diff --git a/drills/0.2.0.md b/drills/0.2.0.md
new file mode 100644
index 0000000..a2e6c9f
--- /dev/null
+++ b/drills/0.2.0.md
@@ -0,0 +1,119 @@
+# 0.2.0 — drill record
+
+Run 2026-07-24, by `claude-bot-andresmgsl`, against the release candidate
+`heavy-duty/ceremony@682b9cb8929aa8c50a3101b64ca57fee7b09fef1` (PR #129's
+head at drill time — the ceremony PR adds only `VERSION`, the assembled
+section with its fragment deletions, the self-ref stamps, and this record;
+no machinery differs between the drilled tree and the released one).
+
+Where: scratch private repo `claude-bot-andresmgsl/ceremony-drill-0.2.0`,
+carrying the `docs/CONSUMERS.md` caller stubs (modulo the pin deviation
+below) and a fixture consumer tree in the **fragment shape** this release
+exists to ship: `VERSION` `0.5.0-dev`, `CHANGELOG.md` preamble plus a
+shipped `## 0.4.0` section, `changelog.d/` with its `README.md` marker and
+two fragments, `drills/`, and the five guard actions in CI at the drill
+ref. Archived after this record was written; deletion needs the
+`delete_repo` scope this box's token deliberately lacks, so the operator
+owns the final delete — the repo is the evidence's scaffolding, this file
+is the evidence. Run ids below are kept for the record's own audit trail.
+
+## The deviation: the self-drill deadlock recurs past the first release
+
+The spec (#118) expected the 0.1.0 record's fork deviation to be moot —
+"`0.1.0` exists, so the pinned path runs straight." It does not, for
+ceremony's *own* drill: the candidate tree pins
+`CEREMONY_SELF_REF: "0.2.0"` — the very stamp this PR makes — and the
+consumer path checks out `heavy-duty/ceremony` at exactly that ref, which
+cannot exist before the merge. What `0.1.0` existing does make moot is the
+*consumer-side* deadlock: a governed repo drilling its own release pins
+ceremony at a real tag and runs straight. Raised on #118 before drilling;
+proceeded per the deviation triage already ruled for #11, unchanged in
+shape: **no `0.2.0`-named ref on the canonical repo**; the scratch callers
+pin the fork ref `claude-bot-andresmgsl/ceremony@drill/0.2.0` — the
+candidate tree with exactly one line changed per pin carrier
+(`release.yml` and `labels.yml`, both — `self-ref-check.sh` requires the
+carriers to agree), `CEREMONY_SELF_REF: "0.2.0"` →
+`"682b9cb8929aa8c50a3101b64ca57fee7b09fef1"`. Every `.ceremony-src` byte
+the doors executed still came from `heavy-duty/ceremony` at the candidate
+SHA (reachable there via PR #129's pull ref); the fork supplied
+orchestration YAML differing by the one line the ceremony stamps each
+release anyway. Drilling against `0.1.0` instead would not have been a
+drill of the candidate: `lib/changelog.sh`, which both doors source for
+notes extraction, is +214 lines since `0.1.0` (fragment mode; #98).
+
+## The probes
+
+| # | probe | run | result |
+|---|---|---|---|
+| a | merge-door ceremony, fragment shape | 30089779736 (attempt 1) | ✅ exactly one release |
+| b | mislabeled ordinary PR | 30089842547 | ✅ green NOTICE no-op |
+| c | bare-version PR, no label | 30089908740 | ✅ refused, red, created nothing |
+| d | re-run of the completed ceremony | 30089779736 (attempt 2) | ✅ refused, red, created nothing |
+| e | tag door, manual matching tag | 30090053523 | ✅ published, no bump |
+| f | tag door, mismatched tag | 30090082676 | ✅ refused, red, created nothing |
+
+**(a) The ceremony, in fragment shape.** PR `release: 0.5.0` — hand-set
+`release` label, `VERSION` `0.5.0-dev` → `0.5.0`, and the stamp made by
+the candidate's own `bin/changelog-assemble 0.5.0`: two fragments folded
+into `## 0.5.0 — 2026-07-24` (newest issue first), deleted in the same
+commit, `changelog.d/` left holding only its `README.md` marker — no
+re-arm edit, armed by existing. The five guard actions at the drill ref —
+`changelog-armed`, `changelog-monotonic`, `changelog-assembled`,
+`drill-recorded`, `runner-isolated` — all green on that PR (run
+30089703224): the assembled guard replayed the assembly from the merge
+base and matched it byte-for-byte, live, on the mechanism this release
+ships. Merged. The one run: facts
+`ver='0.5.0' base_ver='0.5.0-dev' labeled='yes'` → `ceremony=yes` → tag
+`0.5.0` on the merge commit (`9b9e58e`), release `0.5.0` published with
+the assembled section — verbatim, both bullets — as its body, and main
+re-armed to `0.5.1-dev` by the run itself. The GITHUB_TOKEN tag create
+and bump push fired **no** second run (anti-recursion held).
+
+**(b) Mislabeled ordinary PR.** Docs-only PR wearing the `release` label.
+Green: `NOTICE: the version '0.5.1-dev' is -dev and unchanged by this PR —
+release-flow work under the release label, not a ceremony. Nothing to
+publish.` No tag, no release.
+
+**(c) Bare-version PR without the label.** The ceremony's edits —
+`0.5.1-dev` → `0.5.1`, a section assembled from a fragment — label
+deliberately absent. Red at decide:
+`the version transitioned ('0.5.1-dev' -> '0.5.1') but no merged,
+release-labeled PR is behind this commit — a release is a labeled ceremony
+PR, not a bare push — creating nothing.` No tag, no release; main
+reverted to the armed state afterwards, and the revert push itself ran
+green as a no-op (run 30090017774), as it should.
+
+**(d) Re-run of the completed ceremony.** Re-ran (a)'s run. Red at the
+nothing-may-exist assert:
+`tag '0.5.0' already exists — this release already happened, or a manual
+tag won the race; refusing to re-release, creating nothing.` Release
+count still exactly one.
+
+**(e) Tag door.** Branch with `VERSION` `0.6.0` and its section assembled
+from a fragment; manual `0.6.0` tag pushed at its head (never merged to
+main). The tag door published release `0.6.0` with that section as its
+body; the merge door correctly skipped. Main untouched: `VERSION` stayed
+`0.5.1-dev` — the fallback does not rewrite main.
+
+**(f) Mismatched tag.** Tag `9.9.9` pushed at main (tree version
+`0.5.1-dev`). Red:
+`tag '9.9.9' does not match the tree's version '0.5.1-dev' — creating
+nothing.` No release, and the operator-facing remediation text printed as
+designed.
+
+## Failures
+
+None among the probes; every refusal path was checked for droppings — the
+only releases and workflow-created tags that exist are (a)'s and (e)'s.
+
+One finding outside the probe set, written down plainly: the fixture's
+**initial push** — the very first commit of the scratch repo, caller
+included — ran the merge door red (run 30089682128): `lib/facts.sh`
+resolves the base as `MERGE_SHA^1` when `event.before` is all-zeros, and
+a repo's first commit has no parent, so `git rev-parse` fails loudly
+(exit 128) before deciding. Nothing was created, and every subsequent
+push behaved; but a greenfield consumer bootstrapping caller-first will
+meet one red run on its first push, which the doctrine's "every
+legitimate non-ceremony is a green NOTICE no-op" does not intend.
+`facts.sh` is unchanged since `0.1.0` — not a candidate regression, not a
+blocker for this release; escalated as a discussion, not fixed here.