diff --git a/docs/recovery/fleet-projection-reset-2026-08-28.md b/docs/recovery/fleet-projection-reset-2026-08-28.md
new file mode 100644
index 0000000..bdbd1d5
--- /dev/null
+++ b/docs/recovery/fleet-projection-reset-2026-08-28.md
@@ -0,0 +1,89 @@
+# Fleet projection reset — 2026-08-28
+
+Coordination: `CUST-WP-0068-T09` / `STATE-WP-0083`. Central State Hub image
+`forgejo.coulomb.social/coulomb/state-hub:main-54b09ee` (Helm revision 48,
+schema `e2b3c4d5f6a7`). Reset ran **in-cluster** from pod
+`state-hub-5bd5ffd656-plx7z` via `/tmp/sh-projection-reset.py`. The HTTP
+tunnel does not reach central Postgres; this is the only sanctioned write
+path (`ADR-012` decision 7).
+
+Refuse-mode applies creates and updates when nothing would be retired. It
+refuses only when a record would be tombstoned, or when an identifier is
+already held by another repository. Acknowledgement cannot authorise the
+second class.
+
+## the-custodian (ack)
+
+Status `applied` at commit `989806a`. Retired
+`interactive-the-custodian` (archived `ADHOC-2026-03-29` already live as
+`cust-wp-adhoc-2026-03-29`). Created 13 identified tasks
+(`CUST-WP-0010-T1`/`T2`/`T3`, `CUST-WP-0054-T01`..`T10`). Updated 125
+tasks. Cancelled none. 74 workplans updated (mostly
+`derived_from_commit` / backing path).
+
+A subsequent refuse-mode pass on the same repository was `noop`.
+
+## Fleet refuse-mode
+
+125 repositories. No errors. Zero workplans retired. Zero tasks cancelled.
+
+| Status | Count |
+|---|---|
+| applied | 52 |
+| noop | 69 |
+| refused | 2 |
+| unreadable | 2 |
+
+Totals: 122 workplans updated, 149 tasks created, 745 tasks updated,
+0 retired, 0 released.
+
+Fleet-ack was **not** run. Nothing was refused for retirement, so there
+was no stale set to acknowledge. Re-running with acknowledgement would
+not create the colliding records below.
+
+## Refused — identifier collisions (ack does not authorise)
+
+`railiance-bootstrap` currently holds **no** hub workplans. The forge still
+derives four identifiers already live on `railiance-cluster`:
+
+| Record | Held by | Holder status | Holder path |
+|---|---|---|---|
+| `RAIL-BS-WP-0003` | `railiance-cluster` `7ee9ee22-…` | finished | `workplans/archived/260622-RAIL-BS-WP-0003-pgpool-ha-failover-fix.md` |
+| `RAIL-BS-WP-0005` | `railiance-cluster` `b236de41-…` | finished | `workplans/archived/260622-RAIL-BS-WP-0005-kubeconfig-delivery.md` |
+| `RAIL-BS-WP-0008` | `railiance-cluster` `5032c55c-…` | finished | `workplans/RAIL-BS-WP-0008-activity-core-wp0016-triage-output-deploy.md` |
+| `RAIL-BS-WP-0009` | `railiance-cluster` `2c9e8e96-…` | finished | `workplans/RAIL-BS-WP-0009-activity-core-admin-sync-smoke.md` |
+
+`railiance-hosts` likewise holds no hub workplans. The forge derives
+`RAIL-HO-WP-0004`, already live on `railiance-infra`
+(`cee078e9-…`, finished,
+`workplans/archived/260702-RAIL-HO-WP-0004-production-readiness.md`).
+
+These are extraction leftovers: the files still exist in the older
+repositories, and the identifiers belong to the successor. Disposition
+is a source-file decision in those repos, not a projection acknowledgement.
+
+## Unreadable (STATE-WP-0084)
+
+Forge 404, classified as unreadable rather than empty — nothing retired:
+
+- `vergabe_teilnahme` (underscore slug). The hyphenated `vergabe-teilnahme`
+  applied.
+- `markitect-project`
+
+Private repositories remain unread without a forge token. Absence here is
+not evidence of absence in the files.
+
+## Left on purpose
+
+- `adhoc-2026-08-25@retired-20260827` still has hub `status: active`.
+  The live file is already `cust-wp-adhoc-2026-08-25` (`finished`). Reset
+  correctly skips a tombstoned slug; C-07 still flags the leftover
+  status. That is a status-finish, not another reset.
+- 557 task rows with null `record_id` (34 open) are out of scope: the
+  reset must not guess identity.
+- Three `railiance-wp-0001`/`0002`/`0004` rows remain live because archived
+  files still back them. The unbacked four from T09 (`0017`/`0018`/`0022`/`0023`)
+  are already tombstoned.
+
+Full refuse-mode JSON: `/tmp/custodian-reset/fleet-refuse-2026-08-28.json`
+(pod copy `/tmp/fleet-refuse.json`).
diff --git a/workplans/CUST-WP-0068-cache-only-work-record-recovery.md b/workplans/CUST-WP-0068-cache-only-work-record-recovery.md
index 6c83de8..a17e045 100644
--- a/workplans/CUST-WP-0068-cache-only-work-record-recovery.md
+++ b/workplans/CUST-WP-0068-cache-only-work-record-recovery.md
@@ -468,6 +468,17 @@ longer derives. Rows with no `record_id` are left alone. Tests in
 are all closed (0 open). Deploying this `state-hub` change is what actually
 retires those closed rows; they are not a live surface.
 
+**Central run (2026-08-28).** Image `main-54b09ee`, Helm revision 48.
+`the-custodian` reset-ack applied first (retired `interactive-the-custodian`;
+13 tasks created, 125 updated, 0 cancelled). Fleet refuse-mode then ran
+in-cluster across 125 repositories: 52 applied, 69 noop, 2 refused, 2
+unreadable, 0 errors. Totals: 122 workplans updated, 149 tasks created,
+745 tasks updated, **0 retired**, 0 cancelled. Fleet-ack was not run —
+nothing was refused for retirement. The two refusals are identifier
+collisions (`RAIL-BS-WP-0003/0005/0008/0009` held by `railiance-cluster`;
+`RAIL-HO-WP-0004` held by `railiance-infra`) which acknowledgement does
+not authorise. Evidence: `docs/recovery/fleet-projection-reset-2026-08-28.md`.
+
 ## Disposition of duplicate task identities (2026-08-26)
 
 Establishing task identity (`STATE-WP-0083-T06`) exposed 35 duplicate identities