diff --git a/core/src/ldml/C9134_ldml_markers.md b/core/src/ldml/C9134_ldml_markers.md
new file mode 100644
index 0000000000..7a0f5d6959
--- /dev/null
+++ b/core/src/ldml/C9134_ldml_markers.md
@@ -0,0 +1,80 @@
+# LDML Markers in Keyman
+
+## Background
+
+- This is the proposed implementation of markers for Keyman
+
+> Markers are placeholders which record some state, but without producing normal visible text output. They were designed particularly to support dead-keys.
+
+- See [tr35-keyboards-markers][] for source spec.
+    - [CLDR-16837][] has been filed to capture feedback on the marker spec, please add to that with any commnets.
+
+### Format
+
+- The general format is `\m{id}` where `id` is presumed to be `[0-9A-Za-z_]{1,32}`
+- `\m{.}` in matching matches a single marker.
+
+### Location
+
+Markers can appear in both 'emitting' and 'matching-only' areas:
+
+#### Emitting
+
+- `key to=` for emitting markers
+- `transform to=` to emit markers
+- `string value=` for matching or emitting markers
+
+#### Match only
+
+- `transform from=` to match markers
+- `transform after=` to match markers
+- `display to=` for matching keys which contain markers
+
+## Theory / Encoding
+
+- Keyman already uses U_SENTINEL `U+FFFF` (noncharacter)
+- The general proposal here is to use the sequence `U+FFFF U+XXXX` to represent marker #XXXX (starting with `U+0001`)
+- `U+FFFF` cannot otherwise occur in text, so it is unique
+- `U+FFFF U+FFFF` to indicate 'any marker'  corresponds to `\m{.}`
+- This scheme allows for 65,534 (0xFFFE) unique markers, from `U+FFFF U+0001` through `U+FFFF U+FFFE`
+
+## Terminology
+- A marker's "number" is its position in the `markers` list, starting at index 1 (U+0001) being the first element in that list.
+Note that this is different from other 0-based indices in KMX+. If there are three markers in a keyboard file, they will be numbered 1, 2, 3.
+- A marker's "id" is its string name such as `a` or `acute`
+
+## Compiler (kmc)
+
+- `U+FFFF` needs to be illegal as a literal or escaped sequence. So `\u{FFFF}` is not allowed, for example, nor as a literal in the UTF-8 .xml stream.
+
+### `vars`
+
+- `vars` compiler will prescan all of the elements listed under [Emitting](#emitting), but not match-only.
+- All found markers will be added to a single master `list`, which will be added under the `vars.markers` section.
+- Subsequent processes/compilers will be able to check this list for any missing markers. For example, a `<display to='\m{nonExistentMarker}'/>` can then trigger an error failing validation, because nothing is defined which emits that marker.
+
+### Other sections
+
+- `string value='\m{…}'` will simply store `\m{…}`, for application when expanded as with other variables.
+- Other emitters, such as `key`, `transform` will include the string `U+FFFF U+XXXX` where XXXX corresponds to the marker's 1-based number.
+- Transforms will need to match against the marker or markers desired, so may need to emit sequences such as `(?:\uFFFF\u0123)` meaning a match to marker #0x0123
+- matching `\m{.}` may then expand to `(?:\uFFFF.)`
+
+## Binary (.kmx plus)
+
+- The `vars.markers` is a pointer into the `list` section with a list (binary order) of the marker names
+- Other strings will be in `U+FFFF U+0123` form etc. as if it was in the original text stream as such.
+
+## Implementation (core)
+
+- Core needs to recognize `U+FFFF …` sequences and convert them to markers in the context stream, with `state->context().push_marker(marker_number)`
+- For normal processing, Core does _not_ need to correlate the marker _number_ with a marker _id_, although this would be helpful for a debugging or tracing facility.  I.e. `U+FFFF U+0123` corresponding to entry 0x0123 in the `vars.markers` -> `list` table.
+- Core needs to remove `U+FFFF …` sequences before they are passed to the OS.
+
+- Transform processing needs to recognize these markers in the context stream and pass them to the transforms appropriately.
+- User-defined backspace processing `<transform type="backspace"/> may specifically operate on backspaces in the context stream, just as with other transform processing.
+- The default backspace processing needs to recognize these markers in the context stream and remove them as appropriate.
+
+
+[tr35-keyboards-markers]: https://github.com/unicode-org/cldr/blob/keyboard-preview/docs/ldml/tr35-keyboards.md#markers
+[CLDR-16837]: https://unicode-org.atlassian.net/browse/CLDR-16837