diff --git a/docs/decisions/0021-what-this-board-publishes.md b/docs/decisions/0021-what-this-board-publishes.md
new file mode 100644
index 0000000..bffbd6c
--- /dev/null
+++ b/docs/decisions/0021-what-this-board-publishes.md
@@ -0,0 +1,87 @@
+# 0021. What this board publishes
+
+## What was decided
+
+This board publishes downloadable release artefacts. They are signed. The
+release notes say, in the notes themselves rather than only here, that the
+artefacts exist for checking this board and are not part of what the
+organisation ships to users.
+
+The reason is the independent-verification argument, which is the whole point of
+handing somebody a runner at all. The runner's job is reading a checkout and
+reporting what it examined, and somebody who has to build that runner from the
+repository they are checking is in a weaker position than somebody who downloads
+it. Most people who might want to check this board's claims have no toolchain.
+Publishing source only would have cost exactly that, and it would have left the
+verification argument resting on the reader trusting the build they made from
+the tree under examination.
+
+The tension with the scope this board opened with is real and is not resolved by
+being unwritten. That scope excludes anything a user is asked to install. What
+separates the two is who the operator is: somebody checking this repository,
+rather than somebody using a media server. The release-notes sentence is what
+carries that distinction to the person holding the file, and its weakness is
+that it depends on being read. Both halves are stated here so that a later
+reader meets the residual rather than the assurance.
+
+The artefacts are signed rather than published with a bare checksum. A checksum
+published next to the file it checksums proves the download arrived intact and
+says nothing about who built it, which is the claim somebody verifying this
+board actually needs. The keys are the ones operations#1609 sets up for the
+working accounts, and that is what makes this one key-custody story rather than
+two: record 0023 requires a verified signature on a commit from the same keys.
+Where those keys are held and how they are rotated belongs to that issue and is
+not restated here, because a custody story written twice is a custody story with
+two answers.
+
+What publishing costs, named rather than discovered later. Signing. Checksums.
+A bill of materials. A vulnerability surface, which a repository publishing
+nothing does not have. And an expectation of continuity, which is the one that
+cannot be withdrawn once somebody is running a version.
+
+## What it applies to
+
+Every release this board publishes, from the commit this record lands on, and
+the issues that build them: the release workflow, the third-party notices and
+bill of materials, the smoke run against a published artefact, and the first
+release.
+
+It does not apply to what an experiment produces. An experiment's prototype is
+not published, is not something anybody is asked to install, and leaves this
+board by promotion, which is
+[0005-how-a-result-leaves.md](0005-how-a-result-leaves.md).
+
+It decides that the artefacts are signed and does not decide what a signature is
+checked with, by whom, or what a reader does when the check fails. A published
+signature nobody verifies is a file beside a file.
+
+## What else was considered
+
+Publishing source only, with the tool run from a checkout.
+
+Publishing artefacts with a checksum beside each one and no signature.
+
+Publishing artefacts with no release-notes sentence separating them from what
+the organisation ships to users.
+
+## What each rejected option would have cost
+
+Source only costs the reader a toolchain, which most people who would want to
+check this board do not have, and it weakens the argument the publishing exists
+to make. Somebody who builds the tool from the repository they are checking has
+verified that the tree builds, not that the tree is what it claims to be. The
+option is cheap for this board and expensive for exactly the person it was
+supposed to serve.
+
+A checksum with no signature costs the claim that matters. It proves the bytes
+arrived as they left and nothing about where they left from, so anybody who can
+place a file can place a checksum beside it. It is also the cheaper option only
+until entry seven of the plan is answered, since a signed commit needs a key
+held somewhere regardless, and taking it would have meant writing the custody
+story once for commits and discovering later that artefacts needed their own.
+
+Publishing with no separating sentence costs the scope exclusion its meaning. An
+artefact with no such note is read as something the organisation ships, which is
+the reading the exclusion was written against, and the cost of correcting that
+reading arrives after somebody has already deployed against it. The sentence is
+weak because it depends on being read; absent, there is nothing to depend on.
diff --git a/docs/decisions/0023-signed-commits-on-the-default-branch.md b/docs/decisions/0023-signed-commits-on-the-default-branch.md
new file mode 100644
index 0000000..84c3cbb
--- /dev/null
+++ b/docs/decisions/0023-signed-commits-on-the-default-branch.md
@@ -0,0 +1,88 @@
+# 0023. A commit on the default branch carries a verified signature
+
+## What was decided
+
+A commit on the default branch of this board has to carry a verified signature.
+It becomes effective as the account keys operations#1609 sets up for the working
+accounts land, and not before, because a rule that refuses every merge before
+anybody holds a key is a rule that gets turned off rather than followed.
+
+The reason is that two claims are otherwise collapsed into one. That an account
+pushed a commit and that a key signed it are different statements, and only the
+second survives somebody else holding the account. Not requiring a signature
+costs nothing to operate and rests authorship entirely on the account, which is
+one credential and one recovery route away from being somebody else's.
+
+The costs are real, they arrive later than the decision does, and each of them
+lands on somebody who did not take it. They are written here rather than
+discovered.
+
+A key has to be held somewhere and rotated eventually. That is the same key
+story record 0021 needs for signed release artefacts, which is why the two
+answers were taken together: one custody story, written once, in
+operations#1609.
+
+An unsigned commit anywhere in a branch's history refuses the merge rather than
+refusing the commit. So the repair is rebuilding the branch rather than adding
+one more commit to it, and it lands at the end of the work rather than at the
+start, which is the moment it is most expensive and least expected.
+
+Edits made through the web interface and anything an automation authors are
+signed by the platform's key or not at all. That decides what those routes can
+still be used for, and it is a consequence rather than a side effect.
+
+**Earlier history does not become signed.** This covers what arrives after the
+day it takes effect, whatever was intended by setting it, so a reader must not
+take a signed default branch for a branch that is signed end to end. Stating
+that here is the point: the rule buys a property from a date, and a property
+from a date read as a property of the whole is worse than no property at all.
+
+The requirement is a setting on the branch protection rather than anything in
+this tree, so nothing here refuses an unsigned commit and this record does not
+claim otherwise. Read the live state rather than trusting this paragraph:
+
+    gh api repos/Flowfin/lab/rules/branches/main --jq '.[].type'
+
+Neither this board nor the board the quality-parity work targets required a
+signature when this was decided, so parity settles nothing here and the ruleset
+walk points at this record instead of deciding it.
+
+## What it applies to
+
+Every commit reaching the default branch of this board from the day the keys
+land, whoever authors it, with no bypass for whoever administers the repository.
+A gate with an exception for the person most likely to be in a hurry is not a
+gate.
+
+It applies to the ruleset work that configures it and to the quality-parity
+document that walks the ruleset.
+
+It does not apply to a feature branch, which no ruleset covers, and it does not
+apply to anything already in history.
+
+It decides that a signature is required and does not decide what happens to a
+contributor who has no key. This board takes experiments from anybody, so that
+question is real rather than hypothetical, and the record answering who may run
+an experiment here is where it belongs.
+
+## What else was considered
+
+Not requiring a signature at all, and resting authorship on the account.
+
+Requiring one from the first release onward rather than from the day the keys
+land.
+
+## What each rejected option would have cost
+
+Resting on the account costs the distinction the whole decision is about. It is
+free to operate and it is exactly as strong as the weakest recovery route on one
+account, which is not a property this board controls or can measure. The cost is
+invisible until the day it is total.
+
+Tying it to the first release costs the same thing the effective date already
+costs, and more of it. A history carrying unsigned commits does not become
+signed afterwards, so the later the rule starts the more of the branch it does
+not cover, and the release is later than the keys. It also splits the key story
+from record 0021 by a stretch of calendar for no gain, since the keys that sign
+the release artefacts are the keys that sign the commits, and having them
+without using them is a cost already paid and not collected.