diff --git a/lexical-sprouting-scroll.md b/lexical-sprouting-scroll.md
new file mode 100644
index 0000000..8007ec1
--- /dev/null
+++ b/lexical-sprouting-scroll.md
@@ -0,0 +1,107 @@
+# Add `Digest` dataclass for machine-readable comparison output
+
+## Context
+
+Currently, `DataFrameComparison.summary()` returns a `Summary` object that renders rich-formatted console output. There's no way to get structured, machine-readable data from a comparison (e.g., for LLM consumption, CI pipelines, or programmatic analysis). This adds a `digest()` method returning a plain dataclass hierarchy that can be serialized via `dataclasses.asdict()` and `json.dumps()`.
+
+## Dataclass Structure
+
+All dataclasses in a new file `diffly/digest.py`:
+
+```python
+@dataclass
+class Digest:
+    equal: bool
+    left_name: str
+    right_name: str
+    primary_key: list[str] | None
+    schemas: DigestSchemas | None       # None when equal, or slim + schemas match
+    rows: DigestRows | None             # None when equal, or slim + rows match
+    columns: list[DigestColumn] | None  # None when equal, no PK, no joined rows, or slim + all match
+    sample_rows_left_only: list[tuple[Any, ...]] | None   # None when no PK or sample_k==0
+    sample_rows_right_only: list[tuple[Any, ...]] | None
+
+@dataclass
+class DigestSchemas:
+    left_only: list[tuple[str, str]]                # (col_name, dtype_str)
+    in_common: list[tuple[str, str, str]]            # (col_name, left_dtype_str, right_dtype_str)
+    right_only: list[tuple[str, str]]
+
+@dataclass
+class DigestRows:
+    n_left: int
+    n_right: int
+    n_left_only: int | None       # None when no primary key
+    n_joined_equal: int | None
+    n_joined_unequal: int | None
+    n_right_only: int | None
+
+@dataclass
+class DigestColumn:
+    name: str
+    match_rate: float
+    changes: list[DigestColumnChange] | None  # None when top_k==0 or column is hidden
+
+@dataclass
+class DigestColumnChange:
+    old: Any
+    new: Any
+    count: int
+    sample_pk: Any | None   # None when show_sample_primary_key_per_change=False
+```
+
+**Design notes:**
+
+- `primary_key` is a top-level field so consumers know what the sample row tuples represent.
+- `sample_rows_left_only` / `sample_rows_right_only` use `list[tuple]` matching the primary key column order.
+- `in_common` uses 3-tuples `(name, left_dtype, right_dtype)` to capture dtype changes (when they match, `left_dtype == right_dtype`).
+- `schemas` is always populated (not `None`) when frames aren't equal and not slim-hidden, even if schemas match -- the caller might want to confirm schemas are identical. **Actually**: mirror `Summary` logic -- `None` when `slim=True` and schemas are equal.
+
+## Files to modify
+
+### 1. New: [diffly/digest.py](diffly/digest.py)
+
+- All dataclass definitions above
+- `_to_python(value)` helper to convert Polars values (date, datetime, timedelta, Decimal) to JSON-safe types
+- Builder function `_build_digest(comparison, **params) -> Digest` containing the logic to extract data from `DataFrameComparison`, mirroring the control flow of `Summary._print_to_console` / `_print_diff`
+- `to_dict()` method on `Digest` via `dataclasses.asdict()`
+- `to_json()` convenience method
+
+### 2. [diffly/comparison.py](diffly/comparison.py) (~line 976)
+
+- Add `digest()` method on `DataFrameComparison` with same signature as `summary()`
+- Lazy import `from .digest import Digest` (same pattern as summary)
+
+### 3. [diffly/cli.py](diffly/cli.py)
+
+- Add `--json` flag (bool, default False)
+- When True, call `comparison.digest(...).to_json()` instead of `comparison.summary(...).format()`
+
+### 4. [diffly/**init**.py](diffly/__init__.py)
+
+- No changes needed -- `Digest` is accessed via `comparison.digest()`, not imported directly. Can revisit later.
+
+### 5. **No changes to** [diffly/testing.py](diffly/testing.py)
+
+- `testing.py` uses `summary()` for human-readable assertion error messages. `digest()` is a data output format, not relevant to assertions.
+
+### 6. New: [tests/test_digest.py](tests/test_digest.py)
+
+- Equal frames -> `equal=True`, all sections `None`
+- Schema differences (left-only, right-only, dtype mismatches in in_common)
+- Row counts with and without primary key
+- Column match rates with `show_perfect_column_matches=True/False`
+- `top_k_column_changes` + `show_sample_primary_key_per_change`
+- `sample_k_rows_only` for `sample_rows_left_only` / `sample_rows_right_only`
+- `slim=True` suppresses matching sections
+- `hidden_columns` hides column changes
+- Validation errors (same as Summary: hidden PK columns, sample PK without top-k)
+- JSON serialization roundtrip: `json.loads(digest.to_json())` is valid
+
+## Verification
+
+```bash
+pixi run pytest tests/test_digest.py -v
+pixi run test
+pixi run pre-commit-run
+```