects values that cannot be represented
+safely, including binary values, arbitrary objects, non-finite floats,
+unsupported mapping keys, and normalized-key collisions.
+
+Published completeness reports and snapshots own immutable nested evidence.
+Mutating a manager or an ordinary caller-owned dictionary after publication cannot
+rewrite previously returned completeness information.
+
+## Compatibility notes
+
+Existing model `as_dict()` methods retain their raw compatibility behavior.
+Read-completeness reporting is an additional public contract for callers that
+need to distinguish a complete observation from a partial one.