diff --git a/README.md b/README.md
index 6c094d9..76a4240 100644
--- a/README.md
+++ b/README.md
@@ -247,14 +247,9 @@ Checked against N1MM 1.0.11031: a log this program wrote opens in N1MM, which
 reads the contest, its categories and the contacts. See
 [`docs/n1mm-interop.md`](docs/n1mm-interop.md) for what that took.
 
-Not yet: voice keying, QTC handling for WAE, call history files, digital modes
-beyond logging them, and the check window's Call History and Exchange columns,
-which are left out rather than shown empty.
+Nothing here has been run against a real radio, keyer, SO2R box or cluster node.
+Every one of those is tested against a fake that speaks the same protocol, which
+proves the two agree with each other and nothing more.
 
-The radio, network and keyer clients are tested against fakes that speak the
-documented protocols — the radio one over a real socket, against a stand-in for
-`rigctld` that answers in the extended form. None has been run against a real
-radio or keyer. The
-cluster client is tested over a real socket against a node fake that sends the
-telnet negotiation, the login prompt and spot lines, but not against a live
-node.
+[`docs/unfinished.md`](docs/unfinished.md) lists what is half-built and what has
+never been tested against the real thing.
diff --git a/docs/unfinished.md b/docs/unfinished.md
new file mode 100644
index 0000000..c9ce077
--- /dev/null
+++ b/docs/unfinished.md
@@ -0,0 +1,71 @@
+# What is not finished, and what has not been tested
+
+Two separate lists. A feature can be finished and untested, or half-built and
+exercised every day. Written 2026-08-27.
+
+## Never run against the real thing
+
+Everything below is written from a published protocol or from N1MM's source and
+tested against a fake that speaks the same protocol. A fake agrees with whatever
+the person who wrote it believed, so these are the places where a wrong belief
+is still sitting there undiscovered.
+
+| What | Tested against | What has never happened |
+|---|---|---|
+| `RigctldRadio` | a stand-in for `rigctld` over a real socket | no real `rigctld`, no real radio. The extended answer format, the `+` prefix and the split commands are from the documented protocol. If the format is wrong, nothing works rather than something being subtly off — check this first. |
+| `OtrspBox` | a `MemoryStream`, checking the bytes | no real SO2R box. Command forms are from N1MM's `N1MMPort.cs`. |
+| `ClusterClient` | a node fake over a real socket, sending the telnet negotiation, the login prompt and spot lines | no live cluster node. Which nodes send bare CR, and which send option negotiation, is guessed from N1MM's code. |
+| `StationNetwork` | the message format, round-tripped | no second station, and no N1MM on the same network. |
+| `CwDaemonSender` | the UDP messages | no `cwdaemon`, no radio keyed. |
+| `WinkeyerSender` | nothing | no test at all, and no WinKeyer. The host-mode open sequence is from the WinKeyer datasheet. |
+
+The Cabrillo output has not been put in front of a contest sponsor's robot.
+
+## Half-built
+
+**Alternating CQ (SO2R).** Needs to know when the keyer has finished sending.
+`MessageSender` is fire-and-forget: it has `SendAsync`, `AbortAsync` and
+`SetSpeedAsync` and no completion signal. Both `cwdaemon` and a WinKeyer can
+report completion, so the interface has to grow first. Timing it from the length
+of the text would be a guess that goes wrong exactly when the contest is busy.
+
+**Voice keying.** `MessageSender` was written to cover a voice keyer playing a
+recording, and nothing implements it. No DVK support either, so the second radio
+of an SO2R station cannot call CQ by voice.
+
+**The check window's Call History and Exchange columns.** Left out rather than
+shown empty: there is no call history file support to fill them from.
+
+**Call history files.** Not read at all. They are what fills a known station's
+name, state or zone in as soon as the call is typed.
+
+**QTC handling for WAE.** Not started. WAE logs will not be complete without it.
+
+**Digital modes.** A contact can be logged as RTTY or another digital mode, and
+the contest rules score it, but there is no digital window: no decoding, no
+transmitting, no interface to fldigi, MMTTY or similar.
+
+**Band-plan colouring of the bandmap scale.** N1MM colours the scale by CW,
+digital and phone segment. The segments differ by ITU region and there is no
+band-plan table here, so it is left out rather than guessed.
+
+**Contest coverage.** Eight contests are built in, plus whatever `.udc` files
+are in the user-defined folder. N1MM ships well over a hundred. Opening a log
+from a contest not in the registry fails with the contest name, which is the
+right answer but it is still a wall.
+
+**The Available Mults and Qs window.** Not started.
+
+## Known rough edges
+
+**The log window's column widths** are the grid's own automatic sizing with a
+character-count floor. A column can change width as rows scroll into view.
+
+**Contacts arriving over the station network** are applied from the socket
+thread, and `AppSession.TakeFromNetwork` reopens the whole contest to do it.
+That is correct but heavy, and it discards what is typed in an entry window on
+another radio.
+
+**Two entry windows and one keyer.** The SO2R box is told which radio to key
+before each message, but nothing stops both radios sending at once if the
+operator asks them to.